How Professional Skills Become Commodities | What Happens When Expertise Becomes Cheap

Professional skills become cheaper when technology, standardisation and wider access reduce the cost of producing work that once required scarce expertise. In Singapore’s changing labour market, this can happen when software automates a calculation, AI produces a competent first draft, a public knowledge base makes specialist information easier to reach, or a standard operating procedure captures what an experienced person once carried mainly in memory. The result is not that professional expertise becomes worthless. The question is which part of expertise has become easier to reproduce, which part still requires judgment, and how workers and institutions respond when the old scarcity changes.

Cheap expertise is not the same as cheap professionals. A tool can lower the cost of one task while increasing demand for verification, integration, responsibility or unusually difficult cases. It can also lower the premium once paid for producing a routine output. A first draft that took an hour may take ten minutes; deciding whether it is correct, appropriate and safe may still depend on deep knowledge. The economic pressure appears when organisations discover which tasks can be standardised, which can be delegated to tools, and which forms of human contribution remain scarce enough to command attention and pay.

The central proposition is that expertise does not disappear when some outputs become abundant; its value moves. The worker who once created every first version may increasingly define the problem, choose the evidence, recognise an exception, verify an answer, combine several domains, communicate a difficult trade-off or take responsibility for a consequential decision. This shift can broaden access to useful capability and compress some old advantages at the same time. It can also create new divides between workers who learn to supervise inexpensive capability and workers whose tasks are valued mainly because that capability used to be scarce.

HOW X WORKS · SINGAPORE · ARTICLE 22
How Singapore Works | The Problems We Need to Understand.
Evidence reviewed: 23 September 2026. All named workers, firms, prices, salaries, workflows and numerical examples below are fictional teaching illustrations unless a source is expressly identified. They are not actual eduKate clients, employer records or labour-market forecasts.

Evidence boundary: current research shows that generative AI can improve performance on some professional tasks and worsen it on others. Singapore’s official 2026 data show uneven AI adoption, reported productivity gains among many adopting firms, and more job redesign than reported headcount reduction so far. None of these findings proves that every professional skill is becoming commoditised or predicts a particular worker’s future. This article uses them to investigate the mechanism without turning exposure into destiny.

How X Works → Singapore capability series → Changing work divide. Previous published article: Article 21 — How the AI Capability Divide Works.

The 50-second route: what exactly became cheap?

A consultant can now obtain a competent outline in seconds. A junior analyst can ask a tool to explain unfamiliar code. A small business can generate a first customer reply without hiring a specialist writer. These changes make a particular output cheaper. They do not automatically make diagnosis, verification, accountability or domain knowledge cheap. Start by decomposing the work. Which step used to consume scarce professional time? Which step has become easier? What error becomes more likely when the cheap step is accepted without judgment? Which human action still changes the quality or consequence of the final result? The worker’s response should follow that map rather than a slogan about AI replacing experts or experts remaining untouchable.

For a professional: what commoditisation means → map the task → find the new scarcity → 90-day capability audit. For a manager: workflow laboratory → quality and verification → preserve the learning route. For students and families: education implications → credentials → Teaching Guide. For policy readers: Singapore evidence → task exposure → shared capability.

Open the complete chapter map

1. What commoditisation means · 2. Price, value and scarcity · 3. Tasks, not titles · 4. Standardisation · 5. Singapore’s evidence · 6. Exposure is not replacement · 7. Professional writing evidence · 8. Jagged-frontier evidence · 9. Tacit knowledge and support work · 10. The first-draft collapse · 11. Verification becomes scarce · 12. Context remains local · 13. Diagnosis · 14. Selection and taste · 15. Responsibility · 16. Coordination · 17. The scarcity shift · 18. Novices · 19. Experts · 20. Wages and recognition · 21. Workflow laboratory · 22. Pricing laboratory · 23. Preserve junior learning · 24. Education · 25. Credentials · 26. Evidence of capability · 27. Small firms · 28. Shared infrastructure · 29. Career response · 30. 90-day capability audit · 31. Three workers · 32. The system question · Reader questions · Sources · Teaching Guide.

1. A professional skill becomes commodity-like when the ordinary version is easier to obtain

Commoditisation does not mean that two pieces of work become literally identical. It means that buyers can obtain a sufficiently acceptable version from many sources, at lower cost and with less dependence on one scarce specialist. The output becomes easier to compare and substitute. In professional work, this can occur through standard templates, widely taught methods, software, online knowledge or AI systems that reproduce patterns previously available only through trained labour.

Consider an invented market for simple product descriptions. Years ago, a small business may have paid a writer because producing a competent paragraph required time and language skill the owner did not possess. If a current tool can create ten plausible drafts immediately, the scarcity attached to first-draft production falls. That does not establish that every draft is accurate, persuasive, legally appropriate or aligned with the brand. It identifies the part of the workflow whose supply has expanded.

The price pressure follows substitutability. If five alternatives produce an acceptable ordinary output, a buyer has less reason to pay a premium merely for access to that output. The writer can still create value by discovering what should be said, recognising unsupported claims, understanding an audience, developing a distinctive voice or handling a difficult case. Those tasks may remain scarce precisely because the generic draft has become abundant.

A professional skill can therefore contain both commodity-like and scarce components. Spreadsheet arithmetic may be routine while deciding which denominator matters is not. Summarising a policy may be easy while recognising that one clause changes the decision is harder. Generating code may be inexpensive while understanding the production environment, security requirements and unintended effects remains consequential. The useful decomposition follows the work rather than awarding the whole occupation one label.

Standardisation can commoditise skill without AI. A tax form, engineering code, template contract or standard clinical pathway can make some decisions more repeatable and accessible. This can be socially valuable. It reduces dependence on scarce personal memory and can improve consistency. The professional role often moves towards interpretation, exceptions and responsibility rather than disappearing. AI is part of a longer history of making previously specialised procedures easier to reproduce.

Do not infer that cheaper production means lower total demand. When something becomes less expensive, more people may use it. Cheap photography did not remove images from economic life; it expanded image production while changing who could create and sell ordinary pictures. A cheaper first analysis could similarly increase how often organisations ask questions, creating more downstream review or decision work. Whether this happens in a particular occupation requires evidence about demand, not analogy alone.

The opposite is possible too. A task may become cheap and demand may not expand enough to preserve all existing labour. If ten people previously produced a routine deliverable and one person with software can now produce the needed volume, the organisation may redesign staffing. The outcome depends on the workflow, market demand, quality standard and new tasks. A serious explanation allows both augmentation and substitution rather than declaring one universal future.

The professional response starts by naming what the customer or organisation actually buys. Is it a document, an answer, a trustworthy decision, a legally accountable opinion, a relationship, a response time or a reduced risk? The easier the output becomes to obtain, the more important it is to know whether the output itself was ever the final source of value. Commoditisation reveals which part of the old bundle was scarce and which part merely looked scarce because the tools were expensive.

2. Price, value and scarcity move together imperfectly

A useful professional service can become cheaper without becoming less valuable to the user. Suppose a small firm needs a basic translation of an internal notice. If technology reduces the cost while preserving adequate accuracy for that purpose, the firm obtains the same useful function at lower price. The translator’s old price premium may shrink, but the social value of accessible translation may increase because more people can use it. Value to the user and scarcity of the producer are different variables.

Now make the notice legally consequential or culturally delicate. A generic translation may be insufficient. The user may pay for a qualified professional who can verify terminology, understand the jurisdiction, discuss ambiguity and accept responsibility under the relevant professional framework. The base capability has become cheap while the high-stakes layer remains scarce. The market can split rather than simply collapse.

Price is also shaped by bargaining, institutions and business models. A worker can create more value without automatically receiving more pay. An organisation can capture part of a productivity gain, reduce prices, increase output or invest in another activity. A consultant can deliver faster work while the client continues to pay for trust and accountability rather than minutes. The relationship between task cost and labour income is mediated by decisions; it is not an automatic equation.

Consider an invented professional task that once required four hours at S$100 per hour, giving S$400 of labour cost. A new tool reduces direct drafting to one hour but adds one hour of verification. The direct professional time falls from four hours to two. If the same hourly rate is used mechanically, cost falls to S$200. But the final price might be S$250 because the firm charges for the tool, S$400 because the buyer pays for the service outcome, or something else entirely. The arithmetic alone cannot determine the market model.

Quality thresholds can create discontinuities. A 90%-accurate draft may be acceptable for brainstorming and unacceptable for a filing, medical instruction or safety decision. A small increase in accuracy can therefore have large value when it crosses the threshold required for use. The professional who knows whether the threshold has been met may retain substantial value even when generating the underlying content is inexpensive.

Scarcity can also move to attention. When everyone can create reports, presentations and analyses cheaply, the receiving organisation faces more material than people can review. The scarce resource becomes the ability to decide which work deserves attention and which output can be trusted. A world of abundant production can create stronger demand for selection. This is why generic output and professional judgment can move in opposite economic directions.

The worker should therefore avoid measuring value solely through how difficult a task feels. A repetitive manual process can be exhausting and easy to automate. A short review decision can be cognitively demanding and consequential. The market may pay for the scarce decision rather than the volume of effort. Learning should focus on the part of the workflow that remains necessary after the easiest production step becomes cheap.

This distinction also protects against prestige stories. A skill is not noble because it is expensive, and a cheap capability is not worthless because it is widely available. Public education deliberately makes important capabilities less scarce. The economic challenge is what happens to people whose income depended on that scarcity. The social opportunity is that more people can now use knowledge once limited by price or specialist access.

3. Start with tasks, because occupations are too large to commoditise all at once

A professional title can contain dozens of tasks with different relationships to technology. A lawyer may research, draft, interview, negotiate, advise, manage evidence and take responsibility for a legal position. A teacher may explain, question, assess, motivate, plan, coordinate and safeguard. A software engineer may design, code, test, review, deploy and investigate incidents. Calling the whole occupation automatable hides the difference between tasks that can be generated and responsibilities that remain institutionally human.

The International Labour Organization’s 2025 refined index assesses generative-AI exposure at task level and concludes that transformation is more likely than wholesale redundancy for most exposed occupations. It reports that one in four workers globally is in an occupation with some degree of exposure, while only a smaller share falls in the highest exposure category. This is a global potential-exposure framework, not a forecast that one quarter of jobs will disappear or a Singapore employment estimate. [1]

Task maps are useful precisely because exposure can be mixed. Imagine a financial analyst whose weekly work contains ten hours of data cleaning, six hours of routine commentary, six hours of investigation, five hours of stakeholder discussion and three hours of approval-related documentation. AI or software might reduce some cleaning and commentary time. It does not follow that the analyst’s entire role falls by the same percentage. New review or exception work may appear, and the organisation may use the saved time differently.

Write each task as a verb and object: reconcile records, explain variance, verify source, interview client, choose method, approve change. Then ask four questions. Can a tool perform the action under the actual conditions? Can another person verify the output cheaply? What happens if the output is wrong? Who has authority or responsibility for the decision? Tasks with low error cost and cheap verification are more likely to become commodity-like sooner than tasks with high consequence and difficult verification.

Some tasks are bundled for institutional reasons rather than technical ones. A professional may be required to sign a document because the institution places responsibility on that role. A tool can assist without acquiring the legal or organisational authority to sign. Another task may be technically difficult but not formally reserved. The future bundle depends on both capability and rules. Technology does not determine institutional authority automatically.

Task decomposition also helps workers avoid generic panic. A writer who sees first-draft production become cheaper can inspect which other actions are important in their work. A junior analyst can ask which questions a senior reviews. A teacher can distinguish generating a worksheet from recognising the misconception behind a student’s answer. The map replaces “my job is disappearing” with a set of smaller questions, some of which may have different answers.

Managers benefit for the same reason. Buying an AI subscription is not job redesign. The organisation still needs to decide which tasks move, what quality standard applies, which exceptions require escalation and how people learn the new responsibilities. A poorly mapped workflow can automate an upstream step while creating hidden review queues downstream. The later laboratory makes this visible with a simplified example.

The most important result of the task map is not a percentage of work that can be automated. It is the identification of changed dependencies. If draft generation becomes trivial, what becomes the bottleneck? If knowledge lookup is easier, what must the professional now verify? If a novice can perform an ordinary task earlier, what experience will still teach them to recognise an unusual one? Commoditisation changes the sequence through which expertise is built as well as the price of its output.

4. Standardisation makes expertise shareable before it makes it cheap

Every checklist captures a small part of expertise. Someone noticed that a sequence matters, wrote it down and made the knowledge easier for another person to use. Templates, taxonomies, protocols and standard forms perform similar work. They move knowledge from an individual’s memory into a shared system. This can improve quality and access even before technology reduces the cost of execution.

Consider an invented incident handover. An experienced employee naturally tells the next person the current state, what has been tried, the unresolved risk and the next owner. A new employee does not know which details matter. A structured handover template makes the relationship explicit. The veteran’s tacit skill has not disappeared; part of it has been converted into a standard that others can follow. The organisation has deliberately made one component of expertise less scarce.

Once the structure is formalised, software can help fill it, check required fields or summarise the content. The standardisation therefore creates a surface on which automation becomes easier. This is not necessarily an argument against standardisation. A hospital checklist, engineering code or accounting standard can be valuable precisely because it makes a safe minimum less dependent on heroic memory. The important question is which exceptions still require judgment and how people learn to recognise them.

Standardisation can fail when the template becomes the work. Employees may learn to complete boxes without understanding why a condition matters. A rare case can fit the form badly. Management may count completed forms instead of checking whether the underlying decision improved. The expert’s role then shifts from remembering the standard to interpreting its boundary. A competent professional knows when the ordinary process is insufficient and how to escalate responsibly.

A standard can also create a market. Once buyers know what an ordinary deliverable should contain, providers become easier to compare. Competition can lower prices. That can benefit users while pressuring providers whose advantage was mainly familiarity with the standard. Providers can respond through scale, specialisation, reliability, service or more difficult work. The market consequence depends on what customers value after the baseline becomes common.

Education is another standardisation system. It teaches language, mathematics, professional methods and common concepts to large populations. That process deliberately expands the supply of capability. A degree can lose some scarcity as more people earn it without becoming educationally worthless. The social success of making knowledge widespread and the labour-market consequence of a larger supply can coexist. This is one reason career planning cannot treat scarcity as the sole measure of educational value.

For a worker, a useful question is whether their current advantage is a method that can be written down easily or a judgment that still requires context and experience. Even judgment can sometimes be standardised partly. The answer can change over time. Treat the map as something to revisit, not as a declaration that one type of knowledge is permanently safe. Professional development becomes more realistic when it follows what is becoming explicit and asks what remains difficult to transfer.

The next chapter applies these ideas to Singapore’s current evidence. The national picture does not show a complete transition in which professional work has already been commoditised. It shows uneven adoption, reported productivity gains among many adopters and considerable job redesign. That is enough to make the mechanism important to study without pretending the future has already arrived everywhere.

5. Singapore’s current evidence points to uneven adoption and changing work, not a completed replacement story

MOM’s inaugural April 2026 report on AI adoption among private-sector establishments with at least ten employees found that 28.5% had started adopting AI, while 71.5% had not. Only 3.8% were described as integrating AI into core processes at that point. Adoption varied sharply by firm size and sector. These figures establish uneven adoption in the surveyed population; they do not describe every microbusiness or prove how deeply every reported tool is used. [2]

Among firms using AI, 70.7% reported improvements in worker productivity. At the same time, 6.2% of adopting firms reported reduced headcount, while 18.9% reported job redesign and 13.9% reported creating new AI-related jobs. The report therefore does not support a simple claim that AI adoption in Singapore had already produced widespread job replacement. Nor does a reported productivity gain establish its causal size or who received the benefit. [2]

An August 2026 parliamentary answer gives a further size comparison: 27.2% of firms with fewer than 200 employees had adopted AI, compared with 54.8% of firms with 200 to 500 employees and 76.4% of firms with more than 500. The answer also states that larger adopting firms were more likely to report job redesign, new AI-related jobs and redeployment. These are descriptive survey findings, not proof that size itself caused the outcomes. [3]

A September 2026 MOM answer on employment pathways amid AI and automation says current surveys show most adopting firms redesigning jobs or creating new roles rather than reducing headcount, and describes practical pathways for workers who may need to move into different roles. This is a current policy statement and a description of support direction, not a guarantee that every worker will make an easy transition or that future employment effects will remain unchanged. [4]

The evidence is consistent with a market in which some professional outputs become cheaper before whole occupations disappear. A firm can redesign a role because drafting or analysis becomes faster. It can move workers toward review, client interaction or exception handling. Another firm may reduce staffing in a task area. A third may not adopt the technology at all. The national average is the combination of different local decisions, not one universal workflow replicated everywhere.

Sector differences matter too. MOM’s April report recorded particularly high adoption in Information and Communications, Professional Services, and Financial and Insurance Services. Those sectors contain many digitised knowledge tasks. This does not establish that every occupation within them has the same exposure or that adoption makes every skill equally replaceable. It does make the professional-skills question especially relevant because the work already travels through digital systems where outputs can be copied, compared and automated more easily.

Official programmes are also designed around job redesign rather than only worker replacement. The Enterprise Workforce Transformation Package and related support for reskilling and organisational capabilities reflect an institutional view that technology adoption and workforce design need to move together. An employer can still implement the programmes poorly. A worker can still face displacement. The policy architecture simply illustrates that the problem being addressed is larger than teaching individuals how to prompt a tool. [5]

The current evidence therefore justifies neither complacency nor inevitability. Professional skills can lose scarcity without every professional losing employment. A worker can become more productive without automatically receiving higher pay. A firm can adopt AI without achieving meaningful integration. The right next question is which part of the work has changed and what new scarcity the redesigned system creates.

6. Exposure means a task can be affected; replacement means the organisation no longer needs the worker for that function

The ILO’s 2025 refined global index assesses generative-AI exposure across nearly thirty thousand tasks and maps them into occupations. It reports that one in four workers globally is in an occupation with some degree of GenAI exposure, while 3.3% of global employment falls into the highest exposure category. The ILO emphasises transformation rather than redundancy as the more likely outcome overall. Exposure is a technological potential, not a forecast that the same share of workers will lose jobs. [1]

This distinction is easiest to see with a spreadsheet. A finance worker may use formulas to automate arithmetic. The arithmetic is exposed and partly automated. The worker’s role can remain because they decide which records belong, interpret unusual entries and explain the result. If another person or system can also perform those decisions at sufficient quality and lower cost, more of the bundle becomes substitutable. The replacement question therefore depends on the whole workflow and the organisation’s demand.

Exposure can even increase demand for the complementary work. If a drafting tool makes it inexpensive to generate many proposals, an organisation may need more review, selection or client discussion. The total amount of labour can rise or fall depending on how much additional output is demanded and how quickly the complementary tasks can be performed. A claim about exposure cannot settle this quantity without an economic model and evidence.

For a worker, the most actionable exposure map is local. List the tasks you perform, the inputs they require, the outputs they create and the consequences of mistakes. Then identify which steps a tool can perform under the actual conditions. Test the tool on representative cases rather than the easiest demonstration. Record where human judgment changes the result. This turns exposure from a frightening percentage into a practical learning agenda.

For education, exposure changes the meaning of mastery. A student may no longer need to memorise every routine transformation that a tool can perform, but they still need enough knowledge to recognise when the transformation is appropriate and to inspect its output. The exact balance depends on subject and purpose. Removing foundational knowledge simply because software exists can make verification impossible. Preserving every historical routine regardless of changed tools can waste learning time. Curriculum design must decide what understanding is required to supervise the new environment.

For employers, exposure raises a governance question. If a task can be performed automatically, who remains responsible for quality? Who handles exceptions? What evidence is kept? Which outputs require review before action? These questions are not administrative leftovers after automation. They define the boundary between cheap production and trusted use. A workflow without them can create fast outputs and slow crises.

Replacement should therefore be demonstrated by observed staffing and task reallocation, not inferred directly from what a model can technically produce. A tool can write a legal-looking paragraph without being the organisation’s lawyer. It can create code without being authorised to deploy it. It can summarise data without knowing whether the source definition changed. Technical capability and organisational substitution are connected through responsibility, standards and demand.

7. Professional writing became faster in an experiment, and the weaker performers benefited more

Noy and Zhang’s 2023 Science experiment assigned occupation-specific writing tasks to 453 college-educated professionals and randomly gave half access to ChatGPT. In the studied tasks, average time fell by 40% and assessed output quality rose by 18%. The study also reported reduced inequality between workers because participants with weaker baseline performance benefited more. These findings concern a bounded set of midlevel professional writing tasks, not all professional work. [6]

The result illustrates one path toward commoditisation. If a person with less writing skill can produce a more competent routine document quickly, part of the premium attached to drafting skill may compress. The organisation can obtain an acceptable baseline from more people. This can broaden participation and reduce dependence on a small set of naturally fluent writers. It can also reduce the labour time previously sold by specialists for ordinary drafting.

What remains scarce depends on the use. A manager may still need someone to define the problem, select the evidence and judge whether a recommendation overstates it. A communicator may need to manage stakeholder sensitivities. A writer may create a distinctive voice that a generic draft does not supply. The experiment does not show that these tasks disappeared; it shows that the studied writing tasks changed in speed and quality under AI access.

The worker response should not be to hide the tool or to use it invisibly for every assignment. Instead, identify which part of the writing task the tool performs and which part you can demonstrate independently. Can you judge whether the tone fits? Can you find a factual error? Can you restructure an argument when the source evidence changes? These are stronger signals of capability in a world where producing grammatical prose is less scarce.

The employer response should also avoid metric confusion. If drafting time falls, the saved time is a potential resource. It can be used for additional work, review, learning or reduced workload. A productivity claim requires a defined outcome. More words per hour are not automatically more value. The process must still deliver the communication job it exists to perform.

Students can use the same distinction. A generated paragraph may demonstrate that a machine can produce the form. The educational task becomes explaining why one claim is supported, revising a misleading sentence or writing from evidence the student has actually interpreted. Schools should preserve authorship requirements and make permitted tool use explicit. The goal is not to pretend cheap language generation does not exist; it is to teach the judgment needed because it does.

8. The jagged frontier means cheap expertise can become dangerously persuasive

Dell’Acqua and colleagues’ study of 758 BCG consultants, formally published in Organization Science in 2026, found large gains on tasks inside the studied AI capability frontier: participants using GPT-4 completed 12.2% more tasks, worked 25.1% faster and produced higher-quality solutions. For a complex managerial task selected to be outside the frontier, AI users were 19 percentage points less likely to produce correct solutions. The same technology therefore improved some professional tasks and worsened another. [7]

This is economically important because cheap output can look expert before it is reliable. A buyer may receive a polished answer and assume the difficulty has been solved. The more fluent the output, the harder it can be for a non-expert to notice a subtle error. The scarce professional contribution may shift from producing plausible content to recognising when plausibility should not be trusted.

The study’s “jagged technological frontier” is a reminder that neighbouring tasks can behave differently. A model may handle brainstorming and writing effectively while failing on a problem that requires integrating details in a way users do not anticipate. Professional development therefore cannot be reduced to learning a fixed list of tasks that AI is good or bad at. The frontier changes as tools, data and workflows change.

A responsible professional learns to test the frontier within their work. Choose representative cases, include rare or consequential variations and compare the output with trusted evidence. Record where the tool needs supervision. Do not let a successful demonstration on one task authorise use on another merely because the interface looks the same. The boundary is empirical and local.

The result also complicates a simple novice-versus-expert story. AI can raise lower performers on some tasks and still make experienced professionals vulnerable to overreliance outside the frontier. Expertise remains valuable partly because it can detect when the task has changed. Yet experts can also be misled when a tool produces an answer that fits their expectations. Verification must be designed into the workflow rather than treated as a personal virtue that never fails.

For managers, the cheapest workflow is not necessarily the one with the most AI. It is the one that achieves the required outcome with acceptable quality and risk. Some tasks may benefit from heavy automation. Others may require human-led analysis with AI support. A single organisational policy can be too crude when the frontier is jagged. The process needs categories, examples and escalation routes that reflect actual tasks.

For students, this evidence strengthens rather than weakens the case for deep knowledge. If a tool sometimes gives better answers and sometimes gives worse ones, the learner needs concepts against which the answer can be checked. Memorising every first draft becomes less important. Understanding enough to know what must remain true becomes more important. This is a shift in the educational value of expertise, not its disappearance.

9. AI can spread tacit expert patterns to less-experienced workers

Brynjolfsson, Li and Raymond’s 2025 Quarterly Journal of Economics study examined the staggered introduction of a generative-AI assistant among 5,172 customer-support agents. Access increased issues resolved per hour by 15% on average, with larger gains among less-experienced and lower-skilled workers and smaller gains among the most experienced. The authors also report evidence consistent with AI helping diffuse some practices of stronger workers. [8]

This finding illustrates how expertise can become partly codified without a formal manual. A system trained on past interactions can expose a newer worker to response patterns that previously took time and experience to acquire. The novice’s ordinary performance improves. The skill gap narrows. This can be good for customers and workers while changing the market value of the experience once required to reach that baseline.

The experienced worker’s value may move toward the cases where the system lacks enough data, where the customer is unusual or where rules conflict. The expert may also contribute to improving the system itself by identifying better patterns. That contribution needs recognition. If the organisation captures expert knowledge into a tool and then treats the underlying expertise as no longer valuable, it may weaken the source from which future improvements are learned.

There is also a training paradox. If novices receive good recommendations immediately, they may perform well without fully understanding why. That can accelerate learning when feedback and explanation are designed well. It can also create dependence if the tool carries the reasoning permanently. A workplace should decide which decisions juniors need to learn and create opportunities where they must explain or verify those decisions rather than merely accept suggestions.

The customer-support study does not establish that AI always compresses skill differences. Other tasks may reward deep expertise more strongly or create new gaps between workers who know how to use the technology and those who do not. Singapore’s own adoption data are uneven across firms. A capability can become cheaper in one organisation while remaining scarce in another. Market effects can therefore unfold at different speeds even within the same occupation.

The practical worker question is, “Which expert behaviour is the tool helping me reproduce, and do I understand it well enough to recognise when it should change?” If the answer is no, the current output may be useful without being a finished professional capability. The learning route should make the logic visible before the person is asked to handle exceptions without assistance.

10. When the first draft becomes almost free, the bottleneck moves downstream

Imagine a fictional team that needs twenty routine briefing notes each week. Before AI, each note takes ninety minutes to draft and thirty minutes to review: forty hours altogether. A tool reduces drafting to fifteen minutes but review rises to forty-five because reviewers must inspect citations and subtle claims. Total weekly time becomes twenty hours. The workflow has halved the time, but review now consumes three quarters of the remaining labour.

If management measures only drafts produced, it may conclude that the team has enormous spare capacity. If it measures only review time, it may conclude that quality assurance has become inefficient. The full workflow shows what changed. Cheap first drafts create a new bottleneck. A rational redesign might improve source controls, narrow the tasks eligible for automation, train writers to pre-check the output or invest in better review tools. Simply demanding more drafts increases the queue.

The same effect appears in education. A student can generate ten essay outlines quickly. The scarce classroom resource becomes the teacher’s attention to whether any outline reflects the text accurately. Producing more drafts can therefore increase the feedback burden without increasing learning. The solution is not necessarily banning the tool. It may be requiring the student to choose one outline, justify the evidence and show where they revised the generated structure.

In software, cheap code generation can increase the volume of code reaching review. Reviewers need tests, context and security knowledge. If a team merges faster than it can verify, defects can accumulate invisibly. The value of the senior engineer may move from typing code toward architecture, review, incident diagnosis and judgment about trade-offs. Some typing skill remains necessary because understanding and modification still matter, but its relative scarcity changes.

For a client-service professional, cheap drafts can increase responsiveness. A customer receives a first answer sooner. The service can be genuinely better if the final answer remains accurate. The economic value may therefore rise even while the cost of draft production falls. This is why professional commoditisation should not be analysed as a simple fall in value. The organisation can create more total value by combining cheaper production with scarce judgment.

The worker should follow the bottleneck. If first drafts are no longer scarce, becoming faster at first drafts may offer less advantage. Learning to verify a source, resolve an ambiguity or make a decision under incomplete information may become more useful. The bottleneck can change again when tools improve. Career development is therefore a repeated search for the part of the workflow where competent human judgment still changes the outcome materially.

11. Verification becomes more valuable when production becomes cheap

Verification is easy to praise and hard to define. It is not simply reading the output again. A useful check compares the result with an independent standard: a source, a calculation, a test, a policy, a client requirement or another defensible reference. When first drafts are expensive, organisations may produce relatively few of them. When drafts become cheap, verification capacity can become the scarce control that determines whether abundance turns into value or noise.

Consider a fictional financial commentary. The tool writes that revenue rose 12% because customer demand strengthened. The underlying table shows revenue rose 12%, but no demand measure appears. The first clause is supported; the causal explanation is not. A reviewer who checks only the arithmetic may miss the unsupported inference. Verification requires understanding what each piece of evidence can establish, not merely confirming that the numbers were copied accurately.

A second layer asks whether the evidence itself is current and defined consistently. Two reports may use the word active customer differently. A perfect summary of both still produces a misleading comparison if the denominator changed. The professional contribution is identifying the definition before accepting the neat chart. In this case, the scarce skill is not prose generation or arithmetic. It is semantic control over the measurement.

Verification also has a cost. If every automated output requires more senior review time than the old workflow consumed in total, the system has not created the expected capacity. The organisation may need better input controls, narrower automation scope or stronger junior verification skills. The answer is not automatically more senior reviewers. A well-designed workflow reduces the number of cases that require expensive judgment by making routine checks reliable at the appropriate level.

Workers can learn verification through contrast. Give an accurate output and a plausible output containing one material error. Ask what evidence distinguishes them. Then change the error type. A learner who memorises that citations are the problem may miss a numerical definition; one who learns to ask what must be true can adapt. Verification is a family of decisions, not a single checklist item.

The labour-market implication is that expertise attached to verification may gain relative importance even if its direct production role shrinks. That does not guarantee higher wages. Employers may standardise verification too, or the market may not pay a separate premium. The claim is narrower: when production cost falls faster than checking cost, the workflow’s bottleneck moves toward checking, so the organisation has stronger operational reasons to value competent verification.

12. Context remains local even when the generic answer becomes global

Generative systems can produce language about many industries, jurisdictions and practices. That breadth is useful. It can also disguise a missing local condition. A global explanation of employment law does not establish an individual contractual obligation in Singapore. A generic marketing template does not know the organisation’s actual claim approval process. A software suggestion does not know which legacy dependency makes one implementation unsafe unless that information is supplied and interpreted correctly.

Context includes definitions, authority, timing and purpose. A rule may differ by cohort or date. A field name may have a company-specific meaning. A customer request may be constrained by an earlier agreement. An experienced worker often appears slow because they are carrying this context. A faster generic answer can therefore be less useful when it ignores the conditions that make the local decision difficult.

This is one source of continuing scarcity for professionals. The organisation needs people who know what cannot be inferred from the surface. They understand which person has authority, which exception requires escalation and which old decision still shapes the current task. Some of that knowledge can be documented. Some is discovered only when a changed case reveals an assumption nobody had written down.

Local context should not become a defence of unnecessary opacity. If the same hidden rule causes repeated confusion, document it. If a process relies on one veteran remembering every exception, the organisation has a fragility problem. The goal is to make useful context shareable while preserving the professional ability to recognise when the documented model is incomplete.

A student encounters the same distinction. A formula learned in one question does not automatically apply to another merely because the page looks similar. The learner needs to identify the conditions. Professional judgment is often the adult version of this educational skill: knowing what has to remain true before a familiar method can be used safely.

AI can help surface context by asking questions, comparing documents or summarising prior decisions. That can make local knowledge easier to use. But the person deciding which documents are authoritative and which details matter still performs an important function. The more information the system can retrieve, the more the human may need to distinguish relevance from availability.

13. Diagnosis is valuable because the wrong problem makes cheap solutions expensive

A cheap answer is useful only when it answers the right question. Diagnosis identifies the mechanism producing the observed problem. In education, a low mark can reflect vocabulary, representation, knowledge, timing or answer scope. In business, a delayed process can reflect demand, approval, data quality, staffing or an unclear handover. A tool that produces solutions before the mechanism is understood can accelerate the wrong intervention.

Imagine a fictional service team with slow response times. Management asks an AI system for ways to write replies faster. Drafting time falls by half, but overall response time barely changes. A process map reveals that most delay occurs while waiting for information from another department. The writing problem was visible; the handover problem was binding. The cheap solution improved a non-bottleneck.

Experienced professionals often earn trust because they can identify which question matters before solving it. This skill can be difficult to observe when everything goes well. Its economic value becomes visible when a poor diagnosis produces costly rework. As routine solution generation becomes cheaper, problem framing can become a larger share of what differentiates useful expertise.

Diagnosis itself can be assisted. A system can suggest hypotheses, detect patterns or ask structured questions. That can broaden the search. The professional remains responsible for testing the hypotheses against evidence and knowing when the system’s list omits a local possibility. The division of labour can be productive precisely because the human does not need to generate every candidate explanation unaided.

A good training exercise separates symptoms from causes. Give the learner three different mechanisms that produce the same observed delay and ask which additional evidence would distinguish them. The answer should not be a longer list of solutions. It should be a better observation or test. This builds the habit of reducing uncertainty before committing resources.

The market may eventually commoditise some diagnostic patterns too. Standard decision trees and models already do so. That is socially useful when the patterns are reliable. The professional frontier then moves again toward exceptions, integration and responsibility. Expertise is dynamic because successful professional knowledge often becomes tomorrow’s standard procedure.

14. Selection and taste become scarce when everyone can generate many acceptable options

Abundance changes the decision from “Can we produce something?” to “Which of these should we use?” A designer with a hundred generated concepts still needs a criterion for selecting one. A manager with ten draft strategies needs to know which assumptions fit the organisation. A student with several essay structures needs to choose the one supported by the text. Selection is not simply preference; it is judgment against a purpose and constraints.

Taste is sometimes treated as mysterious intuition. Parts of it can be made explicit. A good explanation may be clear, accurate and appropriately detailed for its audience. A useful design may satisfy functional constraints while creating the intended experience. A strong recommendation may make the trade-off visible. Naming criteria makes taste teachable and reviewable without pretending every judgment can be reduced to a formula.

Cheap variation can improve quality by expanding the set of candidates. It can also overwhelm the chooser. If generating twenty alternatives costs almost nothing, the organisation may spend more time comparing them than it once spent creating one. A disciplined process limits generation to the number needed for the decision and defines the criteria before the team becomes attached to a polished option.

A worker can demonstrate selection skill by explaining why an apparently attractive option was rejected. This evidence is especially useful because it reveals the standard being applied. “I chose this because it looked best” is weak. “I rejected the shorter draft because it removed the condition that determines eligibility” shows a relationship between purpose and decision.

The educational consequence is that assignments should sometimes assess comparison and rejection, not only production. Ask students to choose between two plausible explanations and justify the choice from evidence. A world of abundant candidate answers increases the value of being able to distinguish a merely fluent answer from one that satisfies the task.

15. Responsibility can remain scarce even when the technical output is easy to reproduce

Some professional roles exist partly because institutions need a person or organisation to stand behind a decision. A qualified individual may be required to sign, approve or certify particular work. The requirement may reflect training, regulation, liability or governance. A tool can assist the reasoning without automatically receiving that authority. The human role therefore cannot be analysed solely by asking whether the system can generate the same-looking document.

Responsibility also appears outside regulated professions. A manager approving a change may need to consider consequences across teams. A teacher decides whether a student’s response demonstrates the intended learning. A project lead confirms that a handover is ready. The final decision can take seconds after hours of preparation. Its value comes from authority and judgment, not the duration of the click.

This creates a temptation to keep the professional as a ceremonial approver while the system performs the substantive work. That arrangement is fragile if the person lacks time or knowledge to review the output meaningfully. Accountability without review capacity becomes a fiction. Organisations should match the volume of automated production to realistic human oversight or redesign the standard so routine outputs can be trusted through other controls.

Workers should understand what they are being asked to own. If an AI-assisted recommendation is presented under their name, they need access to the evidence and authority to question it. “The system generated it” is not a complete answer when the role requires human approval. Conversely, a worker should not be held responsible for a process they were neither trained nor authorised to control.

The scarce capability here is not abstract responsibility as a character trait. It is the ability and institutional permission to make a defensible decision, explain the evidence and accept the consequences within a legitimate role. That combination can remain valuable even when much of the technical production becomes inexpensive.

16. Coordination matters more when capabilities are cheap but distributed

A team may have easy access to writing, coding, analysis and design tools while still failing to deliver a coherent outcome. Someone needs to decide what each part is for, which version is authoritative and how dependencies fit. Coordination converts separate capabilities into a working system. Its importance can increase when the cost of producing each component falls and the number of components rises.

Imagine a fictional product launch. The marketing team can generate copy quickly, the engineering team can generate interface code and the finance team can model prices. None can independently decide whether a claim is technically accurate, commercially permissible and consistent with the actual product state. A coordination failure can allow three high-quality local outputs to create one misleading public message.

Coordination is not simply scheduling meetings. It includes defining decision rights, resolving conflicting definitions and making handovers explicit. A useful coordinator asks what must be true before the next team acts. The ability to preserve that condition across several specialised systems is a form of professional expertise that often becomes more visible when production itself is easy.

Tools can help coordinate through shared records, reminders and summaries. They can also multiply communication if every system produces more updates. The scarce resource can become attention to the one unresolved dependency that blocks the whole plan. A strong coordinator reduces information while preserving what changes the next decision.

For junior workers, coordination provides a path to deeper understanding because it exposes why local work matters. A junior analyst who sees only their spreadsheet may learn less about the decision it supports. A supervised handover that explains who uses the result and what error would matter can turn routine production into professional judgment. If AI takes the routine task away, organisations may need other ways to expose juniors to the downstream context.

17. The scarcity shift: from producing the answer to deciding what deserves trust

Professional scarcity used to attach to access. Only people with a library, a licence, a codebase, specialist software or years of training could produce certain outputs. Digital tools changed many of those constraints. Generative AI accelerates the change by making language, code and structured analysis available through ordinary interfaces. The scarcity does not necessarily vanish. It can move toward trusted data, accurate context, judgment, accountability and access to the person who controls the next decision.

This shift is visible in ordinary work. A small business can produce a contract-shaped document quickly but still needs to know whether the document is appropriate for its jurisdiction and transaction. A parent can obtain an explanation of a school pathway but still needs the current official requirement. A junior employee can generate a plan but still needs authority and organisational context. Cheap creation does not eliminate the gate where legitimacy enters.

The worker who understands this shift asks a different learning question. Instead of “How can I be the fastest person producing the standard output?” they ask “Which part of the result is hardest to verify, integrate or take responsibility for?” That does not mean production skill is irrelevant. It means production is no longer the only or always the most scarce part of the bundle.

The firm asks a parallel question. Which tasks now have excess capacity, and which remain bottlenecks? If every employee can draft a proposal, review time may constrain output. If search becomes cheap, trusted internal data may become the constraint. If routine coding accelerates, testing and deployment governance may become the limit. Investment should follow the bottleneck rather than the excitement of the new tool.

Scarcity can also move outside the organisation. When many firms can produce similar digital outputs, customer trust, distribution and reputation may become more important. A technically competent product can fail to attract attention. A professional who understands users and relationships may create more value than one who only produces another acceptable document. The market rewards the complete route from capability to use, not the existence of capability in isolation.

Public institutions can shift scarcity deliberately. Education makes literacy and numeracy widespread. Libraries make information accessible. Professional standards make methods legible. These are successes even when they reduce the private scarcity premium attached to knowledge. The worker’s challenge is adapting income and identity to a world where useful capabilities become more common. The social challenge is ensuring that people can move toward the new scarcity without losing access to a decent life in the transition.

18. Novices may become useful faster, but they still need a route to become experienced

The evidence from professional writing and customer support suggests that less-experienced workers can benefit substantially from AI assistance on some tasks. This can lower the cost of reaching competent baseline performance. Organisations may be able to give newer workers meaningful responsibilities earlier. That is potentially valuable for productivity and inclusion. It also raises a developmental question: what experiences will teach the judgment that used to emerge while producing the routine work?

Consider a fictional junior analyst. Before AI, they spent several months preparing simple summaries. A senior corrected the definitions and taught them which anomalies mattered. With AI, the junior can create the summary immediately. If the senior only reviews the final answer silently, the junior may receive less exposure to the reasoning. The task has been accelerated while the apprenticeship has been accidentally removed.

A redesigned apprenticeship makes the reasoning explicit. The junior receives two outputs—one ordinary and one containing a subtle denominator error—and must explain which can be used. Later, they inspect a real permitted case under supervision. They also see what the senior escalates and why. The organisation no longer relies on repetitive drafting to teach judgment. It builds judgment directly.

This can be better than the old system. Routine work was not always a good teacher. Juniors could spend years formatting documents without understanding the final decision. AI creates an opportunity to remove low-value repetition and design more purposeful learning. But this improvement requires attention; it does not happen automatically when the old task disappears.

Firms should therefore distinguish junior productivity from junior development. A new worker can produce acceptable outputs with assistance and still need training before handling an exception alone. Performance metrics should not hide the support carrying the result. A manager who concludes that no apprenticeship is needed because the novice’s output looks professional may create a future shortage of people capable of reviewing difficult cases.

The worker’s responsibility is to use the assistance as an explanation, not merely a shortcut. Ask why the recommendation fits. Compare it with a rejected alternative. Attempt a new case without the decisive prompt where appropriate. Build enough domain knowledge that the tool’s suggestion can be questioned. The goal is not proving independence by refusing technology; it is becoming able to supervise technology rather than being supervised by it.

19. Experts face a different risk: their ordinary output may become abundant while their rare judgment remains invisible

An experienced professional may watch a tool reproduce the visible surface of years of practice. The first reaction can be that the profession has been copied. Yet much expert value can lie in knowing what not to do, which exception matters and when an apparently correct output is unsafe. These contributions are often invisible because they prevent errors rather than create obvious deliverables.

Imagine a senior engineer who reviews twenty routine suggestions and rejects one because it violates an old dependency. The rejected suggestion might look excellent to a non-expert. The senior’s value appears as a decision that produces no new code. If the organisation measures lines generated, the expert looks slow. If it measures avoided failures and reliable delivery, the decision becomes visible. Metric design determines whether the new scarcity is recognised.

Experts should avoid claiming that their intuition is impossible to explain. Some of it can and should be made explicit through examples, checklists and teaching. Sharing expertise strengthens the system and develops successors. The expert’s continuing value is not preserving mystery. It is recognising where the explicit model breaks, improving the model and carrying responsibility for cases that remain difficult.

The organisation should also avoid extracting knowledge without maintaining the conditions that produce it. If experienced workers train a system or codify a process, that work is itself a contribution. Future expertise still requires challenging cases, feedback and time for reflection. A firm that removes every opportunity for deep learning after capturing today’s knowledge may discover that tomorrow’s unusual problem has no one left who understands the underlying system.

An expert’s career response can involve moving closer to the source of standards, difficult cases, integration or teaching. It can also involve becoming better at using tools for routine work while preserving time for high-value judgment. There is no universal need to become a manager. Technical expertise can remain scarce where the domain is deep and the consequences of mistakes are high.

The psychological challenge is real but should not be confused with the analytical claim. A professional can feel that something valuable has been cheapened even when their deeper contribution remains important. The appropriate response is to map the work and evidence rather than dismiss the feeling or make a prediction from it. Identity and labour-market scarcity can change at different speeds.

20. If a skill premium compresses, wages still depend on institutions, demand and who captures the gain

Suppose a routine professional task once required a scarce skill possessed by one in ten workers. A tool lets five in ten perform it acceptably. The supply of the capability has expanded. Other things equal, scarcity pressure on the premium can fall. But wages are not determined by scarcity alone. Demand may expand, the role may gain new responsibilities, bargaining arrangements may change and the organisation may capture or share the productivity gain differently.

Use an invented firm with ten employees producing one hundred reports a week. A new tool lets the team produce one hundred and fifty with the same headcount and quality. Revenue does not necessarily rise 50%. There may be demand for only one hundred and ten. The firm could reduce backlog, lower prices, redeploy staff or leave the extra capacity unused. The labour-market consequence depends on the business decision following the technical capability.

A worker can therefore create more output and see no immediate pay increase. Another can receive higher pay because their review role becomes critical. A third may face fewer available hours because routine work shrinks. These possibilities are not contradictions. They occur at different parts of the value chain. The article does not predict which outcome applies to a particular employer.

Recognition can lag behind the new scarcity. Performance systems may continue rewarding document volume even after verification becomes the bottleneck. Professionals should make the changed contribution legible with appropriate evidence. Managers should review whether incentives still correspond to the work’s purpose. Otherwise employees rationally optimise an obsolete metric while the real risk accumulates elsewhere.

Public policy can affect the distribution of the transition through training support, job redesign and income support, but it cannot freeze every old premium indefinitely. Singapore’s 2026 workforce-transformation measures reflect an attempt to support firms and workers through redesign. The effectiveness of particular programmes needs its own evaluation. Their existence does not establish that every displaced premium is replaced one-for-one elsewhere. [5]

For households, this uncertainty argues for capability and flexibility rather than fear-driven overinvestment. A worker does not need to predict the exact wage effect of AI to keep a record of useful work, understand the changing task bundle and avoid building commitments that assume one premium can never change. This is a risk-management principle, not a forecast that their income will fall.

21. Workflow laboratory: a 70% cheaper draft can create a more expensive final decision

This laboratory is entirely fictional. A professional team prepares ten weekly client briefs. Under Process A, each brief requires sixty minutes of research, sixty minutes of drafting and thirty minutes of review. Total labour is twenty-five hours: ten hours research, ten drafting and five review. For the exercise, labour is valued at S$60 per hour, making S$1,500 of stated internal labour cost. The amounts are not market rates or an estimate of an actual professional service.

Process B uses an AI-assisted workflow. Research preparation falls to thirty minutes and drafting to eighteen minutes per brief, but review rises to sixty minutes because the reviewer checks sources and unsupported claims. Total labour becomes eighteen hours: five research, three drafting and ten review. At the same flat illustrative rate, cost is S$1,080, a 28% reduction. Drafting time alone fell 70%, but total labour did not fall 70% because the bottleneck moved.

Now add a quality event. In Process B, two briefs require an extra ninety-minute senior review because they concern unusual cases. That adds three hours, raising total labour to twenty-one hours and the illustrative cost to S$1,260. The process remains cheaper than A under the stated assumptions, but the gain is smaller. If the team had advertised a 70% cost reduction from draft time alone, it would have measured the wrong unit.

Change the consequence again. Suppose one unsupported claim reaches a client and requires eight hours of corrective work across several staff. Process B’s total rises to twenty-nine hours, now exceeding A. This does not prove AI-assisted drafting is worse. It shows why expected rework and error consequences belong in workflow evaluation. A process can be faster on ordinary cases and less economical when one rare error is sufficiently costly.

The team can respond by narrowing the AI-eligible briefs, improving source constraints or teaching the first reviewer to detect the known error earlier. Suppose those changes reduce ordinary review to forty-five minutes and eliminate the two extra senior reviews in the next comparable period. Total labour becomes fifteen and a half hours. The improvement comes from workflow learning, not from a new model alone. Process design determines how technical capability becomes economic value.

Ask what the laboratory cannot establish. It uses equal hourly rates across roles, fixed demand, ten comparable briefs and invented error events. It does not include software fees, client value, worker learning or different salary structures. The result should therefore remain a teaching example. Its purpose is to show that the percentage reduction in one task is not the percentage reduction in the completed professional service.

22. Pricing laboratory: cheap production can change the business model before it changes the final price

A fictional consultancy sells a standard analytical brief for S$600. Before automation, direct labour costs S$300 and other specified costs S$100, leaving S$200 before omitted overheads and tax. A new tool reduces direct labour to S$180 while other specified costs rise to S$120. The same S$600 price would leave S$300 before omitted items. The firm has gained S$100 of margin under the exercise’s assumptions.

That does not tell us what the firm should do. It might reduce the price to S$500, leaving S$200 of the stated margin. It might keep S$600 and invest the gain in review or faster service. It might increase the scope of the product. Competition might eventually force price down. Demand and strategy determine which path is viable. Technical cost reduction creates options; the market decides which persist.

Now suppose competitors can produce a similar standard brief for S$350. The old S$600 price becomes difficult to defend for ordinary cases. The firm may specialise in complex cases where clients value judgment, or bundle ongoing advice, or exit the standard product. The professional skill that created the commodity-like brief is not useless; it has become less differentiating at that market tier.

A junior employee should understand this economics because their development strategy depends on it. Becoming marginally faster at a commodity-like task may not create a strong career advantage. Learning how to handle exceptions, communicate with clients or improve the system may. The worker does not need to become a salesperson. They need to understand which part of the service customers cannot obtain cheaply elsewhere.

The firm should also recognise social value in lower prices. A service once accessible only to large clients may become available to smaller organisations. That expansion can create new demand and new work. Commoditisation is therefore not synonymous with decline. It can be a diffusion of useful capability whose benefits and adjustment costs are distributed unevenly.

The final question is whether the professional can move with the business model. If the standard brief becomes cheap, what new contribution can they make in the expanded market? Perhaps smaller clients need help interpreting the result. Perhaps specialised cases remain underserved. Perhaps the firm can teach clients to perform routine tasks while charging for complex advice. The answer requires evidence about actual demand, not a universal theory that every professional service must become premium consulting.

23. Preserve the junior learning route when the old beginner task disappears

Entry-level work has historically contained a mixture of useful apprenticeship and low-value repetition. A junior might prepare first drafts, reconcile simple records or perform initial research. Some of those tasks taught the domain by forcing repeated contact with ordinary cases. Others taught mainly patience. When automation removes them, organisations need to identify which developmental function is worth preserving rather than defend every old task simply because seniors once endured it.

Suppose a fictional junior lawyer previously summarised ten cases before discussing one with a senior. The repetition helped them see how similar principles applied under different facts. A tool can now summarise all ten. The organisation could ask the junior to accept the summaries and save time. A stronger learning design asks the junior to verify three, compare where the principle changes and explain one case the model handled badly. The learning goal shifts from compression to discrimination.

In software, a junior may once have written routine functions and learned from code review. If a tool writes the functions, the review can become the teaching surface. Ask the junior to explain tests, edge cases and dependencies. Give them a controlled bug and ask which output should not be trusted. The organisation should not preserve manual typing for its own sake, but it must preserve opportunities to learn the system behind the generated code.

In client service, a new employee may receive suggested responses immediately. Their next learning task can be identifying when a suggestion violates policy or misses the customer’s actual question. A supervisor can expose the reason for a change instead of silently editing the reply. The novice’s productivity is valuable; the novice’s development is a separate organisational asset.

This is especially important when senior judgment is scarce. If every difficult case goes directly to a small group of experts, juniors may never see the reasoning required to become experts themselves. Create graded responsibility. Let newer workers handle bounded cases, observe escalations and gradually take over decisions as evidence supports it. Apprenticeship should become more intentional when experience is no longer produced automatically through routine work.

Performance systems should recognise learning quality as well as output volume. A junior who asks a precise question before acting can be more valuable than one who produces many unchecked drafts. A manager who explains one recurring exception can increase the team’s future capability. These contributions may not appear in conventional productivity metrics. Redesigning work without redesigning evaluation can therefore reward the wrong behaviour.

24. Education should move from producing standard answers toward understanding, comparison and verification

When ordinary professional outputs become cheap, schools face a temptation to react at the surface: either ban the new tool and continue unchanged, or embrace the tool and remove foundational learning. Both can miss the deeper educational question. What must a learner understand to use abundant capability responsibly? The answer varies by subject, but it generally includes concepts, evidence, representation and the ability to recognise when a method does not fit.

Mathematics illustrates the point. A calculator can produce arithmetic cheaply, yet learners still need number sense and representation to choose the operation and recognise an impossible answer. Generative AI extends the same challenge into language and reasoning. A system can produce an essay, but the learner still needs to know whether the evidence supports the claim and whether the response answers the question.

Assessment can adapt by asking for comparison, critique and changed conditions. Give two plausible explanations and ask which is better supported. Provide a generated answer containing one subtle mistake and ask the student to locate it. Change the data and ask whether the earlier conclusion still holds. These tasks make understanding visible without pretending that producing a polished first draft is still scarce.

Foundational knowledge remains important because verification requires a reference frame. A student cannot reliably judge an algebraic transformation without understanding equality. They cannot identify a misleading historical claim without relevant background. Tool access can reduce the memory burden for some details while increasing the value of knowing which details deserve checking.

Schools also need explicit authorship rules. A student should know when AI assistance is permitted, what must be disclosed and which work is expected to be their own. Hidden tool use makes the assessment uninterpretable. Overly broad prohibitions can make ordinary learning diverge sharply from later work. The rules should follow the educational purpose rather than the novelty of the technology.

The connection to opportunity is significant. If affluent learners receive early coaching in how to supervise AI while others are taught only to avoid it, a new capability gap can emerge. Public education can reduce that gap by teaching common principles of verification, responsible use and task decomposition. The goal is not to make every child an AI specialist. It is to make the changed environment understandable enough that access to good judgment is not inherited privately.

Education should also preserve work that has value beyond labour-market scarcity. Reading literature, learning history, understanding science and writing clearly matter for participation and thought, not only employment premiums. A skill becoming cheaper in the market does not erase its human or civic value. Schools can teach both intrinsic understanding and the changed professional context without making one justification replace the other.

25. Credentials become weaker signals when the underlying task changes faster than the certificate

A credential can certify completion, assessed knowledge, professional eligibility or another defined achievement. Its labour-market value depends partly on what employers believe it predicts. If the work changes, the relationship between certificate and performance can change too. This does not make the credential worthless. It means the institution and employer need to understand what evidence the credential still supplies.

Imagine a course whose final assessment requires producing a standard analytical report without AI. A graduate later enters a workplace where the first draft is automated and the scarce task is verification. The course may still have taught useful domain knowledge, but the assessment provides limited direct evidence about the new bottleneck. Updating the curriculum could preserve the foundational analysis while adding evaluation of tool-assisted work.

Professional licences are a different case because they may carry legal or regulatory authority in addition to signalling skill. The technology cannot erase that authority merely by producing a similar-looking output. The profession may still need to reconsider which competencies the licence assesses and how continuing competence is maintained. Formal recognition and technical capability must not be confused.

Short credentials can be useful when they teach a specific new task and are recognised for a defined purpose. They can also become commodity-like themselves if many providers issue similar certificates with unclear assessment. Learners should ask what the credential allows them to do, how competence is assessed and who recognises it. A certificate’s design and institutional connection matter more when obtaining another certificate is easy.

Employers can reduce overreliance on credentials by using appropriately designed work samples, structured interviews and evidence from prior responsibilities, while still respecting formal requirements. This can broaden opportunity for people who have capability through less conventional routes. It can also create burdens if every employer invents an opaque assessment. Shared standards can reduce search costs while preserving room for relevant evidence beyond the certificate.

For students, the sensible position is neither credential worship nor credential dismissal. Verify the requirement for the intended route. Then investigate what the programme teaches and what evidence of capability you will leave with. A qualification can open an institutional door and still need to be complemented by later learning. Treating it as one part of a career system is more robust than treating it as a permanent guarantee.

26. Evidence of capability becomes more important when polished outputs are cheap

When almost anyone can generate a polished document, the document alone becomes weaker evidence of who can perform the underlying work. Portfolios and work samples therefore need context: what was the problem, what did the person decide, what assistance was used and how was the result checked? The explanation does not need to be long. It needs to make authorship and judgment visible.

Consider a fictional data example. A worker presents a dashboard showing improved completion rates. A strong portfolio note explains that the worker discovered two definitions of completion, reconciled them with the responsible teams and rebuilt the comparison on a consistent basis. The chart itself could be generated easily. The professional value lies in identifying the definition problem and establishing a usable measure.

Another worker uses AI to draft a client memo. They can still show capability by explaining which sources were approved, which generated claim they rejected and how the final recommendation changed after checking. The tool use does not invalidate the example. Concealing the tool or claiming independent authorship of the generated reasoning weakens the evidence because the reviewer cannot see what the person can actually do.

Confidentiality remains a boundary. Do not remove private data from an employer merely to make a portfolio more impressive. Use permitted summaries, recreated synthetic examples or explicitly authorised material. A worker who demonstrates judgment by respecting information boundaries is supplying evidence of professionalism even when the artifact is less visually dramatic.

Portfolios should include at least one example of revision after error or feedback. An immediate success story can hide the person’s ability to update. Show what changed when evidence contradicted the first approach. This demonstrates a capability likely to remain useful as tools change: the ability to recognise that the current method is not working and redesign it.

The same principle can be applied in education. A student can annotate a generated explanation, identify an unsupported claim and revise it. The submitted artifact includes the reasoning behind the correction. This is stronger evidence of understanding than a perfect paragraph of uncertain authorship. The assignment’s rules should be explicit so the evidence remains interpretable.

27. Small firms can gain access to capability they could not previously afford—and still need judgment

Commoditisation can benefit small firms by lowering the fixed cost of expertise. A business that could not justify hiring a full-time specialist may use software for bookkeeping support, basic design, translation, analytics or customer communication. This can raise the quality of ordinary operations and let the firm experiment at lower cost. The opportunity is especially important when technology spreads capabilities once concentrated in large organisations.

Singapore’s adoption data show smaller firms lagging larger ones, however. In 2026, MOM reported substantially lower AI adoption among smaller firms than among large firms. This reflects more than subscription price: organisational capability, data, implementation and workforce readiness matter. A cheap tool does not automatically create cheap transformation. [3]

A fictional small firm can illustrate the gap. The owner uses a tool to create a customer FAQ. The draft is good, but it includes a refund condition that does not match the firm’s actual terms. A large firm may have a legal or compliance reviewer; the small firm may not. The same access to generation can therefore produce different risk. Shared guidance, clearer product terms and appropriate professional support can matter more as production becomes democratised.

Small firms should begin with low-consequence tasks where verification is manageable. A tool can help organise internal notes, brainstorm questions or draft material from approved facts. As the consequence rises, the review standard should rise too. This is not a universal rule that small firms must avoid AI in consequential work. It is a risk-based way to decide where scarce professional attention belongs.

The firm can also build its own expertise through use. Keep examples of errors, record which inputs improve reliability and teach staff what should be escalated. The tool becomes part of organisational learning rather than a black box owned by one enthusiastic employee. If that employee leaves, the firm retains a clearer operating model.

Public support for job redesign and AI adoption can help smaller firms develop these organisational capabilities. The value should be evaluated through actual work changes, not adoption counts alone. A firm with one well-designed AI use may be more capable than another with many subscriptions. Depth of integration and quality of work matter alongside access.

28. Shared infrastructure can make professional capability cheaper without making professional judgment unnecessary

Public digital services, open standards, libraries, education and professional guidance all reduce the private cost of obtaining useful knowledge. A well-designed public form can eliminate the need to hire someone merely to understand an administrative step. A searchable official database can reduce dependence on an insider. These are forms of beneficial commoditisation: capability becomes easier to access because the system has made the ordinary route legible.

The same design principle applies to AI-era work. Shared examples of responsible use, sector guidance and accessible training can reduce the cost of learning basic practices. Singapore’s 2026 National AI Impact Programme and workforce-transformation efforts are examples of public support aimed at firms and workers, including capability building and job redesign. Their actual effects require evaluation, but the infrastructure reduces the need for each organisation to invent every practice alone. [9]

Shared infrastructure should focus on functions that many people need: identifying data boundaries, checking sources, understanding tool limitations and mapping changed tasks. It should not pretend that a generic module teaches the local context of every profession. The public layer can establish a common floor while employers and professional bodies teach the specialised layer.

When basic capability becomes widely available, access to advanced judgment may become the new inequality. A well-resourced firm can afford senior reviewers and specialised data; a small organisation may rely on generic output. Public policy cannot supply a private expert to every decision, but it can make standards, escalation routes and trusted information easier to reach. This reduces the degree to which safety depends on private network knowledge.

There is also a distributional question about the knowledge used to build systems. Experienced workers contribute patterns that tools may later disseminate. Institutions should consider how that contribution is recognised and how workers share in the productivity gains. This article does not prescribe a particular bargaining arrangement. It identifies a legitimate question created when private expertise becomes organisational infrastructure.

The social objective is not to preserve artificial scarcity. Making useful expertise cheaper can improve welfare. The challenge is building routes through which workers whose old scarcity premium falls can develop and demonstrate the new scarce capabilities. Education, public guidance, employer learning and income support during transitions can all contribute without guaranteeing identical outcomes.

29. The career response is not to chase every new scarce skill

A scarcity map can become another source of anxiety if workers treat every emerging premium as an instruction to reskill immediately. That would recreate the destination problem in a more frantic form. Today’s scarce skill can become tomorrow’s standard feature. A more durable response begins with the work a person already understands and asks which adjacent capability would let them handle a changed bottleneck.

For one worker, that may be verification. For another, client diagnosis. For another, the ability to combine domain knowledge with an AI-assisted workflow. The right target depends on the role, the organisation and the worker’s goals. A useful development plan should be able to explain why this capability matters in a concrete workflow, not only why commentators describe it as future-proof.

Prefer capabilities that produce evidence quickly. A worker can practise reviewing generated summaries, documenting an exception, comparing two analytical approaches or leading a small handover. The result reveals whether the learning is useful before the person commits to a long programme. This does not mean every important skill can be learned cheaply or quickly. It means uncertainty should be reduced where possible before a large commitment is made.

Keep the household visible. A theoretically valuable skill can be the wrong current project if the timetable is unsustainable or the cost displaces an essential need. The earlier middle-class article explains why career development and household security need to be analysed together. Professional resilience that destroys the rest of life is not a robust outcome.

A worker should also maintain capabilities that are already valuable rather than assuming novelty is always better. Deep subject knowledge, reliable communication and trust can remain useful across tool changes. New AI fluency can complement rather than replace them. A person who understands both the old domain and the new workflow may be more useful than someone who has only learned the newest interface.

The review question is simple: what task has changed, what capability now matters more, and what evidence would show I can perform it? If those three answers remain vague, more research may be more useful than immediate enrolment. Career adaptation should be answerable to observed work rather than a permanent fear that somebody else has already learned the next thing.

30. A 90-day capability audit for work whose ordinary outputs are becoming cheap

This is an original planning framework, not an evaluated career programme. It does not promise employment, promotion or salary change within ninety days. Use it to inspect one changing part of work. An urgent employment, contractual or financial question may require action sooner through the appropriate route.

Days 1–15: map one workflow honestly

List the main tasks in one recurring piece of work and estimate where time currently goes. Mark which outputs a tool can generate, which inputs are sensitive, which decisions require approval and which mistakes would be costly. Do not begin by assigning an automation percentage. Begin by understanding the sequence and the purpose. A map of one real workflow is more actionable than a broad prediction about the occupation.

Choose one task that appears to be becoming commodity-like. Record what counts as acceptable performance and what the ordinary output is used for. If the standard is unclear, clarify it before evaluating a tool. A team cannot know whether a faster draft is useful if nobody agrees what the completed service must preserve.

Days 16–35: test assistance and capture the failure modes

Use approved tools on a small set of representative cases. Include at least one easy case, one ordinary case and one case containing a known complication. Compare time, quality and review effort. Record the important error types rather than only the best output. The purpose is to find the local frontier and the new review burden, not to produce a promotional demonstration.

Ask what the worker had to know to detect the errors. That knowledge becomes a candidate for the new scarce capability. If nobody can verify the output cheaply, the task may not be ready for the proposed automation. If a simple rule catches most errors, the rule can become part of the standard process. The work should become safer as knowledge is captured, not more mysterious.

Days 36–60: redesign the learning route

Identify what juniors and experienced workers now need to learn. Remove repetition that no longer serves a purpose, but replace any lost apprenticeship with structured comparison, review or exception handling. Give workers examples of acceptable and unacceptable outputs. Make escalation routes clear. Allocate time for the learning rather than assuming everyone can absorb the new responsibility outside work.

Each worker should practise one changed decision, not merely attend a presentation. A junior might verify a generated summary. An expert might codify one recurring exception and explain its boundary. A manager might redesign a handover. The task should create evidence of capability and reveal what support remains necessary.

Days 61–90: evaluate the whole completed service

Measure the final outcome rather than the automated step alone. Include rework, queue length, quality, completion time and relevant human effort. Ask whether the new process has shifted work to a hidden bottleneck. Compare like with like and state changes in case mix. A productivity claim should survive the denominator and the workflow it describes.

Then decide what follows. Scale the workflow if the evidence supports it. Narrow its use if only certain cases work well. Invest in verification or context where those have become scarce. Stop an unhelpful use without concluding that every AI use will fail. For the worker, choose the next capability based on the observed bottleneck. The 90-day audit succeeds when it makes the system easier to understand, not when it guarantees a dramatic transformation.

31. Three workers show why the same technology can produce different career effects

The following cases are fictional. They do not describe typical Singapore workers or predict earnings. Each person uses the same general class of AI assistance, but their roles, organisations and starting knowledge differ. The examples show why a single technology label is too coarse to determine who benefits or which skill becomes scarce next.

Alicia’s mother: routine drafting becomes cheap, but trusted review becomes visible

She works in a professional services team where standard client summaries once consumed much of the week. An approved tool now creates first drafts. Her drafting time falls sharply. At first, management interprets this as simple spare capacity. She begins recording the kinds of errors that require correction and finds that one recurring definition creates disproportionate risk. She proposes a shared check and trains two colleagues to use it.

Her visible output shifts from writing every paragraph to improving the review process and handling unusual cases. The tool has commoditised part of her old production work without making her experience irrelevant. Whether the firm changes her pay or title is not specified. The career gain in the story is a clearer, demonstrable responsibility that sits closer to the new bottleneck.

Tricia’s father: the tool raises output but the business does not need more of it

His team produces routine internal analyses. AI makes them faster, but demand for the analyses is fixed. The firm does not need twice as many reports. It reduces time spent on the task and reallocates part of the team to another function. Tricia’s father chooses to learn the data-definition work used in the new function because it connects with experience he already has.

The transition is not automatic. He spends time learning unfamiliar systems and initially produces slower work. His prior knowledge helps him recognise inconsistent measures, while a colleague teaches the new workflow. The story demonstrates substitution in one task and complementarity in another. It does not claim that every worker can be redeployed or that every firm will choose redeployment rather than headcount reduction.

Kai Kai’s future workplace: cheap code changes the first rung of the ladder

In this future-looking teaching case, a junior developer enters a team where code generation is ordinary. The junior can produce routine functions quickly. The team deliberately requires them to explain tests, dependencies and one failure case before merging. They also rotate through incident review under supervision. The apprenticeship does not depend on manually typing every line that previous generations typed.

The changed entry-level role may demand judgment earlier, which can benefit learners who receive good support and disadvantage those expected to supervise tools without strong foundations. The education and employer response is to make those foundations and review practices explicit. The point is not to preserve an old job description. It is to preserve a credible route from novice performance to professional competence.

32. The system question: can society enjoy cheap expertise without making expertise careers disposable?

Cheap capability can be a social gain. Small businesses can access better tools. Students can receive explanations. Routine services can become faster. Knowledge can spread beyond elite networks. Preserving artificial scarcity merely to protect a professional premium would deny these benefits. The policy problem is not how to keep ordinary expertise expensive forever.

The adjustment problem is how people move when the scarcity premium shifts. Workers need time and routes to develop the next capability. Firms need incentives and support to redesign jobs rather than simply bolt tools onto old workflows. Education needs to teach verification and judgment without abandoning foundations. Income and employment support can matter when transitions are not immediate. These are institutional questions because the worker cannot control all the conditions alone.

Singapore’s 2026 policy direction around AI, job redesign and skills reflects this broader challenge. The Tripartite Jobs Council, Enterprise Workforce Transformation Package and National AI Impact Programme are examples of coordinated responses. Their existence does not prove success; programme outcomes need evidence. They show that the transition is being treated as a workforce and organisational problem, not only a matter of individual tool adoption. [5] [9]

A fair system should also protect standards where errors matter. Making advice cheap is useful only if people can distinguish generic assistance from professional authority. Public information can clarify when a qualified person is required and when ordinary self-service is appropriate. This lets technology reduce unnecessary cost without allowing a fluent interface to impersonate every kind of legitimate professional judgment.

The deepest educational response is to make the new scarcity teachable. Problem framing, evidence evaluation, comparison, responsibility and coordination are not mystical traits reserved for senior people. They can be practised through appropriately designed tasks. Experience remains important because judgment develops across varied cases, but institutions can make the route more explicit and less dependent on private mentorship.

The goal is not a labour market in which nobody’s skill ever becomes common. That would require stopping education and technology from spreading useful knowledge. The goal is a labour market in which wider capability improves society while workers have realistic ways to keep contributing as the location of scarcity changes.

Questions readers ask when expertise becomes cheap

Does this mean professional salaries will fall?

Not as a universal conclusion. A task becoming cheaper can reduce a skill premium, expand demand, shift workers into complementary tasks or change the organisation’s margins. Wages also depend on bargaining, institutions and local demand. The article identifies mechanisms and evidence about task performance; it does not forecast a particular occupation’s salary path.

Which skills are safest from AI?

No permanent list is defensible because tool capabilities and work design change. Skills involving difficult verification, local context, responsibility, coordination and unusual cases may remain scarce longer in many settings, but parts of them can also be standardised. A better question is which task in your actual workflow remains difficult to reproduce or verify cheaply.

Should students stop learning writing or coding because AI can generate them?

No. Learners still need the underlying concepts, language and representations required to inspect, adapt and use generated work responsibly. The curriculum can change what it emphasises and how it assesses understanding. Tool availability changes the educational task; it does not erase the need for knowledge.

Is prompting a durable professional skill?

Prompting can be useful, but interfaces and model behaviour can change. The more durable capability is expressing the task clearly, supplying relevant context, recognising a bad answer and verifying the result. Those actions can survive changes in the exact command syntax.

What should an expert do if juniors can now produce similar-looking work?

Identify where your experience changes the outcome: diagnosing the problem, handling exceptions, verifying evidence, integrating domains, teaching others or accepting responsibility. Make those contributions visible and help redesign the apprenticeship so juniors can develop them too. Preserving mystery is a weaker strategy than demonstrating judgment.

Can commoditisation be good for society?

Yes. Lower-cost access to useful capability can benefit households and firms. The challenge is quality, legitimate authority and the distribution of adjustment costs. A good transition captures the access benefit while giving workers routes toward the new scarce tasks.

Expertise survives by moving closer to the part of the problem that remains difficult

A professional skill becomes commodity-like when an ordinary acceptable version can be obtained from many sources at low cost. AI can accelerate that process by making drafts, explanations, analyses and code easier to produce. This is a genuine economic change. It is not the same as saying that every profession disappears.

The evidence shows why the distinction matters. Professional writing and customer-support studies found meaningful productivity gains and larger benefits for some lower-performing or less-experienced workers. Consulting research found strong gains inside an AI capability frontier and worse performance on a task outside it. Singapore’s current firm data show uneven adoption, reported productivity gains, job redesign and limited reported headcount reduction so far. The picture is transformation with substantial uncertainty, not one completed outcome.

As production becomes cheap, value can move toward problem definition, verification, context, selection, responsibility and coordination. Those activities can also be assisted and partly standardised. The frontier continues to move. The useful career strategy is therefore not to identify one permanently safe human trait. It is to understand the workflow deeply enough to see where trustworthy contribution is becoming scarce next.

For education, the answer is to preserve foundations while making judgment visible. For employers, it is to redesign apprenticeship and metrics along with tools. For public institutions, it is to make basic capability widely available while supporting workers through changing roles. For professionals, it is to let cheaper expertise spread without confusing the spread of a capability with the end of a meaningful career.

When expertise becomes cheap, the professional advantage is no longer merely possessing the ordinary answer. It is knowing which answer matters, why it can be trusted, where it fails and what should happen next.

Continue the Singapore capability series

Article 1 and the Singapore capability-series roadmap · Previous: Article 21 — How the AI Capability Divide Works · How X Works: Singapore series.

Article 22 examines the commoditisation of professional skills. The next planned article is How Human Scarcity Works | Which Capabilities Become More Valuable When Intelligence Becomes Abundant. That continuation is planned here, not represented as already published.

Sources, dates and evidence boundaries

The sources below support the attributed research findings and Singapore policy or labour-market descriptions. The original workflows, prices, worker stories, 90-day audit and teaching tasks are illustrations designed to make mechanisms inspectable. They do not estimate prevalence, predict wages or certify a particular provider, employer or technology.

[1] International Labour Organization, 20 May 2025. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. Global task-level exposure framework. Exposure is not equivalent to job loss, and the estimates are not Singapore-specific employment forecasts.

[2] Ministry of Manpower, 30 April 2026. Inaugural Release of Report on Adoption of Artificial Intelligence Among Firms. Private-sector establishments with at least ten employees; reported adoption, productivity and workforce outcomes retain that scope.

[3] Ministry of Manpower, 4 August 2026. Written Answer on AI-Adopting Firms by Firm Size and Outcomes. Descriptive adoption and workforce outcomes by firm size; not a causal effect of size.

[4] Ministry of Manpower, 9 September 2026. Oral Answer on Employment Pathways Amid AI and Automation. Current policy description and survey interpretation, not a guarantee of transition outcomes for individual workers.

[5] Ministry of Manpower, 3 March and 30 April 2026. Enterprise Workforce Transformation Package overview and Tripartite Jobs Council announcement. Programme availability, funding and outcomes require their own current conditions and evaluation.

[6] Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. Preregistered professional-writing experiment; findings should not be generalised to every professional task.

[7] Dell’Acqua, F., and colleagues (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science. The experiment involved 758 BCG consultants and studied specific tasks using GPT-4.

[8] Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. Study of 5,172 customer-support agents; the published 15% average productivity estimate is specific to the studied deployment.

[9] Infocomm Media Development Authority, 2 March and 8 May 2026. National AI Impact Programme and AI upskilling for tech professionals. Announced programme ambitions are not evidence that every participant or firm achieves a specified outcome.

Return to the quick route · Use the workflow laboratory · Use the 90-day capability audit · Return to the Singapore capability-series roadmap.

Teaching Guide: find the human decision hidden inside the cheap output

This original guide is for older students, adult learners and facilitated professional-learning discussions. It uses fictional work so participants do not need to share confidential employer material, private salaries or real client records. The objective is not to prove that AI should or should not be used. It is to identify which part of a workflow has become easier, which part remains consequential and what evidence would justify changing the work.

Use three passes through each task. First, let the learner make an independent attempt appropriate to their level. Second, discuss the first consequential misunderstanding rather than rewrite the whole response. Third, change a relevant condition and ask for a fresh decision. The contrast makes the learner’s reasoning visible. A calculator or approved tool can be used where the target is interpretation, but the participant should still explain what the number or output represents and why it matters.

Module 1: the polished answer with the wrong denominator

Provide this synthetic dataset: Period A has 12 defects among 200 items; Period B has 15 defects among 500. A generated summary says, “Quality worsened because defects rose from twelve to fifteen.” Ask participants what is true and what is missing. The defect count rose. The defect rate fell from 6% to 3%. Neither measure is automatically the right one for every decision. Repair the summary by naming the quantity and the purpose for which it is being used.

Ask which professional skill became cheap. The arithmetic can be automated. The scarce action is recognising which denominator the decision requires. Then give a second dataset in which both count and rate worsen, so learners cannot memorise that rates always reverse the conclusion. The task is choosing the measure and defending the choice, not performing one familiar calculation repeatedly.

Extend the task by changing the unit. Suppose the first period records defects per item and the second records complaints per customer. The two rates cannot simply be placed on the same line as though their denominators were equivalent. Participants should identify what additional information is needed before a comparison becomes meaningful. This makes a general principle visible: professional fluency is weak when the measurement itself has changed underneath the polished explanation.

Module 2: a generated client reply that removes one condition

Use this fictional source: “Customers may request a replacement within fourteen days if the item is unused and the original seal remains intact. Requests are reviewed after the returned item is received.” A generated reply says, “You are entitled to a replacement within fourteen days.” Ask participants which conditions have been removed and which stage has been turned into a guarantee.

A better reply preserves the distinction between eligibility to request, review of the returned item and the eventual outcome. The exercise is not about making the response longer or more defensive. It is about retaining the conditions that determine what the business can honestly promise. The participant should be able to point to each phrase in the source that supports the final wording.

For transfer, change the source so that replacement really is automatic under a different stated condition. The learner must update rather than carry over the previous caution mechanically. A professional who has learned only to hedge every claim has not learned the deeper skill. The target is matching the strength of the message to the actual rule.

Module 3: the workflow with a hidden review queue

A fictional team previously drafts twelve reports a day and reviews twelve. A new tool raises drafting capacity to thirty while review remains twelve. Demand rises to twenty reports daily. Ask participants to calculate the daily queue if every draft reaches review: eight reports accumulate per day. After five identical days, forty await review. The tool improved upstream capacity while the completed-service bottleneck remained downstream.

Ask for three different responses: reduce unnecessary drafting, increase suitable review capacity, or improve draft quality so review becomes faster. Participants should state what evidence each response would need. “Use more AI” is not a complete plan because the constraint is no longer generation. “Hire more reviewers” is also incomplete until the team knows whether review demand is structural or being created by preventable errors.

Now add a case-mix change. Suppose four of the twenty daily reports are complex and each requires twice the ordinary review time. A count of reports no longer describes the review burden adequately. Ask participants how they would represent workload without pretending that all cases are interchangeable. The lesson is to follow the actual scarce resource rather than optimise the most visible stage.

Module 4: the junior who looks expert before understanding the exception

A fictional junior uses an approved assistant to produce an accurate ordinary summary. A second case contains a changed definition that the assistant misses. The junior accepts it. Ask what learning task would help: another prompt-writing lesson, a comparison of definitions, or supervised review of changed cases? Participants should justify the choice from the observed error rather than from a generic belief about AI skills.

Then provide a third case where the definition is ordinary but the source is outdated. The learner must inspect source currency rather than definitions. The exercise demonstrates that verification is not one memorised checklist. Different failures require different evidence and different questions. A strong novice is not the person who never asks for help; it is the person who increasingly recognises which uncertainty requires help.

Ask the group to design one graded-responsibility step. The junior might handle ordinary cases independently, discuss changed definitions with a reviewer and observe one high-consequence exception. The task should develop judgment rather than leave the senior silently correcting everything. The group should also specify what evidence would justify giving the junior greater responsibility later.

Module 5: price falls while value expands

An invented service once costs S$400 and serves 100 customers. Automation lowers cost enough for the provider to charge S$250, and demand rises to 180 customers. Revenue changes from S$40,000 to S$45,000 under these simple assumptions. The unit price fell while total revenue rose. Ask participants why this does not prove the provider is more profitable: costs, capacity, quality and demand conditions remain incomplete.

Now change demand to 120 customers. Revenue becomes S$30,000. The same price reduction produces a different business outcome. Commoditisation can expand access without guaranteeing the old provider’s income. The learner should distinguish social value from the distribution of gains and avoid assuming that a lower price must either benefit or harm every participant in the same way.

Add a verification cost of S$40 per customer in both periods. Ask participants to compute the added review burden and explain whether the business model still looks attractive. The purpose is not to discover one correct pricing strategy. It is to show that a cheap production stage can coexist with another costly stage and that the final service price depends on the whole workflow.

Module 6: map a capability whose scarcity has moved

Give participants a harmless fictional occupation and five tasks: collect information, create a first draft, verify it, decide between alternatives and communicate the decision. Ask which stages a general-purpose tool can assist, which stages need local context and which carry the largest consequence if wrong. Then ask what training a novice would need if first-draft production became almost automatic.

The learner should avoid labelling every human task as safe and every generated task as automatable. They should state conditions. Verification may itself be automated partly. A decision may become standardised. A generated draft may still require substantial expertise when the source material is difficult. The map is useful because it can change when evidence changes.

For a final variation, lower the cost of verification as well. Ask what becomes scarce next. Perhaps trusted data, customer attention, coordination or institutional authority. The exercise demonstrates the article’s central mechanism: technology can move scarcity several times, so career strategy should follow the bottleneck rather than defend a permanent list of uniquely human tasks.

Facilitator review: assess the changed decision, not the number of AI terms used

A strong answer identifies the task that became cheaper, the new bottleneck and the evidence needed before acting. It preserves the limits of research and does not turn exposure into replacement. It can recommend automation, human review, a narrower scope or no change, provided the recommendation follows from the supplied case and names what would cause it to be revised.

Do not reward participants for using fashionable terms such as augmentation, agentic or future-proof without explaining the mechanism. Ask what the tool actually does and what the person still decides. Ask what happens when the output is wrong. Ask which institution has authority. These questions keep the discussion grounded in work rather than technological theatre.

Do not ask participants to reveal confidential work or household information. A synthetic example is sufficient to assess the reasoning. If somebody introduces a real employment, legal or financial concern, recognise the boundary and direct the person toward the appropriate institutional or professional route rather than turning the classroom into an improvised advisory service.

For a final independent task, ask learners to map one harmless, non-confidential activity from school or work into five stages: input, first production, verification, decision and consequence. They may propose where a tool belongs, but they must also say how success would be checked. The Teaching Guide has done its job when the learner can see that cheap expertise is not the end of expertise; it is a change in where professional judgment must enter the system.