Job Redesign and AI | Tasks, Roles & Human Judgement

eduKateSG · Career & Adulthood · Job Redesign and AI · 29 September 2026

The easiest question about AI and work is whether a job will disappear. It is also one of the least useful starting points. Jobs are bundles of tasks, relationships, responsibilities, tools, decisions and institutional obligations. AI rarely enters all of them at once. It changes some tasks, creates new controls, shifts handoffs and alters what counts as valuable human judgement. The real design problem is not job versus machine. It is how work should be reorganised when machines can do more.

This guide owns the job-redesign mechanism inside eduKateSG’s Career & Adulthood Hub. It connects to How AI Changes Singapore’s Labour Market, How the AI Capability Divide Works and How Skills-First Hiring Works. It does not predict which occupations will win or lose. It provides a method for examining tasks, controls, capability and uncertainty.

Singapore’s current policy environment makes the issue concrete. IMDA reported in its 2025 Digital Economy release that 73.8% of surveyed workers used AI tools at work and that 63% of AI-adopting firms expected to redesign jobs to better integrate AI. In 2026, the National AI Impact Programme expanded the focus on AI-fluent workers and role-specific workflow transformation. SkillsFuture also lists Job Redesign+ support for workforce transformation. These are current signals, not proof of one inevitable labour-market outcome.

Start with tasks, not job titles

Start with tasks, not job titles is mainly about breaking a role into work units before asking whether AI replaces or augments the job. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is separating research, drafting, checking, client explanation and final sign-off inside one professional role. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is saying a whole occupation is automated because one visible task can be generated quickly. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the analysis identifies which tasks change, which remain human-accountable and which new tasks appear. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

The human-capability question follows immediately. What must a worker know now that AI is present? Sometimes the answer is less manual production and more diagnosis, source checking, system understanding, client explanation, exception handling or responsibility for final sign-off. Training should follow the redesigned task. Generic AI fluency is useful as a foundation, but role competence is demonstrated only when the person can use the tool inside domain constraints and recognise when not to trust it.

Automation and assistance are different

Automation and assistance are different is mainly about distinguishing tasks handed to the system from tasks where the system proposes and a human decides. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is using AI to draft a first analysis while a professional verifies sources, assumptions and consequences. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is calling any AI use automation. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the workflow makes the locus of judgement and responsibility explicit. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Job redesign is workflow redesign

Job redesign is workflow redesign is mainly about changing sequence, handoffs, controls and role boundaries when new tools enter the process. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is moving routine document classification to AI while giving staff more time for exceptions and client judgement. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is adding an AI tool on top of the old workflow without removing duplicated work. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the redesigned process is simpler, safer or more capable rather than merely more technologically crowded. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

The baseline must be measured first

The baseline must be measured first is mainly about understanding current time, quality, error and handoff patterns before claiming improvement. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is recording how long a process takes and where rework occurs before an AI pilot. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is declaring productivity gains from impressions after a new tool arrives. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that there is a comparable before-state and an agreed success measure. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

AI can speed the wrong process

AI can speed the wrong process is mainly about recognising that efficiency is not valuable when the underlying task is unnecessary or poorly designed. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is automating a report that nobody uses and then questioning why the report exists. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is optimising output volume without examining purpose. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the workflow starts from the decision or service that matters. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Human judgement is a designed control

Human judgement is a designed control is mainly about deciding where review is mandatory because consequence, uncertainty or accountability is high. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is requiring human approval before a model-generated legal, financial or customer decision becomes operational. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is using a vague human-in-the-loop label without specifying what the human must check. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the reviewer has information, authority, time and criteria to intervene. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Verification is a separate task

Verification is a separate task is mainly about making checking visible rather than assuming the user will notice errors while consuming output. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is requiring source checks, calculations, policy validation or physical inspection after AI-assisted generation. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is treating generated fluency as self-verifying. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that verification has an owner and an evidence trail. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Exception handling

Exception handling is mainly about designing for cases the model or automated flow cannot safely resolve. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is routing low-confidence, novel or high-risk cases to a capable person. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is forcing every case through the automated path to maximise adoption. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the system recognises when normal processing should stop. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Professional accountability

Professional accountability is mainly about preserving responsibility even when a tool performs substantial cognitive work. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is a lawyer, accountant, engineer or manager remaining accountable for advice or decisions within professional scope. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is blaming the model after an unverified output causes harm. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the accountable person can explain the basis of the final decision. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Data quality becomes job quality

Data quality becomes job quality is mainly about recognising that AI performance depends partly on records, labels, access and context. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is improving document structure and data governance before expecting reliable automation. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is treating model quality as independent of organisational information quality. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the workflow has trustworthy inputs and known provenance. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Prompting is not the whole skill

Prompting is not the whole skill is mainly about seeing AI use as problem framing, context selection, verification, iteration and integration rather than clever wording. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is a domain expert deciding what information the model needs and how the result will be checked. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is reducing AI capability to memorised prompt templates. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the worker can adapt when the tool, task or interface changes. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Domain knowledge becomes more important in some workflows

Domain knowledge becomes more important in some workflows is mainly about using expertise to notice subtle errors that fluent systems make plausible. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is an accountant recognising that a generated treatment conflicts with current standards or context. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is assuming AI removes the need to understand the domain because it can produce an answer. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the worker can independently judge whether the output makes sense. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Junior work changes first in some roles

Junior work changes first in some roles is mainly about examining whether tasks traditionally used to train beginners are being compressed, automated or moved. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is a junior analyst receiving fewer first-draft tasks and needing new supervised ways to learn reasoning. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is removing entry-level tasks without replacing their developmental function. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the redesigned role includes a route for novices to acquire judgement. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Senior work also changes

Senior work also changes is mainly about recognising that managers and specialists must learn to supervise AI-enabled systems, not only delegate to people. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is reviewing exceptions, setting policy, monitoring quality and redesigning work around new capabilities. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is treating AI adoption as a junior productivity programme. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that leaders understand enough to govern the changed workflow. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Team boundaries shift

Team boundaries shift is mainly about redistributing tasks across operations, technology, risk, legal, domain and customer functions. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is creating a shared process where technical teams do not own business judgement and business teams do not ignore system constraints. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is assuming the AI team can redesign work alone. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that roles and escalation paths are clear across disciplines. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Training follows the redesigned task

Training follows the redesigned task is mainly about teaching people what they now need to do rather than offering generic AI awareness alone. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is training staff on verification, exception handling and domain-specific tool use after the workflow is defined. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is sending everyone to the same introductory AI course and calling the workforce transformed. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that learning objectives match the changed work. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

The skill gap can move upward

The skill gap can move upward is mainly about recognising that automation of routine work may increase demand for framing, interpretation and decision quality. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is freeing time from repetitive reporting but requiring staff to explain anomalies and recommend action. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is assuming less routine work automatically makes the job easier. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the new role has explicit capability requirements and practice. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Workload can increase during transition

Workload can increase during transition is mainly about accounting for parallel systems, training, debugging and uncertainty while the new workflow stabilises. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is running a bounded pilot instead of expecting immediate full-scale savings. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is counting future productivity while ignoring implementation labour. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that transition cost is included in the decision. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Productivity must include quality

Productivity must include quality is mainly about measuring more than throughput. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is tracking turnaround alongside error, rework, customer outcome and staff judgement. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is claiming success because more documents were produced. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the metric set reflects the purpose of the work. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Creativity can be widened or flattened

Creativity can be widened or flattened is mainly about using AI to generate alternatives while preserving selection, taste and purpose. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is asking for multiple options then having a skilled person evaluate which actually solves the brief. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is accepting the first polished option and converging on generic output. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that creative work becomes more exploratory without surrendering judgement. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Customer service redesign

Customer service redesign is mainly about using automation for routine requests while protecting access to human resolution for ambiguity and vulnerability. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is answering common status questions automatically but escalating unusual cases with context intact. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is hiding human access behind an endless automated loop. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the customer can reach appropriate judgement when the case exceeds automation. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Compliance and governance work

Compliance and governance work is mainly about using AI to scan, classify or draft while preserving current-rule verification and accountability. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is flagging documents for review without allowing the model to determine legal compliance alone. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is turning regulation into a keyword-matching exercise. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the final compliance judgement traces to current authoritative requirements. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Knowledge work and source reliability

Knowledge work and source reliability is mainly about separating synthesis from evidence. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is using AI to generate a research outline, then retrieving and reading the primary sources before making a consequential claim. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is citing sources the model invented or never checked. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that every important claim can be traced to evidence the worker actually inspected. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

AI in software work

AI in software work is mainly about using generation, testing and agentic tools while maintaining architecture, security and review. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is automating repetitive code generation while engineers own requirements, system boundaries and validation. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is equating code output with finished software. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the system works safely under real conditions and can be maintained. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

AI in accounting and legal workflows

AI in accounting and legal workflows is mainly about using domain-specific AI to transform routine work while preserving professional judgement, confidentiality and standards. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is automating parts of document review or reporting while professionals focus on risk, interpretation and advice. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is feeding protected information into unapproved tools or accepting outputs without standard checks. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the workflow respects professional and data obligations. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

AI in education work

AI in education work is mainly about using AI to support planning, explanation or feedback without outsourcing assessment of student understanding blindly. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is drafting examples and then checking them against the syllabus, learner needs and subject accuracy. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is equating polished feedback with valid diagnosis. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the teacher still knows what evidence supports the learning decision. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Job enrichment is not automatic

Job enrichment is not automatic is mainly about recognising that removing routine work can create better jobs or simply intensify the remaining workload. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is redesigning roles so saved time becomes analysis, learning or client work rather than more volume. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is assuming automation automatically makes work more meaningful. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the worker has more valuable responsibility without unsustainable load. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Job loss and task loss are different questions

Job loss and task loss are different questions is mainly about separating disappearing activities from net employment outcomes. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is noting that a task can shrink while the role expands in another direction. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is converting task automation into a confident employment prediction. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that claims stay bounded to the evidence available. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Skills-first hiring under AI

Skills-first hiring under AI is mainly about using work samples and current capability to judge readiness for redesigned roles. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is testing a candidate’s ability to verify AI-assisted work rather than only asking whether they used a named tool. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is screening by buzzword lists that age quickly. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that selection examines underlying judgement and adaptable capability. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Organisational learning

Organisational learning is mainly about using incidents and near misses to update AI workflows, controls and training. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is reviewing why an AI-assisted error escaped rather than blaming the last user. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is treating every mistake as individual misuse. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that the system learns and changes after evidence. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Stop rules and rollback

Stop rules and rollback is mainly about deciding in advance when an AI workflow should be paused, narrowed or reverted. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is setting thresholds for unacceptable error, privacy risk or customer harm. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is keeping automation live because reversal would look like failure. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that governance includes the ability to reduce or remove a tool. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

Job redesign as continuous work

Job redesign as continuous work is mainly about accepting that models, policies, markets and user behaviour change after launch. The redesign team should write the workflow down before debating technology. What enters the process? Which decisions occur? Who currently checks them? What information is used? Where does work wait? Where does it return for rework? Which failure would be merely inconvenient and which would be consequential? AI changes the design only after the work is visible enough to reason about.

A practical example is reviewing the task map regularly instead of treating implementation as complete. Notice that the example does not ask whether AI is good or bad in the abstract. It asks which part of the work changes, what control moves, what capability the human still needs and how the result will be checked. This task-level framing is more useful than a headline prediction because organisations can actually redesign tasks, training and accountability.

The common failure is freezing the workflow around one generation of technology. This usually creates hidden work. People copy outputs into old templates, verify manually after the fact, maintain two systems, correct model errors or spend time explaining a process that was never simplified. The organisation may see a new tool and old costs at the same time. Good redesign removes or changes work deliberately rather than layering technology on top of every previous step.

The acceptance test is that roles continue to evolve through measured revision. Evidence should include quality as well as speed. Where possible, compare a baseline with the new process: turnaround time, error, rework, exception rate, customer outcome, staff learning, escalation and risk. Avoid universal productivity claims from a narrow pilot. A workflow can improve in one task and deteriorate elsewhere if the saved time creates new bottlenecks or if verification becomes harder.

A six-step job redesign method

Map the current role into tasks, decisions, handoffs and controls.
Identify where AI can assist, automate or create new capability without assuming adoption is automatically beneficial.
Define the human judgement, verification and escalation points that remain necessary.
Pilot on bounded work with measurable baseline and outcome criteria.
Train workers for the redesigned tasks, including failure modes and recovery.
Review the whole workflow after implementation and remove, narrow or expand AI use based on evidence.

A worker-level audit

An individual can use the same method without waiting for an organisational transformation programme. List the tasks in a normal week. Mark which are repetitive production, which require judgement, which depend on trust, which involve confidential information, which are regulated and which teach valuable capability. Then experiment with AI only where the task and boundaries are understood.

Record what actually changed. Did the tool save time? Did verification consume the saving? Did quality improve? Did the worker learn more or merely produce more? Did the task become easier but the final decision become harder? This personal evidence is more useful than broad anxiety or broad enthusiasm.

Current Singapore signals and the boundary of inference

IMDA’s 2025 Digital Economy release reported rapid workplace AI use and substantial employer interest in job redesign. Its 2026 National AI Impact Programme aims to support AI-bilingual workers who combine domain expertise with practical AI capability. In May 2026, IMDA also announced an industry workgroup with WSG and sector partners to study how tech teams, jobs and skills are changing. These are meaningful signals of active transformation.

They are not proof that every role will change in the same way or that adoption guarantees productivity. Survey measures, sectors and implementation conditions matter. The correct conclusion is narrower: many organisations are redesigning work, and workers increasingly need enough domain and AI fluency to participate intelligently in that redesign.

Questions for leaders before approving an AI workflow

What decision or service are we improving?
Which current task is being removed, assisted or changed?
What evidence describes the baseline?
What is the highest-consequence failure mode?
Which human has authority and time to verify or stop the process?
What data enters the system, and is that use permitted?
What will happen to junior learning if routine tasks disappear?
Which skills become more important after redesign?
What metrics will show quality, not only speed?
What threshold would trigger rollback or narrower use?

Frequently asked questions

Will AI replace my job?

A whole-job answer is usually too coarse. Map the tasks in the role. Some may be automated, some assisted, some unchanged and some newly created. Employment outcomes also depend on demand, regulation, organisation design and whether saved capacity creates new work. Treat predictions as scenarios rather than certainties.

What skills become more valuable with AI?

The answer depends on the role, but domain knowledge, problem framing, verification, exception handling, communication, system understanding and accountable judgement often become more visible when routine production is easier. Avoid universal skill rankings without role context.

Is prompting the main AI skill?

No. Prompting can matter, but durable capability includes understanding the task, selecting context, checking sources, evaluating outputs, protecting data, integrating results into workflow and knowing when to escalate or stop.

How should companies measure AI productivity?

Use a baseline and measure the purpose of the work. Include quality, error, rework, turnaround, customer outcome, risk and staff capability where relevant, not only volume. A faster process that creates more correction work may not be an improvement.

What is Job Redesign+ in Singapore?

SkillsFuture currently lists Job Redesign+ under employer support for workforce transformation, including consultancy, capability building and workforce technology components. Use the live SkillsFuture employer page for current details.

What current Singapore AI workforce initiatives are relevant?

IMDA’s National AI Impact Programme and related TeSA initiatives are current examples. Programme details can change; use the live official pages for present conditions.

The quiet conclusion: redesign the work before judging the future

AI makes some outputs cheaper and some workflows faster. That does not by itself tell us what a good job should become. The design question remains human and organisational: which tasks matter, which decisions need accountable judgement, where verification belongs, how novices learn, how quality is measured and what the saved capacity is used for.

The strongest organisations will not be the ones that simply use the most AI. They will be the ones that understand their work well enough to redesign it without losing the capabilities that make the work trustworthy. The strongest workers will not be those who compete with a model at producing the first draft fastest. They will be those who can frame the right problem, use assistance intelligently, verify what matters and take responsibility for the result.

For the broader AI-to-human-capability map behind this question, see the Super Intelligence master guide, which connects verification, learning, work, judgment and system design.

Job Redesign and AI | How Tasks, Roles and Human Judgement Change When Machines Can Do More