VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

How to Start Using Super Intelligence Without Redesigning Your Entire Company | A Practical Workplace SI Pilot

Three students work together at a classroom table, reviewing open textbooks and notes.

You do not need to redesign your entire company to start using Super Intelligence at work. The strongest first step is usually much smaller: choose one repeated, information-heavy task that people already understand, keep the existing workflow visible, add SI to one bounded part of it, measure the result and expand only after the new method proves useful.

In the eduKateSG workplace series, Super Intelligence is the practical machine-intelligence layer commonly described as artificial intelligence, generative AI, assistants, copilots, agents and connected automation. This article owns the incremental adoption path: how to begin in a real company, school, professional practice or small business without launching a giant transformation programme before anyone knows which workflows actually benefit.

The complete architecture sits in the Super Intelligence Workplace and Workflow Hub. Before this page, the series established task fit, human–SI collaboration, workflow position, maturity levels, role types and delegation boundaries. Here, the question becomes practical: what can an organisation do on Monday morning?


The Short Answer

Start with a thin slice of work. Pick one task that happens frequently, has visible inputs and outputs, can be checked by someone who already understands it, and has low-to-moderate consequences if the first experiment is wrong. Use SI in Assist or Collaborate mode before automating anything. Keep the current process as the fallback. Record time, quality, corrections and exceptions. Standardise only what repeatedly works.

This approach avoids two expensive mistakes. The first is waiting for a perfect company-wide strategy before anyone learns. The second is automating everything quickly and discovering that the organisation never understood the work it was trying to automate. Incremental SI adoption creates evidence before architecture.

Why Whole-Company Redesign Is Usually the Wrong Starting Point

A company is not one workflow. It is a network of jobs, systems, documents, approvals, relationships and exceptions. Trying to redesign the whole organisation around SI at once creates too many unknowns: different data quality, different risks, different employee skill levels, different systems of record and different consequences of error.

The organisation may also mistake tool deployment for operating change. Hundreds of licences can be purchased quickly while the actual work remains unchanged. Employees still copy information manually, rebuild context, wait for approvals and fix the same handoff problems. The technology is present, but the workflow has not improved.

A narrow pilot is not timid. It is an information-gathering method. It reveals what context the system needs, which errors matter, how much review is required, how employees respond and whether the supposed value survives the full process.

The Brownfield Principle

Most organisations are brownfield environments: work already exists. Systems are already running. Employees have habits, spreadsheets, templates, informal rules and customer expectations. SI must enter that reality rather than an idealised blank sheet.

A brownfield strategy asks: what can we improve without breaking the current system? That usually means adding SI beside or inside one known task first. The old workflow remains available while the new one is compared against it.

This is different from designing a new business from scratch. A greenfield company may choose SI-native architecture from day one. An established organisation must preserve continuity while it learns. The transition path matters as much as the target state.

Start With Work People Already Understand

A first SI task should usually be performed today by someone who can explain what good work looks like. If nobody understands the task, SI is unlikely to repair the ambiguity. The first project should not be the most mysterious problem in the company.

  • Good early candidates: meeting summaries, recurring reports, document comparison, routine drafting, knowledge retrieval, intake classification, research preparation and structured handoffs.
  • Harder early candidates: rare strategic judgments, disputed processes, sensitive employment decisions, complex legal conclusions, destructive technical actions or workflows with no reliable source of truth.
  • Often not SI problems: simple arithmetic, stable deterministic rules, missing system integration, duplicate data entry that a direct API can remove, and policies that humans have not yet agreed on.

The first project should teach the organisation how to use intelligence, not merely demonstrate that a model can produce impressive text.

The Thin-Slice Method

A thin slice is the smallest part of a workflow that can produce measurable value without requiring the organisation to redesign everything around it. The slice should have a clear before-state, after-state and reviewer.

For example, do not begin with “automate customer service.” Begin with “prepare a draft response for routine delivery-status questions using the current order record and approved shipping policy.” That slice is easier to test. The organisation can measure draft time, correction rate and policy errors before considering automatic sending.

Do not begin with “build an AI finance department.” Begin with “draft management commentary on monthly variances from approved figures, with every figure traceable to the reporting system.” The finance team remains intact while one information-processing burden is reduced.

Choose a Task With a Visible Acceptance Test

A first pilot needs a way to tell whether the output is good enough. A visible acceptance test might be a source-backed summary, a correctly classified case, a report that matches an approved structure, a draft that requires only minor editing or a comparison table whose fields can be checked against the original documents.

Avoid first pilots whose only success criterion is “the output looks clever.” A polished answer can hide omissions. The acceptance test should belong to the real work.

The Three Doors Into Workplace SI

Door 1 — Personal productivity

One employee uses SI for familiar low-risk tasks: drafting, summarising, explaining, comparing or planning. This is the fastest way to learn because the user supplies context manually and reviews the result directly. The organisation can start with approved-tool guidance and basic data rules.

Door 2 — Team collaboration

A team standardises one repeated interaction. It agrees on input fields, source documents, output format and review. This turns private prompting into shared capability without requiring automation.

Door 3 — Process improvement

A workflow owner identifies a repeated bottleneck and adds SI at a specific transition: intake, context retrieval, interpretation, production or handoff. This creates stronger operational value but requires more process discipline.

These doors can be entered in parallel, but they should not be confused. A personal assistant programme and a connected operational workflow have different controls, evidence and support needs.

The Minimum Safe Starting Kit

A company does not need a hundred-page governance manual before a low-risk pilot, but it does need a few explicit rules. The minimum kit should answer which tools are approved, what information may be entered, what tasks require human verification, where employees report problems and who owns the pilot.

  • One approved tool or environment
  • One named process or pilot owner
  • One defined use case
  • One approved context/source set
  • One acceptance test
  • One exception rule
  • One baseline metric
  • One fallback to the existing method

This kit is intentionally small. The purpose is to create enough control for learning without turning experimentation into bureaucracy.

Do Not Connect Tools on Day One

The first version of a workflow often works best with manual context. The user uploads or pastes approved information, SI produces the candidate output and the person reviews it. This exposes what the task actually needs.

Connecting email, databases, calendars or business systems too early can hide the real learning inside integration complexity. If the workflow is not valuable manually, automation will not make it valuable. Prove the cognitive step first, then connect the sources that remove repeated manual work.

Run in Shadow Mode

Shadow mode means the SI workflow runs alongside the current process without controlling the final outcome. It produces a recommendation, draft or classification that can be compared with what the human team actually did.

Shadow mode is especially useful when the task has medium consequence or when the organisation wants evidence before changing responsibility. The team can measure disagreement, missed information and correction burden without exposing customers or systems to unproven automation.

The limitation is that shadow mode adds temporary work. It should be time-bounded and used to learn, not become a permanent duplicate process.

Move to Assisted Mode

Once the shadow output proves useful, let employees use it as part of the real workflow while keeping final authority human. The person can accept, edit or reject the SI contribution. This is Level-1 or Level-2 workplace SI: Assist or Collaborate.

At this stage, record material corrections. If employees repeatedly change the same phrase, classification or assumption, fix the instruction or context. The workflow should learn from repeated human effort rather than make people correct the same weakness forever.

Standardise the Context Packet

A successful pilot often reveals that the most important variable is not the prompt but the context. One user gets excellent results because they know which policy, example or customer history to include. Another gets generic output.

Turn that expert intuition into a simple context packet: objective, audience, current source, relevant history, constraints, output format and stop conditions. The packet makes performance more transferable across employees.

Standardise the Acceptance Test

The reviewer needs a shared definition of done. A meeting summary might require decisions, owners, deadlines and unresolved questions. A supplier comparison might require normalised pricing, service scope and source references. A customer draft might require confirmed facts and no unsupported promises.

A review checklist is not bureaucracy when it removes ambiguity. It lets the organisation measure whether SI quality improves and whether reviewers agree.

Measure the Whole Workflow

Do not measure only how fast the model responds. Measure from trigger to accepted outcome. Include time spent preparing context, reviewing output, correcting errors, waiting for clarification and fixing downstream problems.

  • Cycle time: how long from start to accepted completion?
  • Human touch time: how much active employee effort is required?
  • First-pass acceptance: how often is the output accepted without material changes?
  • Correction burden: how long does repair take?
  • Error rate: what important mistakes survive review?
  • Receiver effort: how much work does the next person need to do?
  • Throughput: does the team complete more useful work?
  • Capacity redeployed: where does saved effort actually go?

A workflow that saves ten minutes in drafting but creates fifteen minutes of review is not an improvement.

Start With One Owner

Every first pilot needs a person who can decide what the workflow is for, which sources are current and whether a correction should become a system change. Without ownership, employees can identify problems but nobody repairs the process.

The owner does not need to be a senior executive. A team lead, operations manager, teacher, accountant or other domain expert can own a bounded workflow. Technical teams can support the system while the process owner owns the outcome.

Use Existing Approval Paths First

A low-disruption pilot should fit into existing authority rather than invent a new governance structure. If a manager already approves a customer exception, SI can prepare the evidence while the same manager remains the approver. If a finance report already has sign-off, SI can draft commentary before that sign-off.

This keeps the organisational change small while the team learns. New authority structures should be created only when the workflow actually requires them.

The 30-Day Start Plan

Week 1 — Observe

Choose one team and one workflow. Record the current trigger, inputs, steps, waiting, handoffs, output and reviewer. Measure a simple baseline. Identify the repeated information burden.

Week 2 — Assist

Use SI manually on representative cases. Keep external actions human-controlled. Record what context the system needs and which errors recur. Do not optimise for elegance yet; optimise for learning.

Week 3 — Standardise

Create a reusable instruction, context packet, output format, acceptance test and exception rule. Test with more than one employee. Confirm that performance does not depend on one enthusiast’s private technique.

Week 4 — Decide

Compare the new method with the improved baseline. Decide whether to retain, modify, connect, automate or retire the pilot. Write down the next permission or integration only if the evidence justifies it.

A successful first month produces one understood workflow and reusable organisational learning. It does not need to produce a company-wide AI operating model.

The 90-Day Start Plan

Days 1–30 — Prove one workflow

Follow the 30-day plan above. Build a reliable Assist or Collaborate workflow and capture its failure modes.

Days 31–60 — Connect selectively

Connect one approved source or tool if manual copying is now the main friction. Add structured validation. Improve source ownership. Train a small additional group of users.

Days 61–90 — Automate one bounded transition

If the workflow is stable, automate a trigger or low-risk step. Add exception handling, logs and recovery. Keep consequential actions gated. Re-measure the complete workflow.

After ninety days, the organisation should know more about its own data, workflow and control requirements. That knowledge is more valuable than a generic transformation deck.

Start Without Reorganising the Org Chart

SI adoption does not require creating a new department immediately. Early learning can happen inside existing teams with a small central support function or working group. The organisation can formalise governance and centres of excellence later if the portfolio grows.

The first priority is to attach responsibility to real workflows, not invent titles. A Head of AI cannot substitute for process owners who know what good customer service, accounting, teaching or software engineering looks like.

Start Without Replacing Core Systems

Many first wins can sit around existing systems of record. SI can retrieve, interpret and prepare information while the CRM, accounting platform, project system or learning management system remains authoritative.

Replacing core systems creates migration risk unrelated to SI. Unless the current platform blocks the use case, keep it. Add intelligence at the interface or workflow layer, then reconsider architecture after value is proven.

Start Without Migrating Every Document

A company does not need to clean its entire knowledge estate before testing retrieval. Choose the small set of documents required for one workflow, establish which version is current and give them an owner.

This creates a miniature source-of-truth discipline that can expand later. Trying to fix every file repository before one use case is proven can stall adoption for months.

Start Without Automating External Actions

The system can create value before it is allowed to send, change, buy, deploy or approve. Preparation and recommendation often capture much of the cognitive benefit while keeping risk low.

This is especially useful for organisations that are still learning. A customer-response draft, research brief or code patch can be valuable even if a person performs the final action.

Start Without a Perfect ROI Model

Early pilots need a measurable baseline, but not every benefit must immediately be converted to dollars. Time to accepted output, reduced rework, faster response, better knowledge access and lower exception rates are legitimate operational measures.

Once the mechanism is understood, the organisation can connect those measures to financial outcomes. Forcing uncertain assumptions into an ROI spreadsheet too early can create false precision.

Start Without Choosing One Model Forever

The model is one component. The workflow should be designed around its outcome, sources, permissions and verification. Where practical, avoid binding the business logic so tightly to one model that changing providers requires rewriting the process.

This does not mean every system must be model-agnostic from day one. It means the organisation should know which parts of the workflow depend on model-specific behaviour and which parts are stable business requirements.

Personal Productivity Pilot

A professional who spends thirty minutes each morning turning notes into a daily plan can test SI immediately. The baseline is the current preparation time and usefulness of the plan. The pilot gives SI the calendar, approved tasks and notes, asks for a structured plan, then the user checks priorities. If the plan saves time without increasing missed commitments, the method is useful. No company redesign is required.

Meeting Workflow Pilot

A team can begin with post-meeting continuity. SI turns approved notes or transcripts into decisions, action items, owners and unresolved questions. The chair verifies the result before distribution. The real value is whether actions become clearer, fewer commitments are lost and follow-up time decreases.

Customer-Service Pilot

Choose one narrow request class such as delivery status or standard product questions. SI retrieves approved policy and order context, drafts a response and highlights missing information. The human agent reviews and sends. Do not begin with disputes or exceptions where policy and relationship judgment are harder.

Sales Pilot

Use SI to prepare account briefs before calls: current opportunity, recent communication, open commitments and relevant product information. The salesperson checks the brief and uses it to plan questions. Automatic sending should wait until the organisation understands which claims and commitments are safe to automate.

Finance Pilot

Use SI to draft narrative commentary around authoritative monthly figures. The finance professional checks that numbers reconcile and that the explanation matches known business context. This preserves the accounting system as the source of record while reducing the cost of turning data into management language.

HR Pilot

Start with administrative or learning tasks: draft role descriptions, answer routine policy questions from approved sources, prepare onboarding material or organise interview notes. Avoid making the first pilot an automated hiring or termination decision.

Legal Pilot

A legal team can test clause extraction, version comparison or first-pass issue spotting. The system returns source references and uncertainty. Qualified professionals retain interpretation and advice.

Engineering Pilot

An engineer can use SI for explanation, test generation, bounded code changes or documentation. Keep production deployment under existing engineering controls. Measure review time, defects and accepted changes.

Operations Pilot

Choose a repeated operational report or exception class. SI can assemble current information, classify the issue and prepare a response packet. Operators keep control of external actions. The pilot should reveal whether the real bottleneck is interpretation or missing system integration.

Education and Training Pilot

A teacher or training team can use SI to generate candidate explanations, practice items or differentiated materials from approved objectives. Human educators verify correctness and fit. The pilot succeeds when learning improves, not when content volume increases.

The Small-Business Starting Pattern

Small businesses often have one person performing several roles. The best first SI use case may therefore be cross-functional but still bounded: convert meeting notes into tasks, prepare customer follow-up, organise quotations or create first-draft marketing content from approved facts.

Avoid enterprise architecture before the business needs it. Use approved tools, simple templates and clear review. Add integrations only when repeated copying becomes the bottleneck.

The Large-Organisation Starting Pattern

Large organisations face the opposite problem: too many experiments. The first task is often not encouraging use but creating enough shared rails that experiments can compound. Approved tools, data rules, source ownership, reusable patterns and a small evaluation method help teams learn without creating hundreds of incompatible shadow workflows.

Do not centralise every use case. Centralise foundations; let workflow owners retain domain responsibility.

The Change-Management Minimum

Employees need to know why the pilot exists, what work it should reduce, what remains their responsibility and how their performance will be judged. If management introduces SI but leaves workload and expectations unchanged, workers may experience the system as extra work or surveillance.

A good first pilot explicitly names what employees can stop doing when the new method succeeds. That makes the benefit tangible.

Do Not Make Prompting the Whole Training Programme

Prompting is useful, but the deeper skills are task definition, context selection, evidence checking, exception recognition and workflow improvement. A clever prompt cannot repair an obsolete policy or an unauthorised action.

Train people on the work around the model, not just the words typed into it.

The Pilot Failure Catalogue

  • Too broad: automate marketing instead of one bounded workflow.
  • No baseline: nobody knows whether the new process is better.
  • Wrong bottleneck: drafting becomes faster while data gathering remains the real delay.
  • Weak context: users blame the model for missing information the workflow never supplied.
  • Review overload: output volume grows faster than human checking capacity.
  • Tool-first: integrations are built before task fit is proven.
  • Shadow forever: the pilot never transitions into or out of real work.
  • No owner: repeated problems are corrected manually but never repaired.
  • No stop rule: unusual cases are forced through the routine path.
  • No capacity destination: time saved is immediately filled with more low-value output.

How to Decide Whether to Keep the Pilot

At the end of the first test period, make an explicit decision. Keep the workflow when it improves the accepted outcome and its controls are manageable. Modify it when the mechanism is promising but context, review or scope needs repair. Retire it when the SI layer creates little value or ordinary process improvement solves the problem better.

A retired pilot is not failure. It is evidence that prevents the organisation from scaling an unsuitable design.

How to Expand Without Redesigning Everything

Expand adjacently. If a drafting workflow works, add retrieval. If retrieval works, add structured fields. If the structured output is reliable, automate a trigger. If the automated workflow is stable, consider a bounded action.

One adjacent capability at a time keeps cause and effect visible. The organisation knows which change created new value or new risk.

When a Company-Wide Redesign Eventually Becomes Necessary

Incremental adoption can reveal deeper structural constraints. If many workflows need the same identity, permissions, retrieval, logging or evaluation services, shared infrastructure becomes justified. If knowledge is fragmented everywhere, a broader knowledge programme may be necessary. If roles have changed materially, job design and governance may need formal redesign.

The difference is that the redesign is now evidence-led. It responds to patterns observed across proven workflows rather than abstract enthusiasm for AI transformation.

The Bridge to Integration

The next stage after repeated pilots is integration. Personal interactions become reusable instructions. Reusable instructions become team workflows. Workflows connect to approved knowledge and tools. Automated workflows share controls. Eventually, Super Intelligence becomes part of organisational infrastructure.

That progression is the subject of the next article: From AI Tools to SI Systems: Why Integration Matters. This page deliberately stops before full integration so the first step remains small and practical.

A First-Pilot Checklist

  1. Choose one repeated task people already understand.
  2. Name the outcome and receiver.
  3. Record the current baseline.
  4. Use an approved SI environment.
  5. Limit the source material to what is necessary.
  6. Keep external actions human-controlled.
  7. Define the acceptance test.
  8. Run representative cases, including should-stop examples.
  9. Record material corrections and exceptions.
  10. Decide whether to retain, modify or retire.
  11. Only then add a connector, trigger or new permission.

Frequently Asked Questions

Do I need an AI strategy before employees start?

You need enough policy to use approved tools safely, but you do not need a complete enterprise transformation blueprint before a low-risk pilot. Start with bounded tasks and build strategy from evidence.

Should we buy enterprise licences first?

Not necessarily. Define the use cases, data requirements, security needs and support model first. Procurement is easier when the organisation knows what workflows the product must support.

How many pilots should we run?

Enough to learn without fragmenting attention. One to three well-owned pilots are often more informative than dozens of uncontrolled experiments. Large organisations may run more, but they should use shared evaluation and governance rails.

Should the first pilot use an agent?

Usually not unless the use case genuinely requires multi-step tool use. Assistive or collaborative workflows reveal task fit with less complexity.

Can a small business start with consumer AI tools?

Only if the tool and account are appropriate for the information being used and the business accepts the relevant terms and data handling. Sensitive business data should follow approved organisational controls.

What if employees already use AI informally?

Treat existing use as discovery data. Identify repeated valuable patterns, move sensitive work into approved environments and standardise the workflows that deserve organisational support.

How long should a pilot run?

Long enough to encounter representative variation and measure a meaningful number of cases. The period depends on task frequency. A daily workflow may reveal patterns in weeks; a monthly workflow needs a longer observation window.

What if quality is good but review takes too long?

Improve the evidence packaging, narrow eligible cases, automate deterministic checks or keep SI at a lower autonomy level. Review burden belongs in the value calculation.

When should we connect company data?

After the manual version demonstrates value and the missing connection is now the main friction. Connect only the sources required for the task and manage permissions explicitly.

What should we read next?

Continue to From AI Tools to SI Systems: Why Integration Matters, then return to the Workplace Super Intelligence Hub for the full route.

The Core Starting Rule

Do not redesign the company before you have redesigned one useful piece of work. Start with a thin slice. Keep the existing process visible. Use Super Intelligence first as assistance or collaboration. Measure the complete outcome. Capture what works. Expand only when the evidence makes the next step obvious.

The purpose of the first pilot is not to prove that the organisation is AI-first. It is to become better at deciding where intelligence belongs. One reliable workflow is a stronger foundation than a hundred demonstrations.


What the First Pilot Must Leave Behind

A useful pilot should leave organisational memory, not just enthusiasm. The team should be able to show what task was tested, which source material mattered, what the system did, what humans corrected and whether the accepted outcome improved. If all of that disappears into one person’s chat history, the company has learned less than it thinks.

The minimum pilot record should contain the task definition, baseline, current source list, reusable instruction, accepted examples, correction categories, reviewer checklist, exception conditions, owner and decision about what happens next. This record becomes the starting point for the next person and the next workflow.

This is how incremental adoption compounds. The second pilot should inherit lessons from the first: source dates matter, status must distinguish draft from sent, review capacity can become a bottleneck, and unusual cases need an owner. Organisational learning is the mechanism that turns experiments into capability.

The Repair, Stabilise and Extend Path

Repair

Some teams discover that the existing process is too weak to automate. The source documents conflict, nobody agrees on the correct category, or the output varies because hidden knowledge lives in one employee’s head. In this state, SI should be used diagnostically: compare versions, identify repeated questions, expose missing context and help document the workflow.

The goal of repair is not automation. It is to establish a reliable floor. Once the process has a current source, owner and acceptance standard, the team can resume the pilot.

Stabilise

A stabilisation pilot already works, but only some of the time. One expert gets good results while another does not. The task is to make the context, instruction and review transferable. Capture the expert’s hidden decisions and convert them into shared infrastructure.

Test across several users and different case types. The workflow is stable when acceptable performance no longer depends on one person’s private prompting style.

Extend

A dependable workflow can gain one adjacent capability: retrieval, structured output, a trigger, a tool or bounded automation. Add one extension at a time so failures remain attributable. Expansion is a controlled experiment, not a reward for a successful demo.

The Minimum Viable Governance Ladder

Governance should grow with capability. A personal Assist workflow needs approved-tool guidance, data boundaries and user responsibility. A team Collaborate workflow adds shared sources, review criteria and ownership. An automated workflow adds permissions, logs, exception routing and recovery. An operating portfolio adds lifecycle review, incident management and shared control services.

This avoids two extremes: over-governing tiny experiments until nobody learns anything, and under-governing tool-using systems that can change external state. The organisation should add governance at the same rate that it adds authority.

The Minimum Viable Knowledge Ladder

A company does not need perfect enterprise knowledge before starting. It needs enough reliable knowledge for the chosen task. Begin by identifying the current authoritative documents for one workflow. Add ownership and version dates. Retire obvious obsolete copies. If retrieval becomes valuable, build a bounded corpus. Expand only after repeated workflows justify a wider knowledge service.

This is better than connecting SI to every file repository at once. An unrestricted knowledge layer can increase confusion because old and new documents compete silently. Small, governed source sets teach the organisation how retrieval should work before scale.

The Minimum Viable Technical Ladder

The technical path can be equally incremental: manual context → reusable instruction → structured output → approved retrieval → one tool connection → automated trigger → bounded agent → shared infrastructure. Each stage should remove a specific bottleneck observed in the previous stage.

Do not add an agent because the architecture is fashionable. If a fixed workflow is easier to understand and performs the task reliably, use it. Do not add a model to a deterministic transition that a direct API or formula handles better.

The Minimum Viable People Ladder

Human capability must scale alongside machine capability. The first user needs literacy and verification skill. The first shared workflow needs reviewers who agree on acceptance criteria. Automation needs a process owner and technical support. A growing portfolio needs security, governance, change management and workforce planning.

This prevents a common mismatch: sophisticated systems are deployed while employees are trained only to write prompts. Prompting is one small part of operating Super Intelligence.

The No-Big-Bang Rule

A big-bang transformation changes roles, tools, systems and policies simultaneously. When the outcome improves or degrades, the organisation may not know why. Incremental adoption changes fewer variables at a time, making learning causal rather than anecdotal.

Continuity is especially valuable in established businesses. The old workflow can remain available while the new one earns trust. If a pilot fails, the company has learned without putting the entire operating model at risk.

A Pilot Scoring Canvas

A simple scoring canvas helps compare candidate pilots without false precision. Use qualitative bands rather than pretending a decimal score can solve uncertainty.

  • Value: frequency, time consumed, bottleneck effect, service impact or strategic importance.
  • Feasibility: source availability, task clarity, user expertise, integration difficulty and review method.
  • Risk: consequence of error, data sensitivity, reversibility and external action.
  • Learning value: whether the pilot teaches a reusable pattern such as retrieval, structured review or exception handling.
  • Owner readiness: whether someone can maintain sources, define acceptance and make process decisions.
  • Transfer potential: whether the mechanism could help other teams after it is proven.

The ideal first pilot scores high on value, feasibility and learning while remaining low-to-moderate on risk. A glamorous project with low verifiability and no owner should move down the queue.

The Transfer Test

When a pilot succeeds, another team may want to copy it. Before transfer, compare five things: object, evidence, receiver, consequence and authority. A meeting-summary workflow may transfer easily across departments. A customer-refund workflow may not transfer to expense approval because money rules and evidence differ.

Copy the mechanism, not the surface form. If the success came from structured context and source-backed drafting, preserve those conditions. Do not assume that the words “classification” or “agent” make two workflows equivalent.

Shadow Mode: How to Learn Without Exposing the Outcome

Shadow mode deserves deliberate design. Choose cases the human process will handle anyway. Let SI produce its own classification, draft or recommendation. Store the result without allowing it to control the actual decision. Compare later against the accepted human outcome and source evidence.

Use shadow mode to learn disagreement, not to prove humans are always right. If the SI result appears better in some cases, inspect why. Perhaps it found a source the human missed. Perhaps the human knew tacit context the system lacked. The disagreement itself is useful diagnostic data.

Shadow mode should end. After enough evidence, promote the workflow to assisted use, redesign it or retire it. Permanent duplicate processing wastes capacity.

The Assisted-Mode Promotion Gate

Promote from shadow to assisted use when the system contributes enough value that employees should see the output during real work, and when the review burden is manageable. The system may still be wrong; the question is whether people can use it safely and productively.

  • The task and source set are stable enough for repeated use.
  • Employees can recognise common failure modes.
  • Important claims can be checked without redoing everything.
  • The system does not require unauthorised data.
  • The workflow has an exception path.
  • The old process remains available if the SI layer fails.

The Integration Promotion Gate

Do not connect company systems simply because assisted use works. Connect when repeated manual context transfer is now the main bottleneck and the organisation knows exactly which source is required. The new connection should remove friction rather than expand scope.

For example, if employees repeatedly copy the same customer history from CRM into an approved assistant, read-only CRM retrieval may be the next step. Do not simultaneously add sending authority, payment tools and broad customer-data access. One new permission at a time keeps the learning visible.

The Automation Promotion Gate

Automation becomes appropriate when the workflow has stable triggers, well-defined eligible cases, observable validation and an exception route. The manual collaborative version should already be dependable.

If reviewers still disagree about what good output looks like, or users must repair source data on every case, automation is premature. The machine will make ambiguity faster.

The First-SI Team Meeting

A manager can start with a practical one-hour session. Ask the team to list repeated tasks from the last two weeks. Group them into reading, searching, drafting, comparing, classifying, calculating, deciding and acting. Mark which tasks consume time and which require human authority.

Choose one candidate whose output the team can verify. Define the first experiment and baseline. Do not spend the meeting debating the future of every job. The purpose is to find one work unit worth testing.

The First-SI Manager Brief

Managers need to tell employees what the pilot is for and what it is not for. State which task is being tested, what data may be used, who reviews outputs, what work should become easier and how feedback will be treated.

Employees should know that reporting errors is part of the experiment, not evidence that they failed to use the tool correctly. A learning culture surfaces weaknesses early.

The First-SI Executive Brief

Executives do not need to choose the final architecture during the first pilot. They need to set strategic boundaries: which outcomes matter, which data or actions require stronger control, and what evidence will justify further investment.

A useful executive question is: “What will we know after this pilot that we do not know now?” If the answer is vague, the experiment is probably too broad.

The First-SI Technical Brief

Technical teams should resist over-engineering. Clarify the expected inputs, output, data class, tool boundary and fallback. Start with the simplest implementation that can test the hypothesis.

Build reusable foundations only when repeated demand appears. A small pilot does not need an enterprise orchestration layer, but it should not violate security or identity rules simply because it is temporary.

The First-SI Risk Brief

Risk, security or compliance teams can help define the pilot envelope rather than merely approve or reject AI in the abstract. Which information can be used? Which actions must remain manual? What evidence should be logged? What would require the pilot to stop?

A bounded envelope makes experimentation safer and reduces friction between innovation and governance.

A More Detailed Customer-Service Pilot

Suppose a support team spends large amounts of time on order-status questions. The first version gives agents an approved assistant. The input contains the customer’s message, current order record and shipping policy. The output is a draft response plus a list of missing facts. The agent checks and sends.

The acceptance test is simple: no unsupported delivery promises, correct order state, appropriate tone and a clear next step. Measure time to final send, material edits, policy errors and customer recontact. If stable, add read-only retrieval from the order system. Later, automate only the narrow cases whose evidence is complete.

Notice what does not change initially: the support team, CRM, approval rules and escalation path. The pilot improves one repeated information transition without redesigning customer service.

A More Detailed Management-Reporting Pilot

A manager spends hours combining inconsistent weekly updates. The first pilot standardises five fields: completed work, next work, blockers, risks and help needed. SI converts submissions into a draft briefing and flags missing updates.

The manager checks the source links and adds interpretation. Measure preparation time, missing information detected and number of factual corrections. If the workflow works, connect it to the project system. Do not automate the manager’s prioritisation simply because report assembly became easier.

A More Detailed Sales Pilot

A sales team repeatedly reconstructs account context before meetings. Start with an account brief generated from approved notes and CRM fields supplied manually. The salesperson checks the brief and records missing information.

Once the fields are known, connect read-only CRM access. Add a follow-up draft only after the pre-call brief is stable. Keep pricing and commercial promises under existing authority. Expansion follows evidence, not the desire to create a “sales agent”.

A More Detailed Finance Pilot

A finance team uses authoritative monthly numbers but spends time writing narrative commentary. Give SI the approved variance table, definitions and examples of acceptable management commentary. Ask for a draft that cites each figure back to the provided source.

The finance professional verifies arithmetic in the source system and judges the explanation. Measure time to accepted commentary and material interpretation changes. Only later consider automated retrieval. The ledger remains the source of truth throughout.

A More Detailed Education Pilot

A teacher wants faster differentiated practice. Start with one learning objective and a small set of verified example questions. Provide recent student errors and ask SI for candidate practice that isolates the misconception.

The teacher solves and checks every item during the pilot. Measure preparation time and whether students perform better on an independent transfer question. If the material looks good but learning does not improve, content generation is not solving the educational problem.

A More Detailed Engineering Pilot

An engineering team wants SI coding assistance. Choose one class of low-risk changes, such as tests or documentation. The engineer gives repository context and a clear specification. SI proposes the change, tests run, and peer review remains in place.

Measure accepted changes, review time and defects. If value appears, expand to bounded code changes on branches. Production deployment remains behind existing CI/CD controls until the organisation has evidence to justify a different boundary.

The First 100 Cases

Where task frequency allows it, the first hundred cases are more informative than the first perfect demo. They reveal variation, edge cases, reviewer disagreement, data gaps and operational burden.

Tag each material correction by cause: missing context, stale source, ambiguous instruction, model error, bad taxonomy, wrong permission or reviewer mistake. The correction catalogue becomes the roadmap for improvement.

What to Do With Repeated Corrections

If the same error appears repeatedly, fix the system. Add the missing definition, update the canonical source, change the output schema or narrow the eligible cases. Repeated human correction is not “human in the loop” working successfully; it is a sensor showing where the workflow needs repair.

The best pilots reduce repeated correction over time. If correction stays flat, the workflow may not be learning at the organisational level.

What to Do When Humans Disagree

Reviewer disagreement is also useful. It may reveal that the task standard itself is unclear. Collect difficult examples and discuss why reasonable people disagree. Update the rubric or policy before blaming the model.

SI can only automate a definition the organisation has actually made. Where discretion is legitimate, keep that discretion visible rather than forcing false consistency.

What to Do When SI Beats the Baseline

If the SI-assisted workflow consistently outperforms the old process, do not jump immediately to maximum automation. Ask what mechanism created the gain. Was it faster retrieval, more consistent formatting, broader comparison or better continuity? Preserve that mechanism while expanding cautiously.

This helps the organisation avoid overgeneralising from one success. A model that excels at summarising may not deserve authority to act.

What to Do When SI Loses to the Baseline

If the workflow is slower or lower quality, diagnose before abandoning the category. The task may be unsuitable, the context weak, the review badly designed or the wrong bottleneck selected. Repair the first weak link and retest if the potential value justifies it.

If the mechanism still does not help, retire the pilot. Ordinary process improvement or conventional software may be the better answer.

Capacity Released Is Not the Same as Cost Saved

A pilot that saves employee time creates capacity. It does not automatically reduce cash expenditure. The organisation receives financial value only if the capacity is redeployed productively, avoids new hiring, improves service, increases revenue or supports another measurable outcome.

Keep capacity and cash measures separate until the causal link is clear. This avoids inflated ROI claims and helps managers decide what employees should do with time saved.

The Capacity Destination

Every successful pilot should name where the saved attention will go. A salesperson may spend more time with customers. A teacher may give more feedback. An analyst may investigate anomalies. An operations team may improve runbooks.

If the destination is simply “produce more machine-generated output”, the organisation may increase volume without increasing value.

The First Shared Prompt Library

Do not create a giant prompt library before recurring patterns exist. Capture only the instructions that have proven useful across multiple real cases. Store the purpose, required context, output format, examples and review notes alongside the prompt.

A prompt without context or acceptance criteria is not a workflow asset. The library should teach people how to delegate the task, not merely give them magic words.

The First Shared Knowledge Pack

For each proven workflow, create a small knowledge pack: current policies, definitions, templates and examples. Name an owner and review date. This is enough to support reliable context without launching an enterprise knowledge project.

As several workflows use the same sources, the packs can merge into a broader governed corpus.

The First Shared Evaluation Set

Keep a sample of representative cases, including difficult and should-stop examples. When instructions, models or sources change, rerun the set. Evaluation does not need to be mathematically elaborate to be useful; it needs to detect whether important behaviour changed.

This is one of the first pieces of infrastructure that allows SI adoption to become maintainable rather than anecdotal.

The First Fallback Procedure

Write down what happens when the tool is unavailable or untrusted. The fallback may be the old manual process, a simpler template or a queue for later handling. Critical workflows should not discover their fallback during an outage.

The fallback can be slower. Its job is continuity and control.

The First Stop Conditions

  • Required source is missing.
  • Two current sources conflict.
  • Sensitive data appears outside the approved scope.
  • Output requires an action beyond granted authority.
  • Validation fails.
  • The case belongs to a high-consequence category.
  • The system cannot establish whether a previous external action succeeded.

Stop conditions create confidence because employees know when not to rely on the normal path.

The First Review Trigger

Define what event causes the pilot to be reassessed: a material model update, new data source, new tool permission, rising error rate, policy change, security incident or expansion into a more consequential action.

Without review triggers, a small low-risk experiment can quietly evolve into a high-autonomy system while still operating under the assumptions of day one.

Singapore’s Task-First Guidance

Singapore’s Institute of Digital Government has advised public-sector users to start with the task rather than the tool and highlights everyday uses such as writing, research, meetings and knowledge retrieval. That task-first logic fits incremental workplace adoption: identify a real job to be done, then decide whether SI is the right component. See How do we use AI at work?.

The same source also encourages responsible use and verification rather than treating generated output as automatically authoritative. The exact controls in each organisation will differ, but the operating principle is portable.

Why the First Pilot Should Be Boring

Boring work is often frequent, measurable and expensive in aggregate. Repeated reporting, document comparison, knowledge search and routine drafting may create more value than an impressive autonomous demonstration used rarely.

A boring pilot also reduces organisational politics. The team can discuss whether the workflow works rather than whether SI threatens the identity of an entire profession.

Why the First Pilot Should Not Be Invisible

Employees should know when SI is part of a workflow and what they are expected to verify. Hidden use creates inconsistent standards and makes it harder to learn from errors.

Transparency does not require announcing every trivial assistive use. Focus on workflows where the system materially affects shared work, external communication, data or decisions.

Why the First Pilot Should Not Be Permanent by Default

A pilot is a test with an exit. At the end, decide to keep, change or remove it. Temporary status should not become an excuse for unmaintained systems that remain in production indefinitely.

If the workflow becomes permanent, give it an owner, version, source lifecycle and review trigger.

The Promotion Sequence

Manual task → SI-assisted task → shared collaborative workflow → connected workflow → bounded automation → agentic workflow → shared infrastructure. Not every task needs to reach the end. The sequence simply shows how capability can grow without forcing a whole-company redesign.

The key is that every promotion solves a real bottleneck discovered at the previous stage.

The Reader’s Monday-Morning Exercise

  1. Write down the three tasks you repeated most often last week.
  2. Circle the one with the most reading, searching, rewriting or organising.
  3. Write the accepted output in one sentence.
  4. Name the source material required.
  5. State one error that would matter.
  6. Use an approved SI tool to produce a candidate result on one real case.
  7. Check it against the source and record what you changed.
  8. Repeat on five to ten cases.
  9. Keep the task only if the final accepted work improves.

That is enough to begin. No reorganisation, agent platform or company-wide transformation is required to learn whether Super Intelligence belongs in that piece of work.

The Incremental Adoption Rule

Keep the organisational surface area small until the evidence is large. Start with one task, one owner, one source set and one acceptance test. Let success earn integration. Let integration earn automation. Let repeated automation earn shared infrastructure.

That sequence protects continuity while still moving quickly. The company learns from reality instead of betting the whole operating model on assumptions.


The Pilot Operating Pack

Before a pilot graduates from experiment to real operating method, assemble a compact operating pack. The pack should be short enough that employees actually use it and complete enough that a new authorised person can run the workflow without reconstructing its logic from old chats.

  • Purpose: the outcome this workflow exists to improve.
  • Scope: eligible cases and excluded cases.
  • Sources: current approved documents or systems.
  • Instruction: the reusable delegation pattern.
  • Output: required structure and status labels.
  • Review: what the human checks and what can be validated automatically.
  • Exceptions: observable stop conditions and escalation owner.
  • Permissions: what SI may read, recommend, prepare or execute.
  • Metrics: baseline and current performance.
  • Fallback: how the task continues when SI is unavailable or untrusted.
  • Owner: who changes the workflow when reality changes.

This pack is the beginning of operating discipline. It is also a signal that the workflow has become an organisational asset rather than a personal trick.

The Department Starter Matrix

Executives

Start with decision-brief preparation, report compression and question generation. Keep strategy, resource allocation and final commitments under leadership authority. The pilot should reduce time spent assembling evidence rather than pretend to automate leadership.

Operations

Start with recurring reporting, exception summaries, runbook retrieval or incident timelines. Keep destructive or production-impacting actions behind existing controls. Measure queue time, escalation quality and repeated manual search.

Marketing

Start with research organisation, first drafts, repurposing and content variants from approved facts. Keep regulated claims, major brand positions and final publication under editorial control. Measure accepted content and revision effort, not raw word volume.

Sales

Start with account briefs, call preparation and follow-up drafts. Keep pricing, commitments and relationship-sensitive decisions human-led. Measure preparation time, follow-up consistency and factual corrections.

Customer Service

Start with one routine request class. Use approved policy and current customer state. Keep disputes and high-value exceptions human-led. Measure response time, recontact and escalation quality.

Finance

Start with commentary, document organisation and anomaly preparation around authoritative numbers. Keep source-of-record arithmetic and material approvals under established controls. Measure time to accepted report and interpretation corrections.

HR

Start with onboarding, policy explanation, job-description drafting and administrative organisation. Keep employment decisions under appropriate human and legal governance. Measure administrative effort and knowledge access.

Legal

Start with comparison, extraction and issue spotting with source references. Keep interpretation and advice with qualified professionals. Measure document-review preparation time and missed issues.

Engineering

Start with explanation, tests, documentation and bounded code changes. Keep merge and deployment under existing CI/CD controls. Measure review time, defect rate and accepted changes.

Research

Start with source organisation, extraction and comparison. Preserve citations and researcher judgment. Measure evidence coverage, source quality and time to validated synthesis.

The Pilot Measurement Examples

Example: drafting workflow

Baseline: twenty minutes from request to accepted email. Pilot: five minutes to generate, ten minutes to verify and edit, two minutes to send. The true improvement is three minutes, not fifteen. If later context improves and review falls to five minutes, the workflow creates a larger gain. Always include the whole path.

Example: retrieval workflow

Baseline: employees spend twelve minutes finding the current policy. Pilot: SI answers in thirty seconds but one in ten answers cites an obsolete document. The speed gain is real, but reliability is not acceptable. Repair source ownership before scaling.

Example: classification workflow

Baseline: staff route one hundred requests manually in sixty minutes. Pilot: SI routes ninety-five correctly, flags three uncertain cases and misroutes two. The organisation can decide whether the two errors are tolerable based on consequence and whether deterministic checks or better categories can reduce them.

Example: meeting workflow

Baseline: action items are often missed and managers spend thirty minutes writing follow-up. Pilot: SI creates actions in two minutes, the chair spends eight minutes verifying, and missed commitments decline. The value includes both time and continuity.

The Falsification Test

A serious pilot should state what evidence would make the organisation stop believing the workflow is useful. Examples: correction time remains as high as the manual task; error rates exceed the current process; employees need to reread every source; review queues grow; sensitive data enters the wrong environment; or downstream users spend more effort than before.

Falsification protects teams from sunk-cost bias. Once money and reputation are invested in a project, people may reinterpret every result as progress. A pre-defined failure condition makes retirement an acceptable outcome.

The Independence Test

Ask another authorised employee to run the workflow without help from the original enthusiast. Can they identify the correct sources, give the right context, interpret the output and perform the review? If not, the method is not yet organisational.

The repair may be simple: better examples, a clearer context packet, a labelled source set or a more explicit acceptance checklist. Independence is a powerful signal that the process is stabilising.

The Receiver Test

Give the output to the person who actually uses it next. Do they receive a better handoff? Do they need fewer clarifications? Can they see the evidence and status? A workflow that pleases the generator but burdens the receiver is not an improvement.

Receiver effort should be part of the metric whenever work crosses team boundaries.

The World-Return Test

For any workflow that eventually changes external state, confirm the real state. Was the message sent? Was the calendar event created? Did the record update? Did the customer receive the promised follow-up? Did the learner demonstrate the skill? Generated output is not the same as completed work.

This test becomes essential as pilots move toward automation. The real system of record should return evidence of completion.

The Permission Test

Before adding any connection, list exactly what the system must read and what it must change. If read access is enough, do not grant write. If preparing an action is enough, do not grant execution. If one folder is enough, do not connect the entire drive.

Least-necessary access keeps early experiments easier to govern and easier to explain.

The Review-Capacity Test

Estimate how many outputs the pilot can create and how many the human reviewer can meaningfully inspect. If production can grow faster than review, define which low-risk cases can be validated automatically, which cases should be sampled and which must always receive human judgment.

Do not let a successful pilot create a hidden queue that eventually turns review into rubber-stamping.

The Skill-Preservation Test

Ask which human skills remain necessary for verification, exception handling or fallback. If SI performs the routine task, will employees still know enough to detect a serious mistake or recover during an outage?

Preserve critical capability through explanation, occasional manual practice, peer review or training. The goal is leverage without brittle dependence.

The Fair-Baseline Test

A pilot often reveals ordinary process improvements: clearer templates, better source ownership, fewer unnecessary fields. Apply those improvements to the baseline too before claiming SI created the entire gain.

The fair comparison is the best practical non-SI process versus the best practical SI-enabled process, not an improved pilot versus a deliberately neglected old method.

The Upgrade Ledger

Keep a small ledger of every material change to the pilot: new prompt, new model, new source, new permission, new tool, new approval threshold. Note why the change was made and what representative cases were retested.

This ledger creates traceability without heavy bureaucracy. It also prevents a low-risk pilot from quietly accumulating capabilities until nobody remembers what was originally approved.

The Assumption Ledger

Write down material assumptions: the policy is current, the customer ID is reliable, the input language is supported, the reviewer can handle the queue, the source system is available. When an assumption breaks, the workflow should be reassessed.

Many operational failures begin when an old assumption remains invisible after reality changes.

The Exception Ledger

Track the reasons cases leave the normal path. Repeated exceptions reveal where the workflow may expand safely or where its scope should remain narrow. If forty percent of cases escalate for the same missing field, fix intake. If rare legal questions escalate appropriately, the boundary may already be correct.

An exception rate is not automatically bad. It is information about the relationship between the designed envelope and real work.

The Correction Ledger

Categorise material human changes rather than storing only the final accepted output. Corrections show whether the weakness belongs to context, instruction, taxonomy, source quality, model capability, review or policy.

The correction ledger is one of the strongest inputs for deciding whether to standardise, automate or retire the workflow.

The Pilot-to-Production Handoff

When a pilot becomes permanent, change its status explicitly. Give it a version, owner, source-review date, metrics, fallback and incident route. Move secrets or credentials out of ad hoc user setups into managed configuration. Document tool permissions.

Production status means someone is responsible for maintaining the workflow after the excitement of the pilot has ended.

The Pilot-to-Integration Threshold

A workflow is ready for deeper integration when repeated manual context transfer is clearly the remaining bottleneck, the task itself is already useful, the required source is known, the source has an owner, and connecting it does not expand authority beyond what the team has tested.

If the team still debates what the task should produce, integration is premature. Connectors remove friction; they do not create task clarity.

The Pilot-to-Automation Threshold

A workflow is ready for automation when triggers are stable, eligible cases are distinguishable, output can be validated, exceptions have owners, and the team understands what happens when the system fails. The collaborative version should already have a strong track record.

Automation should reduce repeated human initiation, not remove meaningful judgment simply because approval feels inconvenient.

The Pilot-to-Agent Threshold

An agent becomes reasonable when the workflow genuinely needs dynamic multi-step tool use and a fixed path would be brittle or excessively complex. The organisation should already understand the objective, tools, permissions, stopping conditions and recovery.

If a fixed workflow works, keep it. Agentic freedom is useful only when it pays for its added uncertainty.

The Pilot-to-Infrastructure Threshold

Shared infrastructure becomes justified when several proven workflows need the same capabilities: identity, model routing, retrieval, permissions, evaluation, logs, monitoring or cost controls. Build common rails after repeated demand appears.

This is the moment when company-wide redesign may finally make sense—not as the starting assumption, but as the result of evidence accumulated across local successes.

The Five Outcomes of a Pilot

  1. Retire: SI adds little value or the task is better solved another way.
  2. Repair: the workflow has potential but sources, task definition or review are unstable.
  3. Stabilise: the method works but needs shared context and evaluation.
  4. Integrate: manual context transfer is now the main bottleneck.
  5. Automate: the workflow is stable enough for trigger-based execution within bounded permissions.

A sixth outcome—move toward agentic or shared infrastructure—comes only after one or more automated workflows demonstrate sustained value.

The No-Transformation-Theatre Rule

Avoid measuring progress by the number of licences, agents, prompts, training sessions or generated words. Those are activity measures. Transformation is real when accepted outcomes improve and employees can explain how the new workflow works.

A company with one deeply useful SI workflow may be more operationally mature than a company with hundreds of uncontrolled AI experiments.

The No-Cannibalisation Rule for the Content Series

This article owns incremental starting strategy. It does not own full integration architecture, workplace readiness, workflow mapping or automation engineering. Those subjects have separate pages in the series so each search intent has one clear owner.

The next article takes the handoff from pilot to system: why isolated AI tools eventually need context, connections, controls and measurement to become a durable Super Intelligence operating layer.

Final Pilot Readiness Checklist

  • One repeated task
  • One clear accepted output
  • One owner
  • One approved source set
  • One approved SI environment
  • One baseline
  • One reviewer or validation method
  • One exception route
  • One fallback
  • Representative cases
  • Recorded corrections
  • A decision date
  • A next-step gate based on evidence

If those elements exist, the organisation can learn quickly without redesigning itself first.

The Final Starting Principle

Begin locally, learn rigorously, expand adjacently. That is the workplace SI adoption path. A small pilot should not be small in discipline. It should be narrow enough to understand, rich enough to produce evidence and useful enough to teach the organisation what the next step should be.

A company does not become SI-ready by announcing a transformation. It becomes SI-capable by repeatedly turning real work into better measured workflows.


The Pilot Evidence Card

At the end of the first cycle, compress the pilot into a one-page evidence card. Record the old method, the new method, representative cases, time to accepted outcome, material errors, review burden, exceptions and the decision to retire, repair, stabilise, integrate or automate. This forces the team to distinguish what it observed from what it hopes will happen at scale.

The evidence card should also name the mechanism of value. “AI saved time” is too vague. “Approved retrieval removed repeated document search”, “structured drafting reduced blank-page work”, or “classification moved routine cases out of a manual queue” is specific enough to transfer. Mechanism matters because it tells the next team what conditions must be present for the result to repeat.

The Organisational Return Path

Every pilot should return something to the organisation even if the workflow is retired. A failed retrieval pilot may reveal that policy ownership is broken. A weak drafting pilot may reveal that the team has no shared definition of a good report. A review-heavy coding pilot may reveal that requirements are underspecified. These discoveries are process assets.

The return path asks: what did this experiment teach us about the work, the data, the people and the controls? Feed that lesson into documentation, training, source ownership or the next pilot. A company becomes better at SI adoption when failed experiments improve its operating knowledge rather than disappear.

The Independence-After-Ninety-Days Test

After three months, ask whether the workflow still depends on the person who started it. Can a colleague run it? Can the owner explain which sources are authoritative? Can technical support diagnose a failure? Can management see the outcome metrics? Can the organisation stop the workflow without losing the underlying business process?

If the answer is no, the pilot may be useful but not yet institutionalised. Repair transferability before expanding scope. The goal is not to remove experts; it is to ensure the method can survive ordinary absence, turnover and change.

The Pilot Portfolio Rule

Once several pilots exist, compare them as a portfolio. Which ones share the same knowledge sources, evaluation method or permission pattern? Which ones create the same exception classes? Which ones are merely variants of one mechanism? This is the point where shared infrastructure becomes visible from evidence rather than from architecture diagrams.

  • Merge duplicated source packs.
  • Standardise common review patterns.
  • Share evaluation cases where the task is genuinely equivalent.
  • Centralise security and identity controls that several workflows need.
  • Retain domain ownership for the parts that differ.

The portfolio view is the bridge from local experiments to organisational systems. It should emerge after repeated learning, not replace the first pilot.

The Final Test Before Calling the Start Successful

A successful start means the organisation can answer five questions with evidence: Which piece of work improved? Why did it improve? What still belongs to humans? What failure conditions were discovered? What is the next smallest justified step?

If those answers are clear, the company has started using Super Intelligence properly even if only one workflow changed. If those answers are unclear, a large number of licences, prompts or agents does not compensate. Starting well is an operating discipline, not a publicity milestone.