Human-Centred AI Workflows

AI guidance often ends at the prompt. Real work begins after the response: someone must check the evidence, restore missing context, decide whether the output is suitable, and accept responsibility for using it. This guide treats AI as one participant in a workflow rather than an authority floating above it.

That distinction is increasingly important. The International Labour Organization's 2025 occupational-exposure index concluded that transformation is more likely than wholesale job replacement because occupations contain different tasks and most still require human input. Microsoft's 2026 Work Trend Index likewise describes experienced users pausing to choose whether a person or AI should do a task. Neither source proves that one universal workflow will succeed. Together they support a task-level question: what should this system contribute, and what must a person retain?

The framework has five stages: map the task, set the boundary, state the acceptance test, run a human review loop, and learn from repeated corrections. Use it for ordinary knowledge work such as research, drafting, analysis, planning, and internal operations. In high-consequence fields, follow the professional, legal, security, and organizational controls that apply. A personal checklist cannot replace them.

Choose one recurring piece of work and break it into observable steps. A report might involve gathering evidence, deciding which evidence is relevant, calculating results, drafting an explanation, checking limitations, approving the conclusion, and publishing it. 'Write the report' hides all of these decisions inside one instruction.

Score each step on consequence, context, and learning value. Consequence asks what happens if the step is wrong or disclosed. Context asks how much the answer depends on tacit knowledge, sensitive information, local rules, or current facts. Learning asks whether doing the step builds a capability the worker must retain.

Low-consequence, reversible transformations with clear inputs and inspectable outputs are the easiest assistance candidates. High-consequence decisions with hidden context and high learning value should remain human-led. Between those extremes, use AI to support a defined operation while a named person owns the decision.

Also identify information that must never enter an unapproved system: personal data, customer material, confidential strategy, credentials, protected records, or copyrighted content you are not entitled to use. Check the provider terms and your organization's rules rather than assuming a consumer tool is suitable. This map becomes the workflow's first control and a concise explanation other people can inspect.

A delegation contract is a short statement of the tool's role, boundaries, and acceptance criteria. It can live in a template, team handbook, or project brief. Include the allowed input, permitted operation, prohibited decisions, required evidence, reviewer, and retention rule.

For example: 'The system may create a provisional summary from approved meeting notes. It must separate decisions, actions, and unresolved questions; it may not infer performance judgments. The meeting owner compares the summary with the notes and approves it before distribution. Raw notes are handled under the team's approved retention policy.'

Now write an acceptance test before generation. A useful test is observable: all material claims trace to an approved source; calculations can be reproduced; uncertainty is stated; the tone fits the audience; confidential material is absent; and the named reviewer can explain the result. Avoid criteria such as 'high quality' unless the team has defined what that means.

The contract should be proportional. A private brainstorm may need only a reminder not to enter confidential data. A customer recommendation may need source verification, subject-matter review, version history, and approval. The point is not paperwork around every interaction. It is to prevent speed at the beginning from erasing responsibility at the end.

Review in separate passes so fluency does not substitute for inspection. First, mark factual claims, numbers, quotations, policy statements, and citations. Open the original sources and confirm that they support the exact claim, population, date, and strength of conclusion. Generated references are leads, not proof.

Second, inspect reasoning. List assumptions, plausible alternatives, and observations that would change the answer. Be alert when an association has become a causal claim or a precise number has appeared without a defensible basis. Third, restore context: jurisdiction, audience, accessibility, budget, local exceptions, security requirements, and the history of earlier decisions.

Fourth, test the output. Reproduce calculations, run code in an appropriate environment, try counterexamples, compare a sample with a trusted baseline, or ask a qualified second reviewer. NIST's AI Risk Management Framework and Generative AI Profile emphasize roles, documentation, testing, evaluation, verification, and validation. A small team can apply the principle without copying an enterprise process.

Finally, run the ownership test: who is qualified and willing to approve this result? Record that person beside consequential work. If no one understands the basis well enough to own it, the output is not ready.

A workflow can produce acceptable outputs while quietly weakening the team's ability to judge them. Protect the skills required for review. Learners should make an unaided first attempt before requesting a complete answer. Experienced workers should periodically perform key tasks without assistance or explain the method from first principles.

Use AI for feedback after an attempt: request one hint, a counterexample, test cases, or critique of stated reasoning. Then close the tool and retrieve the lesson from memory. If a person cannot reproduce or explain the central step, record the output as assisted completion rather than learned capability.

Agency also means preserving a real choice not to use the system. Watch for workflows where targets assume AI-level volume, so opting out becomes impossible even when the tool is unsuitable. Measure time to approved work, error correction, and worker confidence—not only the number of generated outputs.

When assistance saves time, decide where the capacity should go. Better service, deeper analysis, learning, and recovery are deliberate choices. Without one, the recovered hour is often absorbed by more status traffic. A human-centred workflow should improve the quality and sustainability of work, not merely its throughput.

Day 1: choose one repeated, moderate-consequence task and map its steps. Day 2: score consequence, context, and learning value; identify data that cannot be used. Day 3: write the delegation contract and acceptance test. Day 4: run the task with the human review loop and record corrections. Day 5: repeat with a second example and compare the failure patterns.

On Day 6, ask another person to follow the workflow. Note any instruction that depends on knowledge only you possess. On Day 7, review four measures: total time to approved output, material corrections, evidence verified, and whether the responsible person can explain the result.

Keep the workflow if it reduces total effort without weakening quality, privacy, learning, or ownership. Redesign it if the same correction repeats. Stop using it for that task if review costs exceed the value, the reviewer cannot inspect the result, or sensitive context cannot be handled safely.

Revisit the contract when the model, data source, policy, audience, or consequence changes. Tool capability is not a permanent authorization. The lasting skill is knowing how to make the boundary visible, test the result, and keep a person accountable for the decision.