In this guide
Choose a task, not a tool
Begin with work you already understand. Look for a recurring task that takes time because you gather, sort, compare or reshape information: preparing a weekly update, turning notes into actions, outlining a presentation or checking a draft against a brief. A good first task has a clear beginning, a recognisable output and consequences you can manage.
Avoid starting with a high-stakes decision or a task you cannot personally evaluate. If you would not know whether the result is wrong, you are not ready to delegate part of that task to AI. Write one sentence: AI will help me ____ so that I can ____. If the second blank does not name a useful outcome, choose a better task.
Define what good looks like
Describe the output before writing the prompt. Who will use it? What decision or action should it support? Which facts must appear? What tone, length and format fit the situation? List the errors that would make the output unusable. This turns quality from a vague feeling into something you can inspect.
Create a short acceptance check. For a meeting summary it might be: every action has an owner; dates match the notes; unresolved points are labelled; no decision is added unless it appears in the source. Your check should reflect the real task, not the model’s idea of a polished answer.
Set the information boundary
Confirm which AI tools your organisation permits and what information may be entered. Do not paste customer records, employee details, contracts, unpublished results, passwords or confidential operating information into an unapproved service. Removing a name may not remove all identifying or sensitive context.
For your first experiment, use fictional or non-sensitive material. If the workflow eventually needs workplace data, ask the appropriate owner about the approved tool, retention settings and handling rules. A useful workflow must fit the organisation’s responsibilities as well as your own convenience.
Build a five-stage loop
Use the same five stages each time: prepare, instruct, draft, verify and deliver. Prepare the safe source material. Instruct the model with the task, audience, constraints and required structure. Generate a draft rather than a final answer. Verify it against the source and your acceptance check. Deliver only after you have edited and taken responsibility for the result.
Keep the human decisions visible. AI may group notes or propose an outline; you decide what matters. It may draft alternatives; you select and revise. It may identify gaps; you decide how to resolve them. The workflow should make your judgement easier to apply, not hide where it is needed.
Use a reusable task brief
A strong task brief is more useful than a clever one-line command. Give the model a role only when it clarifies the perspective. State the outcome, audience, supplied material, constraints and desired format. Tell it how to handle missing information and what it must not invent. Ask for an intermediate structure when the task is complex.
Save the brief beside the acceptance check, not beside a long list of disconnected prompts. The pair belongs together: one explains what to produce; the other defines what you will accept. Update both when the task, audience or workplace rules change.
Practice brief
Task: Turn the fictional notes below into a weekly operating update. Audience: a manager choosing next week’s priorities. Use only the supplied facts. Structure: What changed, Evidence, Risks, Decision needed. Label missing information as a question. Do not invent causes, numbers, owners or deadlines. First propose an outline and wait for review.
Run a controlled comparison
Test the workflow on one task you can complete without AI. Keep the original method as a comparison. Record the time you spend preparing inputs, prompting, checking and correcting—not only the time spent generating. Then compare the two outputs against the same acceptance check.
Ask three questions: Did the AI-assisted version make the result clearer or more complete? Did checking it create new work? Did it introduce a risk or error the old process avoided? A faster draft is not a better workflow if verification becomes difficult. Keep the method only when the total process is useful.
Document the workflow and its stop conditions
Write a one-page procedure: purpose, approved tool, allowed inputs, prompt or brief, review checks, owner and final destination. Add stop conditions. Examples include: the source contains restricted information; the task affects employment or another high-impact decision; the output cannot be verified; the tool behaves unexpectedly; or the task has changed beyond the tested use.
Review the workflow after several uses. Keep examples of failures as well as successes. Small documented improvements make the process more dependable and easier to explain to a colleague. If a manual method remains safer or simpler, keep it manual.
Before you move on
- The workflow begins with one clearly defined recurring task.
- The output has a written acceptance check.
- The tool and input information comply with workplace rules.
- Human decisions and verification steps are explicit.
- The complete process is compared with the previous method.
- The procedure includes an owner, review point and stop conditions.
Sources and further reading
These official resources informed the guardrails and practical method in this guide.
This preview uses fictional examples where indicated. Read our editorial approach.