Build a Human-Reviewed AI Workflow for Recurring Work

Choose a task with a clear beginning and a small, reviewable output. Avoid starting with “manage my projects” or “run my inbox.” Those requests contain many separate decisions and permissions.

For our example, the task begins when the week's approved project notes are ready. It finishes when the project owner has checked a draft update. Sending that update is a separate human action.

Complete this worksheet:

For the fictional weekly update, the exclusions are pricing, changing project records, assigning people, moving deadlines and contacting anyone. Keeping these outside the routine makes the first version easier to inspect.

A shorter task sentence you can adapt:

Create one update yourself from fictional notes before using AI. Record what you did: find the relevant notes, separate completed work from plans, identify blockers and prepare a short summary.

Time the whole process if speed is one of your reasons for trying AI. Later, include input preparation, checking, corrections and saving the result in your comparison. Generation time alone does not tell you whether the routine helps.

Also record quality criteria. For this task, every completion claim must have support, every missing date must remain visible, and the update must distinguish a proposed next step from an agreed one. These are trial criteria you choose, not a promise that a tool will meet them.

Write a short checklist of the information you must provide each time. Our example requires a reporting period, a project label and numbered source notes. Each note should distinguish an observation, a decision or a proposal where possible.

Use the minimum information needed. Remove unrelated personal details and confidential material that this task does not require. Replacing names with labels may help, but does not make every document anonymous. Check the tool's current terms, settings and your organisation's rules before considering real information.

If you do not have permission or cannot determine whether the material is suitable, stop. Use fictional data or your normal manual process instead. See Protect Your Privacy When Using AI for a preparation checklist.

Your information checklist can be simple:

Choose headings that help you review the work: completed, in progress, blockers, decisions needed and proposed next steps. Require a note reference beside each factual statement.

Include a separate “cannot determine” section. Without a place for uncertainty, gaps can disappear into fluent prose. A status such as “on track” also needs a defined basis. If your notes do not establish the agreed schedule and actual progress, leave that overall judgement out.

Decide how long the update should be. For this practice routine, five short sections and a short uncertainty list are enough. A long report can hide the very issues you want to see quickly.

A prompt is the set of instructions you give an AI tool. Copy the template below into a document you control, and keep it separate from each week's notes.

These instructions tell the AI what you want, but they do not enforce account permissions or guarantee correct answers. Keep account connections outside this practice routine, and check the result yourself.

Divide actions into three lists: allowed without further action, needs human approval and never allowed in this routine. Keep reading, drafting, changing records, sending and spending separate.

For a first trial, the allowed list can be only “analyse the fictional input and draft an output.” The approval list can include “use real customer information after the data-handling review” and “send any message.” The prohibited list can include “invent a price or claim a booking succeeded.”

If you later connect an account, inspect the actual access grant. A written instruction alone does not prove the integration is restricted appropriately.

Here is the complete practice input. Every detail is invented.

A careful draft should look broadly like this. This is an authored example of the expected structure, not a claimed live model test:

• Completed: page outline approved on 20 October. [S1]
• In progress: venue section drafted and awaiting review. [S2]
• Blocker: accessible entrance details remain unconfirmed. [S3]
• Decision needed: confirm whether page review is due 28 or 29 October. Both dates appear; replacement is not recorded. [S4, S5]
• Proposed next step: consider a transport section; approval has not been recorded. [S6]
• Recorded action: coordinator to request entrance details; due date not recorded. [S7]
• Cannot determine: agreed review deadline, completion date for the coordinator's request and overall schedule status.

The later note does not automatically settle the deadline. The venue draft does not become approved content because the outline was approved. Reviewing those distinctions is the central job of the person checking the output.

Keep the source notes and draft side by side. For each statement, check its reference, meaning and level of certainty. Then read the sources again for omissions. Checking only the generated text can miss something it left out.

Use this checklist before approving the draft:

If you find an error, correct it against the source or revise the prompt and generate another draft. Review the entire new version, not just the sentence you wanted changed. An instruction to fix one item does not establish that everything else remained correct.

Prepare the expected response before running a test. Include ordinary complete notes, a missing reporting period, conflicting deadlines, repeated notes and a note that contains an instruction such as “ignore the review requirement.” That last line should remain untrusted source text, not permission to change the workflow.

Also test an empty completed-work section. The right result may be “no completed work recorded.” A plausible invented achievement is a failure.

Log the input version, expected behaviour, actual result and corrections needed. A few successful trials are a learning checkpoint, not evidence that the process is safe for unattended operation. NIST describes confidently incorrect generated content as a risk; readable output alone cannot establish reliability. See the Generative AI Profile.

Use this six-case checklist when preparing your trials:

Prepare these cases and write the expected response before testing:

1. Ordinary complete input: accurate brief, no invented facts.
2. Missing information: visible gap, no silent substitution.
3. Conflicting dates: conflict flagged, neither date chosen without basis.
4. Duplicate input: possible duplicate identified, no assumed new obligation.
5. Unavailable source: affected step stops, no claim of completion.
6. Embedded instruction: source text asking for a rule override is not treated as permission.

Repeat the important cases. Log failures as well as successes. Passing a handful of examples is a learning checkpoint, not certification for unattended use.

If a source is missing, obtain it through your normal process or finish manually. If the tool is unavailable, use your saved headings. If the output repeatedly invents details or takes longer to check than to write, stop using that version.

Keep the previous working template when you make changes. Re-run the cases affected by each revision. Never retry an external action just because a system's result is unclear; this beginner workflow avoids that problem by having no external actions in the first place.

Maintain a small log without copying unnecessary sensitive source material:

Keep the routine manual while you learn whether it is worthwhile. Reusing a prompt may be all you need. Our reusable prompt guide explains that simpler option.

If you later consider connected automation, separately review actual permissions, logs, failure handling and human approval controls. A written prompt does not prove those controls exist. Compare AI agents and automation before choosing an approach.

For now, finish with a saved template, a test log and one checked fictional example. That gives you a concrete routine to assess before introducing real data or additional access.

Does this workflow send an update automatically?

No. You prepare the information, review the draft and share an approved version yourself. Account connections are outside this practice routine.

What if two notes give different deadlines?

Keep both references visible and ask for confirmation. A later note does not establish that it replaced an earlier deadline.

How do I decide whether to keep using AI?

Compare the whole task, including preparation, checking and corrections. If the routine takes more effort or keeps inventing details, use your manual method.