How to Keep a Decision Log When You Use AI at Work

An open notebook and pencil sit next to loose papers and a closed laptop on a meeting table.

Three weeks after a community centre changes its closing checklist, a volunteer asks why the team kept paper sign-off. The project chat contains an AI summary, two versions of the checklist and a note that the team liked the new process. The person who approved it is away. Nobody can tell whether paper was a temporary measure or a condition of the rollout. A decision log gives the team somewhere to find that answer.

Start the entry when the decision is made. A brief record can capture the choice, the people responsible, the evidence they checked and the conditions they agreed to revisit. A colleague should be able to read it without hunting through an AI conversation or asking someone to remember the meeting. The example below is fictional and uses an ordinary operational decision.

Record the choice and the authority to make it

Write the date, the named decision owner and one sentence describing what was agreed. In the community centre example, the facilities coordinator approves a two-week trial of a digital checklist for one building, with paper sign-off kept during the trial. Add the options the team seriously considered, such as keeping the existing checklist or switching both buildings at once. A sentence about why each option was set aside gives the next reviewer enough context to understand the trade-off.

Include the limits of the decision. A trial at one building gives permission to try the checklist there for two weeks. The entry should also specify who can extend the trial. When the group settles on a wider rollout later, add a dated entry linked to the trial decision. Keeping the earlier entry intact makes it possible to see when the scope changed.

Show which evidence the team actually checked

Give each important claim a source the reviewer can find. The coordinator might verify that the duty team has access to the digital form by watching two volunteers open it. In the log, name the access check, date it and record what it covered. A sentence saying that all volunteers can use the form would go beyond those observations. Write the remaining question separately, such as whether volunteers on other shifts have the same access.

Record AI help in plain language. For example, write that the assistant grouped comments from the rehearsal and drafted a comparison of the rollout options. Then note that the coordinator checked the relevant comments against the original rehearsal notes. The NIST Generative AI Profile describes the risk of confidently presented false content and recommends checking sources and citations in generative AI outputs. Keep links to the original evidence in the team workspace and use approved access controls for internal records.

Write down which assumptions could change the choice

A useful assumption has a consequence. Suppose the coordinator expects the network to be available at closing time. The team decides to keep paper sign-off so volunteers can finish when the network is unavailable. Write down who will check the assumption and what would count as a reason to change the process. One possible rule is to keep paper sign-off if any trial shift cannot open or submit the form. The rule belongs to the team and can be evaluated from the trial records.

The UK government’s AQuA Book provides guidance on recording analytical assumptions and communicating uncertainty. For a small operational decision, a few lines can do the job. If a proposed decision still has several untested assumptions, use our guide to stress-testing a plan with AI to turn concerns into checks before recording approval.

Give the next reviewer a date and a useful question

Set the review for the end of the two-week trial and assign it to someone who can access the original records. For the community centre, the review question is whether every trial shift could open and submit the checklist. Ask the reviewer to record which shifts were checked and what happened during interruptions. If a shift has no record, ask the shift lead what happened before treating it as a success.

At the review, append the result and the next decision. If the team extends the trial, record the new end date and the reason. If it returns to paper, record what prompted the change. A short entry can use these fields in order: date, owner, decision and scope, alternatives, verified evidence, AI contribution, unresolved assumptions, review date and outcome. Use our guide to keeping a human in the loop at work to decide who checks the AI output before it reaches the person making the decision.

Sources

  1. NIST Generative AI Profile
  2. The AQuA Book