Whether to buy more accounts sits in the Friday report's total minutes

The decision in this example is not to buy more accounts.

The warehouse team’s Friday draft wrote “80% complete” with no source sentence, and the 18 September test milestone stayed as [NEEDS CONFIRMATION] because the notes do not say who grants access to the warehouse. The first measurement is not a smoother sentence. Three older reports, written by hand, took 100, 119, and 101 minutes. These are simulated numbers, adding up the time to gather the input, draft, check, and ask about what was missing. If the writing gets 25 minutes shorter and the checking grows by 30, the total does not fall.

This is for a team lead or the owner of a small business, before paying for several people to share one tool. You need one person who is allowed to send the final version, two to four older reports that were already approved, and a tool the organization allows. Do the work on data that has already been anonymized.

The draft shows where it has to stop

The passage below was written to show where a mark has to go. It is not the output of one run on a specific tool:

3. Obstacles that need a decision
- Data entry is slow because warehouse A access has not been granted.
  Effect: the 18 September test milestone may slip.
  Decision needed: [NEEDS CONFIRMATION: who grants access, and the deadline].

Figure with no source:
- "80% complete" — no source in the input.

What is worth keeping is the [NEEDS CONFIRMATION] mark, not a smooth sentence. If the tool fills in who grants access, or turns “may slip” into a commitment the group already agreed, that draft has not passed the trial.

Add four stretches of time on your own report

For each older report, write down the real minutes for four jobs: gathering the input, drafting, the responsible person checking and editing, and asking for information that is still missing. The total is those four stretches. If two people sit for 15 minutes to edit, count 30 minutes of labor. Record the minutes spent preparing the input separately. A draft that is typed faster is not a gain if fixing the input takes longer than the time just saved.

Use only one kind of document, for example the progress report sent every Friday. Do not mix client email, a revenue forecast, and comments about staff into the same trial. At most two people operate it. Do not put salary, health information, unpublished contracts, client secrets, or passwords into the tool.

The NIST AI Risk Management Framework asks for clear roles and for measures tied to the actual task, repeatable, and written down. For a small team, write the name of the person allowed to send the final version and the name of the person allowed to stop the trial.

The input should use the same frame, so minutes of preparation are not mixed with minutes of writing:

PROJECT/TEAM:
REPORT PERIOD:

1. Finished
- Task:
- Evidence or link:
- Confirmed by:

2. In progress
- Task:
- Status:
- Next milestone:
- Owner:

3. Obstacles
- Problem:
- Effect:
- Decision or help needed:

4. Figures
- Metric:
- Value:
- Source:
- Comparison period:

Same input, two versions in the first week

In the first week, make one version by hand and one version with the tool, from the same notes. If you can arrange it, the person who reviews them does not know which one had machine help.

Create a DRAFT weekly report using only <data>.
Do not fill in numbers, deadlines, owners, or causes on your own.
If something is missing, write [NEEDS CONFIRMATION].

Structure:
1. Results that are finished
2. Work in progress
3. Obstacles that need a decision
4. Figures and their sources
5. Work for next week

Every number must have its source in parentheses. Keep owners
and milestones as they appear in the input. At the end of the draft, list missing data,
sentences that are easy to read as a commitment, and numbers with no source.

<data>
[PASTE INPUT YOU ARE ALLOWED TO USE]
</data>

This is a prompt to start from. It has not been run on one fixed tool. Run it first on older reports that have been anonymized, where you already know which sentences are right. If the tool adds data, do not move on to more sensitive documents.

When the draft is wrong

Keep the first version, mark the errors, then ask it to check against the source:

Check every number, owner, and deadline in the draft.
Make a list: what was written; the source sentence inside <data>;
whether the conclusion has a source or does not.
Anything without a source must be changed to [NEEDS CONFIRMATION].
Do not add guesses.

Write the minutes of that correction into the total. If the same error comes back after one change to the prompt and the input, narrow the kind of report or stop. Stop immediately if sensitive data was sent to the wrong place, if the draft touches money or personnel without a reviewer, or if nobody still has time to read each draft.

Count the errors by hand. You do not need a dashboard: a number the tool added, the wrong person assigned, an obstacle left out, something written as certain that the notes left uncertain, and data exposed. Exposure is a reason to stop, not a reason to tweak the wording and keep going.

The NIST Generative AI Profile says that a check in one narrow situation may not hold when the work changes. Two weeks answer only whether it is worth trying again on this same kind of report.

Write the thresholds before you see a flattering number

The marks below are not an industry standard. Write them down before you look at the results from the week that used the tool:

• Continue when total time falls by at least 20%, no serious error gets past the reviewer, and the reviewer can still handle the number of drafts.
• Revise and try once more only when time falls but a repeated error is tied to the input or the prompt, and you change only one thing.
• Stop when the total does not fall, when you have to rewrite most of it, when an important error gets through, when there is a data risk, or when there is no reviewer left.

The decision slip only needs these lines: kind of report, number of drafts, tool and cost, total minutes before, total minutes after, repeated errors, minutes of the reviewer, data risk, the decision to continue or revise or stop, the scope for next week, who signs, and the date to look again.

If you continue, widen by one step only: one more team doing the same kind of report, or two more weeks. This trial decides whether to buy or not to buy from the total minutes, the errors, and the risk you wrote down. It does not prove that the tool is useful in general.

Sources

1. NIST AI Risk Management Framework 1.0 — roles, responsibility, and measurement tied to the actual task; published January 2023.
2. NIST AI 600-1, Generative AI Profile — checking before a rollout, the limit of one narrow trial, and the person who still has to supervise; published July 2024.