An AI-generated weekly report is ready to share only after a human checks its claims, calculations and commitments against the source material. The most common errors are not strange sentences. They are ordinary-looking statements that turn “ready for review” into “complete”, a proposed date into a promise, or a missing cause into a confident explanation.
This is a fictional workplace exercise. It shows a repeatable workflow without asking you to upload private project records to an AI tool.
The five-minute checking workflow
Use this order each time:
- Define the decision. Write down what the reader needs to decide or do after reading the update.
- Limit the source pack. Include only the material you are allowed to use. Keep confidential records in an approved system.
- Ask for a draft with evidence labels. Tell the assistant to separate facts, proposed actions and open questions.
- Recalculate the numbers yourself. Confirm the denominator and what the percentage measures.
- Check status, owners and dates. “Ready for review”, “approved” and “complete” are different states.
- Remove unsupported explanations. If the source does not name a cause, owner, approval or customer reaction, write “not stated”.
The final reviewer should be able to trace each important sentence back to a source line or calculation.
A fictional source pack
Imagine you are preparing a weekly update for a workshop signup refresh. The permitted source notes say:
| Source | Verified information |
|---|---|
| S1 | There are 24 items: 18 ready for review, four in progress and two blocked. Categories do not overlap. |
| S2 | No items have final sign-off. |
| S3 | Both blocked items need an attendance-limit decision. The reason that decision is missing is not recorded. |
| S4 | 6 November is a proposed release date, not an approved date. |
| S5 | No client feedback has been collected. |
| S6 | The project owner requests reviewer assignments and the attendance-limit decision by 2 November. |
The numbers reconcile: 18 + 4 + 2 = 24. The calculation 18 ÷ 24 = 75%, but that is the share of items ready for review. It is not proof that the project is 75% complete.
The illustrative first draft
Here is a plausible but unsafe draft an assistant might produce:
The project is 75% complete, with positive client feedback. The team is on track to launch on 6 November after resolving a supplier delay. All remaining work has owners and should be finished by 2 November.
The wording sounds tidy. It is also wrong in several different ways.
Check each claim, not the general impression
| Draft claim | Evidence available | Verdict | Safer correction |
|---|---|---|---|
| The project is 75% complete | 18 of 24 items are ready for review | Mislabelled | “18 of 24 items, or 75%, are ready for review. No items have final sign-off.” |
| There is positive client feedback | No client feedback has been collected | Unsupported | Remove the sentence. |
| The team will launch on 6 November | 6 November is proposed, not approved | Overstated | “6 November is a proposed release date.” |
| The blocker is a supplier delay | The reason is not recorded | Invented cause | “Both blocked items need the attendance-limit decision. The reason for the missing decision is not stated.” |
| All remaining work has owners | The source asks for reviewer assignments | Unsupported | “Reviewer assignments are requested by 2 November.” |
The checked version is less dramatic:
Eighteen of 24 items are ready for review, which is 75% of the current item list. Four items are in progress and two are blocked. No items have final sign-off. Both blocked items need the attendance-limit decision. Reviewer assignments and that decision are requested by 2 November. 6 November is a proposed release date, not an approved commitment. Client feedback has not been collected.
That version gives the reader something useful to act on without pretending that the evidence says more than it does.
Prompt for a bounded draft
You can ask an approved assistant to organise the source pack, but keep the boundary explicit:
Create a weekly project update using only the source pack below.
Separate the output into:
1. Verified status, with calculations shown
2. Decisions or actions needed, with owner and date only when stated
3. Open questions and limitations
4. Proposed dates, clearly labelled as proposed
If a fact, cause, owner, approval, customer reaction or deadline is missing,
write "not stated". Do not infer project completion from the number of items
ready for review. Do not turn a proposed date into a commitment.
Source pack:
[paste only material you are permitted to share]The prompt reduces guessing, but it does not replace the six checks. Read the result alongside the source pack and recalculate important figures outside the assistant.
Use an evidence table in the final report
For recurring updates, keep a small table behind the prose:
| Claim | Source and date | What was checked | Limitation or open question |
|---|---|---|---|
| 18 of 24 items are ready for review | S1, reporting date | 18 + 4 + 2 = 24; 18 ÷ 24 = 75% | Ready for review is not final approval |
| Attendance-limit decision is needed | S3 and S6 | Both blocked items require it | Decision owner is not named |
| 6 November is proposed | S4 | Confirmed wording is “proposed” | Approval is missing |
This is more useful than a confidence score. A reader can see what was checked and what still needs a person to decide.
What not to paste into an AI tool
Do not upload a real project pack merely because it would make the example more realistic. Follow your organisation's approved-tool and data rules. Remove personal details only when your policy says that is enough. Names, dates, contract terms, customer information and combinations of harmless-looking fields can still be sensitive.
When in doubt, practise with a fictional pack like the one above, or describe the structure without copying the underlying records. Our guide to confidential work information explains the questions to ask before sharing anything.
Download the practice template
The weekly report template and practice source pack includes the editable headings, reconciliation checks and the fictional facts used in this example. You can also practise the wider workflow in the free lesson on reports and summaries, or follow the professional learning path.
The goal is not to make every update sound polished. It is to make each important sentence traceable, each calculation reproducible and each unresolved decision visible.