AI Project Status Report Checklist (Fact-Check the Draft)

Status reports fail when they sound finished but the evidence does not. AI can turn project records into a readable first draft in minutes. That only helps if every material claim still traces to a source before the pack is sent.

Definition

AI-assisted status report: A project update drafted with AI from supplied records, then checked and approved by a responsible person before circulation.

In practice: Trace each material claim to its source before approving the report.

Not the same as: An automatically verified account of project delivery.

This article is a governed workflow for one reporting cycle: gather approved inputs, draft from bounded evidence, fact-check every colour, date, owner and number, then stamp the result yourself. For where AI helps and fails across status, RAG, RAID and SteerCo work more broadly, start with AI in project management. This piece does not repeat that pillar’s task-by-task prompts.

Gather required inputs before prompting

If the model only sees a chat message and last week’s slide, it will invent a tidy story from thin air. Collect the artefacts first. Do not prompt until you can tick the list below for the reporting period.

Approved plan baseline, current forecast and reporting period — including the cut-off date and whether dates are baseline, forecast or actual.
Open RAID items — owners, due or review dates, and evidence of recent changes. If your register uses Assumptions and Dependencies, or Actions and Decisions, keep the letter meanings consistent with the live log (see the RAID log template).
Last period’s actual results and source artefacts — schedule extracts, cost files, change log entries, test evidence or similar records you would defend in the room.
Decisions needed — approving authority and decision deadlines, written as asks rather than assumed outcomes.
Agreed RAG criteria and evidence dates — so colour language matches the rules you already use (see computable RAG status criteria).
Information-handling approval — confirmation that the chosen tool and the data you will paste are permitted for this project.

Missing inputs are not a reason to “let the model fill gaps”. They are a reason to delay the draft or to leave the claim out of the circulated report.

A practical way to keep the pack honest is to paste extracts, not whole repositories. Prefer a RAID export of open and recently changed rows, a one-page schedule extract for the critical path, and the cost summary the finance partner already signed off. If two sources disagree, resolve the conflict with the owner before you ask the model to narrate either version.

Draft from bounded evidence

Separate three layers in the prompt and in the draft you accept:

Source facts — dated records you supplied.
Interpretation — what those facts mean for delivery confidence this cycle.
Decisions requested — asks that still need an authorised person.

Ask for a draft based only on the records you paste. Require evidence references beside material claims, and instruct the model to flag unsupported statements for review rather than inventing them. One short illustrative prompt is enough for this workflow; use the broader prompts in the AI project management guide when you need task-specific wording.

Draft a status report for [audience] covering [reporting period], using ONLY the records below.
Structure: position, changes since last report, impact, decisions needed, next actions.
For every material claim, add a short evidence tag (e.g. RAID-I12, schedule extract 12 Oct, cost file EAC row).
If a claim is not supported by the records, write [NEEDS REVIEW] instead of inventing detail.
Do not invent RAG colours, owners, dates, percentages or commitments.
Records:
[paste baseline / forecast / actuals / RAID extract / decisions needed / RAG criteria]

Treat every [NEEDS REVIEW] marker as work for you, not for the reader. Resolve the gap from a source artefact, or remove the claim before the report leaves your desk.

Keep the first draft short enough to check in one sitting. A long narrative that restates the whole plan is harder to verify and tempts the model to fill gaps with plausible filler. Prefer a clear position statement, a short list of material changes, impacts tied to evidence, and a small set of decisions with owners and dates.

Illustrative example (invented, not a client case): the schedule extract shows test environment access still unconfirmed, while an older milestone line still says “on track for 18 November”. A useful draft states both facts, tags the issue and the milestone, and asks for confirmation. It must not report that testing will start on 18 November.

Check every material claim against a source artefact

The draft is not the report until you have checked it. Walk the checklist below line by line. If you cannot point to a source, the sentence does not go out.

Every RAG colour — apply the agreed criteria to dated evidence; do not accept a colour justified only by tone.
Every date — distinguish baseline, forecast, actual and requested dates. A forecast written as a commitment is an error.
Every owner — confirm the named person still holds that responsibility.
Every number or percentage — check source, period, calculation and denominator where relevant.
Every “on track” claim — identify the evidence that supports delivery confidence for this cycle.
Every issue, risk and decision — check current status and preserve genuine uncertainty; do not let smooth wording close an open item.
Every commitment — confirm authority and approval. Generated wording cannot create agreement.

Also scan the source pack for material items the draft omitted. A fluent report that drops an awkward issue is worse than a slower, complete one.

Write corrections into the report file you will circulate, not only into the chat. If you change a RAG colour, note the criterion and evidence date beside it in your working papers. If you rename an owner, confirm with that person or with the RACI the project already uses. Speed gained in drafting is wasted if the checking notes disappear when the chat window closes.

When not to use AI for status reporting

Skip the AI draft — or stop after gathering inputs — when any of these apply:

Confidential information without the required approval for the tool and data combination you would use.
Binding commitments that require an authorised person’s judgement and agreement. AI may help you phrase an ask; it must not phrase the answer as already agreed.
Inputs that are too unreliable or contradictory to support a defensible report. Fix the records first. A polished narrative on broken inputs is watermelon reporting with better grammar.

Human approval and a correction trail

Name the person responsible for the final report. Record approval under the project’s normal process — the same way you would for a hand-written pack. Resolve contradictory sources with owners before you send. Keep supporting references with the working papers for the cycle.

If an error reaches distribution, correct it through the agreed reporting route. Do not rely on a later chat history as the audit trail.

When the bottleneck is the reporting process itself rather than the draft wording, use how to reduce project reporting time beside this checklist.

Used this way, AI is a drafting aid inside an existing reporting discipline. It does not replace computable RAG rules, a maintained RAID register, or the person who owns the pack. The value shows up when total effort — drafting plus checking — falls, while the evidence trail stays intact.

Copyable checklist for one reporting cycle

Inputs gathered and permitted for the chosen tool.
Draft produced only from those records, with unsupported claims flagged.
RAG, dates, owners, numbers, “on track” claims, RAID items and commitments checked against sources.
Omissions scanned; material gaps closed or removed.
Responsible person named; approval recorded; pack issued.

FAQ

What is an AI project status report?

It is a status update whose first draft was produced with AI from project records you supplied. It becomes a real status report only after a responsible person has checked material claims and approved circulation.

Can ChatGPT write a project status report?

It can draft one if you supply bounded records and require evidence tags. It cannot verify delivery, invent missing facts safely, or approve the pack. Use an organisation-approved tool and keep human fact-check and approval.

How do I fact-check an AI-written status report?

Trace every RAG colour, date, owner, number, “on track” claim, open issue or risk, and commitment to a source artefact. Scan the sources for material omissions. Remove or resolve anything that cannot be supported.

Using AI to Draft Status Reports Without Losing the Facts