Let AI draft the period report narrative, not decide it

The numbers in a period report come out of the tools. The narrative doesn't, and it's usually written late on the day before the deadline. That narrative is one of the best places to put AI to work, as long as it only writes what the data supports.

Where the time actually goes

Most controls teams can produce a variance table quickly. The slow part is turning it into sentences a project board will read: which variances matter, what caused them, what's being done, and what it means for the forecast. That takes judgement, but a lot of it is also just drafting, and drafting is what language models are good at.

The risk is that a model will fill gaps with plausible-sounding causes. "Delays in supplier deliveries" reads well in a report whether or not it's true. So the setup has to stop that from happening.

Give it structured inputs, not screenshots

Export what you'd normally look at into plain tables. For a typical monthly report that means:

• Cost and schedule variance by control account, with the thresholds that make a variance reportable.
• Milestone movements since last period: baseline, last forecast, this forecast.
• The causes you already know, as short notes from the control account managers (CAMs). Even one line each.
• Changes to the top risks and any approved or pending change requests.

The CAM notes are the part people skip. Without them the model has no causes to work from, so it either invents them or writes vague filler. With them it can do what it's good at, which is turning rough notes into clear, consistent prose.

A prompt pattern that holds up

The structure matters more than the wording. Tell it who the reader is, what to produce, and, most importantly, what it isn't allowed to do.

You are drafting the narrative section of a monthly project report
for a project board.

Inputs below: variance table by control account, milestone movements,
CAM notes on causes, and risk/change updates.

Write:
1. A 3-sentence summary of overall cost and schedule position.
2. One short paragraph per control account that breaches a threshold:
   the variance, the cause, the recovery action, the forecast impact.
3. Milestones that moved more than 2 weeks, with the reason.

Rules:
- Use only figures from the inputs. Do not calculate new ones.
- Only give a cause if a CAM note states it. Otherwise write
  "Cause not yet confirmed".
- If inputs conflict, list the conflict at the end instead of
  choosing one.
- Plain English. No adjectives like "significant" without a number.

[paste tables and notes here]

The "cause not yet confirmed" rule is the important one. It makes gaps in your information visible instead of hiding them under confident prose. A board would rather read that than a guess.

This prompt is also in the prompt library.

Check it like you'd check a junior's draft

Treat the output as a first draft from someone capable who wasn't in the room. Before it goes anywhere:

1. Trace every number in the text back to the input table. Models occasionally transpose figures or round them differently.
2. Check every stated cause against a CAM note. Anything without a source comes out.
3. Read the summary on its own. It's the part people read, and the part where tone drifts most easily into optimism.
4. Ask whether the CAMs would recognise their own account. If not, the draft has smoothed over something.

What you get

Done this way, the time goes into getting good cause notes from CAMs and reviewing the draft, not into writing paragraph after paragraph. The writing gets more consistent from month to month, which makes trends easier to spot. And because every sentence has to trace to an input, gaps in the information show up early rather than in the board meeting.

Before you try this. Cost and schedule data is often commercially sensitive, and on some programmes classified. Use only AI tools your organisation has approved for that level of data, and follow its rules on what can be shared.