What does AI need before it drafts a funder report?

In short: four things: the sources it may use, your data dictionary, the funder's agreement and the funder's taste file. Without them it fills the gaps with plausible guesses.

The first three fix the numbers. The fourth shapes the words.

Add a short brief on top: who reads the report, what they will decide with it, and the period it covers. Frame the question fairly. "What changed for trainees this quarter, and what did not?" leaves room for an honest answer. "Show how our training succeeded" pushes the AI toward only the favorable evidence.

Do not start by uploading everything you have and asking for a strong story. That is how a draft ends up quoting last year's figures, mixing periods or presenting a single participant's experience as typical.

What happens when the AI cannot find a number?

In short: it should say the number is missing and why, never estimate it. A guessed number looks exactly like a found one, and the funder cannot tell the difference.

In a demo dataset, a workforce program with training and job placement workflows, the AI's first draft found total enrolled and female trainees. It flagged three items as missing: training hours delivered, job placements and starting wage by track. The reason was simple to fix: the placements survey had not been selected as a source. The team added it, asked again, and the draft was complete.

This works only because the dictionary told the AI what should be there. Without that list, a draft with two numbers looks finished. With it, the gaps are visible before a program officer finds them.

How do you choose which sources the AI may use?

In short: name the surveys and fields for this report before asking anything, send the fewest personal details the question needs, and make sure every answer points back to its records.

A report on placements needs the enrolment and placement records. It does not need names, emails or case notes. An approved summary table often answers the question with no identified records at all. The Foundations chapter on what the assistant may see covers this in detail.

In Sopact Sense today, the AI Assistant stays locked until you pick which surveys it may use. Field selection lets you choose which fields are sent to the AI, so names and emails stay out. Each team's Assistant sees only its own folder's data. Every line of an answer links to a record you can open. With other tools, keep the same habits by hand: a named export for each report, identifying columns removed, and a note of which files went in.

How do you draft the report with AI?

In short: ask for the funder's format, their words and your numbers, with a source for every figure and a clear flag for everything missing.

A structure most funders can follow: what the reader needs to know or decide; what was delivered against the plan; the outcomes, each with its source and limits; shortfalls and what is still unknown; spending where required; and the next action with an owner. Keep each caveat beside its number, not in a footnote.

Use a participant's words to explain a finding, not to replace it. One account shows what an experience was like; it does not show how common it was. Use quotes only with permission for this audience, and never let the AI write one.

Draft our [PERIOD] report for [FUNDER NAME].

Use only:
- the data sources selected for this report: [LIST SURVEYS OR FILES]
- our data dictionary: [ATTACH]
- our agreement with this funder: [ATTACH]
- the funder taste file and our own taste file: [ATTACH]

Rules:
1. Follow the funder's template and the order in their taste file. Use their words.
2. Report every metric in the agreement, using the definition in our dictionary. Where the funder defines a metric differently, give their number, labelled, and add one sentence on how it differs.
3. Put the source next to every number.
4. If a number cannot be found in the selected sources, write "MISSING: [metric], not in the selected sources". Do not estimate, round up or borrow from another period.
5. Keep each limitation beside the number it limits. Include shortfalls.
6. Do not invent quotations, causes or explanations. Use only quotes marked as permitted.
7. End with a table: claim | number | source | status (found, missing, differs from agreement).

How do you check every number before you send?

In short: open the source of each number, confirm it matches the agreement's definition and period, and resolve or name every gap. Make every number match its source shows the method in full.

Pre-send checklist. Run it on every report, for every funder.

If the report includes spending, have the person responsible for finance explain any variance. An underspend may mean a delayed activity, not efficiency, and the AI cannot know which. A second AI pass that compares the draft with the sources can help find problems, but it can miss the same error twice. It is not an independent check.

An honest limit: AI makes drafting faster, and it makes confident mistakes faster too. The report is still yours. Every number the funder reads is one your team has opened, and every gap is one your team decided how to name.