Using AI on the grant report without putting a client in it

The report is due Friday

The program data is in a spreadsheet, a shared drive, and three people's heads. The funder sent six questions in a portal that times out. Every executive director knows this week. The question is not whether AI can help write the report; it can, and well. The question is how to do it without a single client record ever touching the tool.

Step one: the aggregate, written by a person

Before the tool sees anything, a person writes the summary. Not the log, the summary. It looks like this: fourteen intakes this quarter, nine for housing and five for food; average wait from intake to first service three days; 212 meals served; two households moved to stable housing and are still housed at 90 days; one program coordinator hired in August. Numbers and categories only. No names, no initials, no dates of birth, no case detail, nothing that could identify a person served.

This is the rule that makes everything else safe, and it is the first hour of every session I run. The tool is for the form and the process, never the person.

Step two: the funder's questions, in the funder's order

Paste the funder's questions exactly as written. Funders read hundreds of these; a report that answers question three under question one gets skimmed. The draft should match their structure one for one, and it should say so when the data doesn't support an answer instead of inventing one.

Prompt: the funder report from an aggregate summary

Here is our organization brief: [paste] Here is a one-paragraph description of the program this report covers: [paste your program brief] Here is the aggregate activity summary for the reporting period, written by our staff. It contains counts and categories only, no individual records: [paste] Here are the funder's report questions, exactly as written and in their order: [paste] Draft our report. Answer each question under its own heading, in the funder's order and wording. Use only the numbers in the summary; do not calculate new figures or invent any. Where the summary does not support an answer, write [NEED: what is missing] instead of guessing. Plain English in our voice, no jargon, under [word limit] words. After the draft, list every number you used and where it came from in the summary.

Step three: the check

Read the draft against the summary. Every number should trace back. Every claim about an outcome should be one the data supports; "212 meals served" is an output, "families are more food secure" is an outcome you can only claim if you measured it. Fix the softened language, fill the [NEED] flags from the source, and send it to the program officer with a sentence that says which numbers are final and which are preliminary.

What this saves, and what it doesn't

The writing goes from an evening to an hour. The data work doesn't change: someone still has to count, and the summary still has to be true. What the tool removes is the blank page and the eleven o'clock rewrite, and what the rule removes is the risk of a client's name in a chat window.

This is lesson three of module four in the Adept Library, which comes with every Working Session. The session itself does this on your real report, with your staff, and ends with a one-page policy your board can approve.