Skip to content
·7 min read

how a nine-person charity got nine director-days back from funder reporting

ainonprofitgrant reportingautomationottawaworked exampleai adoptioncharity operations

this is an invented example. no client of ours is described here. we have written it as a full project (scope, build, timeline, cost, result) because "ai can help with grant reporting" is useless and a specific build you can argue with is not.

the research behind the time savings is on our ai adoption page, with the papers.

the situation

rideau valley literacy. nine staff, about $900k a year. adult literacy and newcomer language support, six funders: two foundations, one provincial program, three smaller grants.

maya is the executive director. she writes every funder report herself, because the numbers have to be right and she is the only person who holds the whole picture.

a reporting week looks like this. she opens the finance system for expenditure, the CRM for participants, a shared spreadsheet the program coordinator maintains for attendance, and last quarter's board pack to check she is being consistent. the three systems disagree about how many people were served. the CRM counts registrations, the spreadsheet counts attendance, finance counts anyone a cost was allocated to.

she reconciles them by hand, every time, and makes the same judgement calls she made last quarter without any record of what those were.

two director-days per report. six reports a year. twelve days.

that is five per cent of her working year, spent by the most expensive person in the organisation, and roughly three quarters of it is not writing.

where the two days actually go

activityshare of the two days
pulling numbers out of four places½ day
reconciling systems that disagree½ day
writing prose around the numbers¾ day
formatting into the funder's template¼ day

only one of those four is writing.

what was actually built

three things. the first is not ai, and it is the one that matters.

1. one definition of a participant (weeks one and two)

we sat with maya and the program coordinator and wrote down, in plain language, what counts:

  • a registrant is anyone who completed intake
  • a participant is a registrant who attended at least one session
  • an active participant attended three or more sessions in the quarter
  • someone who attends two programs is one participant, counted once, attributed to the program they attended most

this took two sessions and one argument. it is the single most valuable deliverable in the project and it contains no ai whatsoever. maya would keep it if she fired us in week three.

2. one reporting view (week three)

a scheduled pull from the three systems into a single table, applying those definitions the same way every time. it runs nightly. when the systems disagree, the view records that they disagreed and by how much, rather than silently picking one.

maya can now answer "how many people did we serve last quarter" in four seconds instead of half a day, before any report is written.

3. the draft (weeks four and five)

for each funder, a draft is produced from:

  • the figures, retrieved from the reporting view, never generated
  • that funder's template and last report, so the structure and tone match
  • the program coordinator's notes from the quarter

maya opens a draft with the numbers already in place and the prose already written in her voice, pulled from her own past reports.

she edits and signs. every time.

how long it took

weekwhat happened
1half a day with maya and the coordinator. first pass at definitions.
2definitions checked against last year's four reports. two were found to have used different participant counts for the same quarter.
3reporting view built. nightly pull from finance, CRM and the spreadsheet.
4drafting built against maya's past six reports.
5first real report drafted and compared line by line with how maya would have written it.
6tuning. the draft was overclaiming, it described a flat quarter as "strong growth". we fixed the prompt and the tone guidance.

live in five weeks. trusted by week eight, after she had checked two reports against her own working.

what changed

beforeafter
director-days per report20.5
days a year123
time to answer "how many did we serve?"half a dayseconds
consistency between reportsby memoryby definition

≈ 9 director-days returned.

the days matter, but the thing maya noticed first was different: she stopped dreading reporting weeks. the work that made it unpleasant was the reconciliation, and that part is now done before she sits down.

the second thing: the definitions project surfaced that two reports last year had used different participant counts for the same quarter. nobody had noticed. that is a finding worth more than the nine days.

what it would cost

honest numbers for a build this size:

  • the project: $8,000–12,000, weighted toward weeks one to three
  • running it: roughly $30–60 a month
  • payback: nine director-days a year against a director's loaded cost, before counting the grant applications that get written because there is time to write them

in ontario, the planning portion may be eligible for a matched grant through the digital competence centre, which changes the arithmetic.

the one rule that makes it safe

every figure comes from the reporting view. the model never produces a number.

a language model asked for a figure will produce one, confidently, and it will look exactly like a real figure. in a funder report a fabricated number is not an embarrassment, it is a finding, and findings end funding relationships.

so numbers are retrieved and inserted. the model writes the sentences around them. if a figure is missing, the draft says it is missing rather than estimating it.

what we would not build

  • the narrative judgement. which story this funder needs, what to lead with, what a quiet quarter means. that is maya's job and it is why she is the director.
  • auto-send. the report leaves with a human name on it, after a human has read it.
  • anything touching the funder relationship. the email, the call, the awkward conversation about an underspend.

when this is not worth doing

if your reports are genuinely one draft and done, this saves you very little. the organisations where this pays are the ones where reporting means a week of chasing numbers across systems that disagree.

if you have one system and it is clean, skip to the drafting step and the project gets much cheaper.


if you want this worked out with your real numbers rather than maya's invented ones, that is what a readiness review is: a written assessment, fixed fee, with the arithmetic done on your actual reporting load.

other worked examples: quoting in a trades business · intake and referrals in a clinic