Short answer
AI grant evidence assembly is a workflow where a model gathers the figures, quotes and documents a funder asks for from a nonprofit’s own records and drafts an evidence pack with every claim linked to its source. In New Zealand it suits charities preparing funder accountability reports and the non-financial parts of their annual return. Staff verify each figure, check consent and sign off before anything is submitted.
- Every figure links to the record it came from, and code, not the model, calculates it.
- Privacy principle 10 limits reusing case notes as funder stories without consent or de-identification.
- Registered charities file an annual return within six months of balance date.
- Staff decide which outcomes to claim, explain variations to funders and sign off.
What grant evidence assembly does
AI grant evidence assembly answers each funder question with what your records show and exactly where. A model reads the reporting template, searches approved folders and exports, and drafts each answer with a link to the row, document or photo behind it. It does not decide what the programme achieved.
Evidence usually sits in attendance sheets, case management, survey exports, accounts and a shared drive of photos. The slow part is moving between them. For the wider picture, see AI for nonprofits.
How the workflow runs
- Inputs
- The grant agreement and its promised outcomes, the funder’s template, attendance logs, survey exports, spending by tracking category, a consent register, approved stories and photos, and last year’s report.
- Systems
- SharePoint or Google Drive, an attendance spreadsheet or Airtable base, Xero or MYOB, and a language model that can only search approved folders. The case management system stays out of scope.
- Steps
- Map each funder question to evidence types, calculate every number with a script, retrieve matching documents, draft cited answers, and list gaps such as a promised outcome with no measure.
- Human review
- The programme lead checks every figure against its source, a second person checks stories against the consent register, and an authorised officer signs off.
- Outputs
- An evidence table, cited draft answers, a gap list, and approved photos and quotes.
- Measures
- Staff hours per report, figures corrected at review, gaps found two weeks or more before the deadline, and funder follow-up questions.
- Risks
- Invented or inflated numbers, one person counted under two grants, identifiable details in a “de-identified” story, and outcome claims stronger than the data supports.
The rule that matters most: the model writes sentences but never produces a number. Counts and spending come from code run against the source data, and any figure without a source link blocks export. See structured outputs and validation for the mechanics.
A worked example with synthetic data
Illustrative example: a youth mentoring trust in Kirikiriroa Hamilton with 2.5 full-time staff reports to three funders, two annually and one six-monthly, so four reports a year. Assumptions: manual assembly takes about 14 hours per report, the attendance sheet carries a participant ID, and all figures are synthetic.
| Funder question | Evidence source | Draft answer | Human check |
|---|---|---|---|
| How many young people took part? | Attendance sheet, 2026 tab | 62 attended at least one session; 41 attended eight or more | Coordinator checks nobody is counted under another grant |
| What changed for them? | Pre and post survey export | Self-rated confidence rose from 2.9 to 3.6 (of 5) for the 38 who did both surveys | Lead checks the 38 are not described as all 62 |
| How was the grant spent? | Xero tracking category | NZ$48,200 of NZ$50,000 spent; NZ$1,800 transport underspend | Finance checks against the grant agreement |
| Share a participant story | Consent register and approved story folder | A 180-word de-identified story with one quote | Young person and whānau approve the text |
On these assumptions, drafting drops to about 2 hours and review takes about 4 per report: 24 hours a year instead of 56. The larger gain is timing. If the trust promised to measure school attendance and holds no data, the gap list says so in month two, not deadline week.
New Zealand rules that shape the evidence
- Annual returns. Under section 41 of the Charities Act 2005, registered charities file an annual return within six months of balance date, with financial statements. The External Reporting Board puts not-for-profits with total expenses of $5 million or less (and operating payments over $140,000) in Tier 3, and those with operating payments under $140,000 in Tier 4. Tier 3 reporting includes non-financial information about what the entity achieved, so one evidence table can serve funders and the annual return.
- Reusing client information. Information privacy principle 10 stops personal information collected for one purpose being used for another, unless an exception applies, such as the person’s authorisation or use in a form that does not identify them. Case notes gathered to deliver a service are not raw material for funder stories, so stories enter only through the consent register.
- Māori data. The Privacy Commissioner expects agencies adopting AI tools to engage with Māori about risks to the taonga of their information. Kaupapa Māori providers should agree how outcome data is described before any model drafts it.
What stays with people
Staff decide which outcomes to claim and how strongly. “Contributed to” and “caused” are different claims, and a model drifts towards the stronger one. People also hold consent conversations with participants and whānau, explain underspends and changes of plan to the funder, and sign the report.
Where this goes wrong
| Failure | What it looks like | Control |
|---|---|---|
| Filled gap | A plausible number appears where no data exists | Numbers only from calculated fields; unsourced figures block export |
| Double counting | One participant claimed by two grants | Participant ID and a de-duplication check per grant period |
| Re-identification | A “de-identified” story mentions a small-town school and an age | Reviewer checklist for indirect identifiers |
| Small samples | “75% improved”, from eight responses | Every percentage printed with its sample size |
Reviewers need the habit taught in summarising documents with citations.
What it costs and where to start
The cost sits in connecting sources and building the review screen, not model usage. Drivers: how many systems, whether a participant ID exists, how many funder templates differ, and how much cleaning the attendance data needs. A data inventory of where each kind of evidence lives is a cheap first step.
A good pilot takes one report you have already submitted and checks whether the system reproduces its figures from raw records. Where it disagrees, find out who was right. Our AI document processing work covers extraction and the review queue. The same consented stories can feed donor communication drafting, and volunteer hours logged through volunteer enquiry handling often appear in funder reports as in-kind support.
Sources
- Charities Act 2005 (version as at 6 June 2026), New Zealand Legislation, 2026
- About reporting Tiers and reporting sectors, External Reporting Board, accessed 19 October 2026
- At a glance: Tier 3 Not-For-Profit Reporting Requirements (consultation document), External Reporting Board, May 2022
- Principle 10: Limits on use of personal information, Office of the Privacy Commissioner, accessed 19 October 2026
- Artificial intelligence and the Information Privacy Principles, Office of the Privacy Commissioner, 2023
Written by the Levy Street team in Wellington. Last updated 19 October 2026.