How Can NGOs Write Donor Reports Faster With AI?

TL;DR:

• The fast and safe method is to build an evidence file first, clean out personal details, then ask AI to draft only from that file.
• AI helps most with structure, first drafts, shortening and checking gaps. People still own every number, claim and judgement.
• Few donors have published rules about AI in reports. Spencer Foundation and MacArthur Foundation are two that have. Your grant agreement and programme officer come first.
• Tell the donor when AI did more than light editing, or when it made an image, audio or video.
• Do the sums in Excel, check every sentence against your evidence, and never let AI add a result you did not measure.

Who this guide is for

This guide is for programme officers, M&E staff, grants managers and country directors who write narrative reports for funders. It suits a small NGO with one report writer as well as a larger team reporting to several donors.

You need no technical background. You need a report that is due, the notes behind it and a tool your organisation allows.

What donors have actually published about AI in reports

Most published funder rules cover applications, and only a few cover reports. As of November 2026, the two clearest rules we could check that reach reporting are from the Spencer Foundation and the MacArthur Foundation.

We also searched for public AI rules on grant reporting from the UK's FCDO and the EU's international partnerships department, and found none. That does not prove none exist. A rule can sit inside a grant agreement, a reporting template or an email from your programme officer.

So use this order:

1. Read your grant agreement and reporting guidelines. Search them for "artificial intelligence", "AI", "automated", "confidential" and "data".
2. If they are silent, ask your programme officer. A sample email is further down.
3. If you still have no answer, follow the strictest common practice: keep personal data out, check everything, and be open about any use beyond light editing.

Why AI-assisted reports go wrong

AI tools write confident sentences even when the facts are wrong. A report can look finished and still contain things that never happened.

A well-known case came from Australia. In 2025 Deloitte produced a review for the Department of Employment and Workplace Relations under a contract of about AU$440,000. A University of Sydney academic found references that did not exist and a made-up quote from a court judgment. The corrected version removed more than a dozen bad references and footnotes, and it disclosed that a generative AI tool had been used. Deloitte agreed to repay the final instalment of the contract. The department said the substance and recommendations of the review had not changed.

An NGO has less money and less reputation to spare than a global firm. A single invented beneficiary number can lead to awkward questions at your next audit or renewal.

The second problem is blandness. If you type "write a quarterly report for our education programme", the tool has nothing specific to use. It fills the page with phrases that could describe any project anywhere. Your donor has read those phrases many times.

Both problems have the same cure: give the tool real evidence, and tell it to stop when the evidence runs out.

The six-step method

Good AI reporting starts weeks before the deadline, when you collect the evidence. Here are the six steps we suggest.

Step 1: Keep an evidence file all period

Create one folder per grant, and add to it every week. It should hold:

• Activity log: what was planned, what happened, what changed and why.
• Attendance totals, split the way your donor needs them.
• Your indicator table with baseline, target and actual, and the source of each number.
• Field updates, including the useful ones sent on WhatsApp.
• Quotes and stories with written consent on file.
• Budget notes that explain variances.
• The donor's last comments on your previous report.

Fifteen minutes each Friday is cheaper than a week of trying to remember what happened in July.

Step 2: Clean the evidence

Remove names, ID numbers, phone numbers, exact home locations and case details before anything goes near an AI tool. Check that the remaining details cannot point to one person. In a small ward, "a 71-year-old widow who runs the kiosk" is an identification.

Our guide to data protection and AI in Kenya was written for public bodies, but its before-and-after cleaning example works for NGO case notes too.

Step 3: Give the tool your template and your voice

Paste the donor's headings and word limits. Then give it a past report you are proud of and ask for a style sheet (prompt 1 below).

Step 4: Draft one section at a time

Short sections are easier to check. Tell the tool to use only your evidence, and to mark every gap with [CHECK].

Step 5: Trace every claim

Run the claim check in prompt 3. Then have a person trace each number and name back to the evidence file. The tool can miss its own errors, so its check does not replace yours.

Step 6: Final read by the person who ran the programme

Only that person knows what the donor asked about last time, which partner relationship is delicate, and what the numbers leave out. They sign off. AI cannot carry that responsibility, and donors like Spencer and Wellcome say so in their own rules.

Which reporting tasks suit AI

AI suits tasks where you hold the facts and a person can check the result quickly. Use this table to decide.

A worked example you can copy

This example is invented for training. Project Maji, its numbers and its quote do not exist. Replace them with your own cleaned material.

The evidence file, after cleaning:

Project Maji, Quarter 3 (July to September). Water committee trainings: 6 planned, 5 held. Session in Ward C postponed because of heavy rain. Attendance: 112 people (68 women, 44 men). Source: signed attendance sheets. Borehole repairs: 4 planned, 3 finished. Fourth waiting for a pump part. Households with water within 1 km: baseline 410, year-end target 700, September figure 590. Source: September household survey. Quote, consent form on file: "Now we keep the book ourselves." (Treasurer of a water committee.) Donor's comment on last report: asked for more on how women take part in committees.

Prompt 1: Build a style sheet from a past report

Below is a past report from our organisation. Study it and do not change it. Write a style sheet of no more than 150 words that describes how we write: sentence length, how we describe communities, how we report problems, and words we use often. List any phrases that are clichés so we can avoid them. Do not comment on the content. Report: [paste a past report with all personal details removed]

Save the result. It becomes part of every later prompt, so your reports keep one voice even when the writer changes.

Prompt 2: Draft from evidence only

This prompt uses the four Cs we teach: Clarity, Context, Constraints and Customisation.

Clarity: Write the "Progress against activities" section of a quarterly donor report. Context: We are a Kenyan NGO. The project helps communities manage water points. The donor asked last time for more on how women take part in committees. Constraints: Use only the evidence below. Do not add any number, name, cause, outcome or quote that is not in the evidence. Where the evidence does not answer something the section needs, write [CHECK: what is missing]. No more than 200 words. Do not claim that one activity caused a result unless the evidence says so. Report the postponed session and the unfinished repair plainly. Customisation: The reader is a programme officer who reads many reports and wants facts first. Follow this style sheet: [paste style sheet]. Evidence: [paste cleaned evidence]

The kind of draft you want:

Between July and September, the project held five of six planned water committee training sessions, reaching 112 people (68 women and 44 men). The sixth session, in Ward C, was postponed after heavy rain and is now set for [CHECK: new date]. Three of the four planned borehole repairs were finished; the fourth is waiting for a pump part. The September household survey found 590 households with water within one kilometre, up from 410 at baseline, against a year-end target of 700. [CHECK: how women are taking part in committee decisions, as the donor asked.]

Look at what the draft does not say. It does not say the training caused the rise in households, because the evidence does not show that. It flags the women's participation gap instead of inventing a line about empowerment.

Prompt 3: Check claims against evidence

Below are my evidence and my draft. Make a table with three columns: Claim in the draft, Exact line of evidence that supports it, Status (Supported, Partly supported, Not supported). Include every number, date, name, cause and outcome. Do not rewrite the draft. Evidence: [paste]. Draft: [paste]

Prompt 4: Find what the donor will ask

You are a careful programme officer reading this report section. List the five questions you would ask the NGO after reading it. Focus on missing explanations, unclear numbers and anything that sounds stronger than the evidence. Do not praise it. Section: [paste]

Prompt 5: Cut to the word limit

Shorten this section to [number] words. Keep every number, every date, every limitation and every problem reported. Remove repetition and general statements first. Show me a list of anything you removed. Section: [paste]

That final request matters, because cuts often remove the honest parts first.

Keep the numbers in Excel

Use a spreadsheet for arithmetic, and ask AI to explain the formula. With columns for Indicator, Baseline, Target and Actual, try this:

I have an Excel sheet. Column A is Indicator, B is Baseline, C is Target, D is Actual. Give me a formula for E2 that shows progress from baseline to target as a percentage, and tell me how to copy it down. Explain it in one sentence. I am comfortable with basic Excel only.

The answer should be close to =(D2-B2)/(C2-B2), formatted as a percentage. For Project Maji that gives (590 minus 410) divided by (700 minus 410), which is 62%. You did the sum in a tool you can audit, and no project data had to go into a chat window.

How to keep your organisation's voice

Your voice lives in the details only your team knows. AI can copy your sentence rhythm, but it cannot supply what you saw in the field.

Five habits help:

• Keep the style sheet from prompt 1, and update it each year.
• Write a banned-words list. Many teams remove "transformative", "empower", "leverage", "robust" and "holistic" unless they can say exactly what they mean.
• Name the actor. "The county water officer joined two sessions" beats "stakeholders were engaged".
• Use your community's terms for groups, places and roles.
• Have a person write the opening and the challenges section. These are the paragraphs donors remember, and they are where honesty shows.

What goes wrong, and how to avoid it

Tools differ in how they treat your data. Google's Gemini help page, for example, says human reviewers may read a sample of chats and asks people to keep confidential information out. If your office has a qualifying Microsoft 365 plan, sign in with your work account for Copilot Chat, because Microsoft's enterprise data protection applies to organisational use. Your team also needs a short, written rule. Our guide to an AI usage policy for Kenyan organisations covers what to put in it.

Should you tell the donor you used AI?

Yes, if your donor's rules require it, and we suggest yes whenever AI drafted real content. Spencer's and MacArthur's rules show where the line usually sits: light editing is one thing and substantial drafting or media creation is another.

A sample statement:

Generative AI ([tool name]) was used to help organise and draft sections of this report from our verified project records. No personal data about programme participants was entered into the tool. Project staff reviewed and verified the full content and take responsibility for it.

Only use that sentence if every part of it is true.

A sample email to your programme officer:

Dear [Name], we are preparing the [period] report for [grant name]. Our team sometimes uses generative AI tools to help organise and draft text from our verified project records. We do not enter personal data about participants, and our staff check all content before submission. Could you confirm whether [donor] has any guidance on this, or any requirement for disclosure? We will follow it. Kind regards, [Name]

A routine for a small team

A small team can run this with one owner per grant and a few fixed moments.

To find out if this saves time, record how long your last report took. Then time the next one in full, including cleaning and checking. Some early reports may take as long as before while the team learns. Judge the method after two or three reporting cycles, and count corrections from your reviewers as well as hours.

Your final checklist before sending

• We checked the grant agreement and reporting guidance for any AI rule.
• No names, ID numbers, phone numbers or case details went into a tool that our organisation has not approved for them.
• Every number matches the indicator table.
• Every total and percentage came from a spreadsheet.
• Every quote is real, and consent is on file.
• Every reference has been opened and read.
• No sentence claims a cause that our evidence does not show.
• Problems and delays are stated plainly.
• Narrative and budget agree.
• Each comment from the donor's last review has been answered.
• No [CHECK] markers remain.
• The programme lead has read every line.
• Any required AI disclosure is included and true.

Questions people ask

Do we have to tell our donor we used AI on a report?

It depends on the donor. Spencer Foundation requires disclosure when AI produced content, and MacArthur's principles call for it when AI is a primary contributor. Where the agreement is silent, ask your programme officer. Light grammar help is usually treated differently from drafting.

Which AI tool is right for donor reports?

No single tool is right for every team. Free versions of the main tools can draft and shorten text. For internal work, the deciding questions are which tool your organisation has approved and what happens to the text you enter.

Can we paste field notes into ChatGPT or Gemini?

Only after removing personal details and checking that nobody can be recognised from what remains. Field notes often contain names, locations and sensitive stories. If they cannot be cleaned safely, leave them out.

How do we stop AI inventing results?

Tell it to use only your evidence, mark gaps with [CHECK], and make no claims about cause. Then trace every number and name yourself. No prompt fully removes the risk.

Can AI write the logframe progress table?

It can lay out a table from verified figures that you supply. Have a person confirm every figure against the source. Do your calculations in a spreadsheet.

Does this method work with weak internet in the field?

Yes, because the drafting happens later. Field staff only need to send short, structured updates, and the report owner can do the AI work from an office connection. Agree a simple format for those updates, and remove names before sharing them.

Sources

• Spencer Foundation, Policy on the Use of Generative AI, last checked 9 November 2026
• MacArthur Foundation, Artificial Intelligence policy, last checked 9 November 2026
• Wellcome, Use of generative AI funding policy (updated 8 April 2026), last checked 9 November 2026
• UK Research and Innovation, Use of generative AI in application preparation and assessment, last checked 9 November 2026
• The Register, Deloitte refunds Aussie government after AI fabrications slip into report (6 October 2025), last checked 9 November 2026
• Google, Gemini Apps Privacy Hub, last checked 9 November 2026
• Microsoft Learn, Enterprise data protection in Microsoft Copilot and Copilot Chat, last checked 9 November 2026

Where FuKazee fits in

FuKazee trains teams to use AI on the work in front of them, and NGO reporting is part of that. We assume nobody in the room knows AI yet, and we avoid jargon. We say plainly that nobody becomes an expert in a few days. We show the free version of each feature, teach the four Cs, and never push one tool.

Practice uses invented or cleaned material, so no beneficiary data enters the training. For Maisha Girls Safe House we trained 13 staff on using AI for communication and reporting; you can read their recommendation. If you are planning a programme, read how to train NGO staff on AI in Kenya and see our page on AI training for NGOs in Kenya.

Training will not repair a weak evidence system. If your team does not collect good data during the period, start there.

To talk through training for your reporting team, email lead@fukazee.com