AI and data protection in M&E: what do donors expect?
Oct 12 2026 · Governance
Donor data-protection rules apply to AI too. A short, practical policy for M&E teams.
Donors expect M&E teams to treat AI like any other data processor: with a written policy, a clear data-tiering rule, and an audit trail. In practice that means public data can go into any approved tool; confidential data needs training disabled and identifiers removed; personal or sensitive data should never be pasted into a general model — use a private or local tool instead. Before prompting, strip names, locations and identifiers, and keep the mapping off the prompt. Record which tool was used and who verified the output. Many donors now ask for this in the data-management plan, so a one-page AI policy is increasingly a condition of the grant.
What do donors actually expect from AI in M&E?
Donors are not asking for a ban on AI. They are asking for evidence that you have thought about data protection. In practice, that means a written policy, a clear rule for which data can go into which tool, and a record of what was used and who checked it. The expectation is usually folded into the data-management plan or the grant agreement, not issued as a separate AI rulebook.
If you can show a one-page policy and a simple log, you are ahead of most teams.
Which data can you put into an AI tool?
Tier your data before you open a chat window. The tier determines the tool and the controls.
| Data tier | Example | Tool allowed | Controls |
|---|---|---|---|
| Public | Published reports, open datasets | Any approved tool | Normal review of output |
| Confidential | Internal drafts, budgets, unpublished findings | Approved tool with training disabled | Remove identifiers; keep mapping offline |
| Personal or sensitive | Names, locations, health data, beneficiary lists | Private or local model only | Do not paste; use anonymised extracts |
If you are unsure which tier applies, treat it as the higher tier. That single habit prevents most incidents.
What should a one-page AI policy contain?
Keep it to one page. A policy people read beats a manual they ignore.
- Scope: which teams and which tools are covered.
- Data tiers: what may go into which tool.
- Prohibited uses: personal data in general models.
- Verification: who checks AI output before use.
- Logging: what is recorded and where.
- Review: when the policy is updated.
How do you handle personal data when using AI?
The safest rule is simple: never paste personal or sensitive data into a general AI model. If you need AI help with a document that contains identifiers, create an anonymised extract first. Replace names with codes, remove locations and dates of birth, and keep the mapping in a separate file that never enters the prompt.
For work that genuinely requires personal data, use a private or locally hosted model that you control. Check with your IT or data-protection lead before you start.
What does a data-management plan need to say about AI?
Add a short section that names the tools you use, the data tiers, and the verification and logging steps. If your donor has a template, map your policy to it. If not, a paragraph and a link to your one-pager is usually enough.
You can see how this fits with evaluation planning in our /store/evalplan resource.
How do you verify AI output in M&E?
AI output is a draft, not a finding. Verification means a named person checks facts, figures and references against the source. For qualitative coding, check a sample of AI-assigned codes against your own coding. For text, check that no invented citations or statistics have crept in.
- Name the verifier for each use.
- Check numbers against the original source.
- Check quotes and citations exist.
- Record the check in your log.
What should you log when using AI?
A simple log is enough. It should let someone else reconstruct what happened.
- Date and tool used.
- Purpose (for example, summarising a report).
- Data tier and whether identifiers were removed.
- Who verified the output.
- Any issues or corrections.
What are the risks of getting this wrong?
The main risks are a data breach, a donor finding out you had no policy, and AI-generated errors entering a report. A breach can end a grant. A missing policy can delay one. Errors can damage your reputation. None of these require a complex fix; they require a habit of tiering data and checking output.
How do you roll this out across a team?
Start with the one-pager. Run a short session where people practise tiering a few real documents. Put the log where everyone can find it. Review after a month and adjust. If you need a tool to help choose the right AI for each task, see /which-ai. For team-wide templates and guidance, /membership covers the basics.
Written for M&E professionals. Not a guarantee of accuracy — always verify every claim against its source, and let a named human sign off. Get the free handbook.