Most nonprofit AI pilots that fall apart weren't badly chosen. They were badly handed over. The demo worked, the first few weeks worked, and then the one person who understood how the tool actually behaved moved on, and nobody else could run it. The failure gets blamed on the tool. The real cause is that the knowledge of how to run it never left one person's head.
Why do nonprofit AI pilots end up depending on one person?
Small teams don't have the slack to carry a second person who already knows how a tool works, just in case. A communications officer, a programme coordinator, or a single tech-comfortable volunteer often becomes the de facto owner of a new tool by default, simply because they were the one curious enough to set it up. Donor accountability makes this worse, not better: a funder who hears that a reporting process now "uses AI" expects that process to keep running, and a quiet reversion to manual work after someone leaves is the kind of thing nobody wants to explain in a grant report. A routine staff change is enough to expose this dependency. Plan the handover while the person who built the workflow is still available.
What does a failed AI handover look like in a mid-size NGO?
Imagine a communications officer at a mid-size NGO who starts using an AI tool to draft the monthly donor newsletter. She feeds it the raw programme updates, a style guide she wrote herself, and a running list of phrases to avoid (jargon the board dislikes, claims legal asked her to stop making). Over four months, she builds an instinct for when the draft is usable and when it needs a full rewrite, but none of that instinct is written down anywhere. It lives in her head as a set of habits: which prompts work, which files to attach, what a bad draft looks like before it's too obviously bad.
When she leaves for a new job, her replacement inherits the tool, a login, and nothing else. The first newsletter draft under the new person reads strangely formal, repeats a claim legal had flagged months earlier, and takes three rounds of frustrated edits to fix, more time than writing it manually would have taken. Within two newsletters, the team quietly stops using the tool, and nobody updates the "we use AI for donor comms" line in last year's grant report, because admitting the reversal feels worse than leaving it inaccurate.
Nothing here was a technology failure. The tool worked exactly as it had for four months. What failed was that one person's tacit knowledge was never converted into something a second person could pick up and run.
What should an AI handover record contain?
A handover doesn't need a training manual or a formal SOP. It needs four short, specific records, written down somewhere a successor will actually find them, not buried in a Slack thread or a single person's inbox.
| Record | What it answers | What it looks like in practice |
|---|---|---|
| Choice | Why this tool, for this task, instead of doing it manually or with another tool | A few sentences: the task, the tool, and the one or two reasons it won over the alternative |
| Boundary | What the tool is allowed to touch, and what it's never given | Which files, data, or systems feed it, and an explicit line on what stays out (donor PII, unreleased board decisions, anything under embargo) |
| Failure | What a bad result actually looks like, from someone who has seen one | Two or three concrete examples of output that looked fine but wasn't, and what gave it away |
| Owner | Who is responsible for the tool right now, and who checks in on it | A named person, not a role or a team, plus a review date already on a calendar |
This maps directly onto Gate 4, Human control, in the Five Gates Test from How Nonprofits Can Choose the Right AI Tools, which asks whether a named owner, a reviewer, a stop rule, and a handover note all exist before a tool is adopted at all. The four records here are what actually fill in that gate, written at the level of detail a real successor needs, not just a checkbox that says "handover note: yes."
The AI Handover Test: a five-day procedure
Writing the four records down isn't proof they work. The test I use is simple: the person who built the pilot steps away for five working days, and a successor runs real tasks with it, using only what's written down.
- Day 1: The builder hands over the four records and nothing else, no verbal explanation, no "just ask me if you get stuck."
- Days 1-4: The successor runs the tool on real, current work, not a practice task, and logs every single point where they had to guess, because the records didn't say.
- Day 5: The builder reviews the log, and for each stall, updates the Boundary or Failure record so the next person wouldn't hit the same gap.
- After the test: The director (not the builder) signs off on a named Owner and a specific review date already placed on a calendar, not a vague "we'll check in sometime."
My rule for whether a pilot is actually finished, not just demonstrated, is this: if only the person who built it can run it for a week, it hasn't been handed over, it's only been shown off. A successful demo proves the tool works. A successful handover proves the organisation can keep using it after the builder is gone, which is the only version of "working" that survives staff turnover.
The same accountability applies to each external claim: use the Signature Test to record who verified a number before publication. A demo without a repeatable workflow risks becoming AI theatre.
My view
I'd rather see a two-person team spend half a day writing these four records than spend a week in a formal AI training session. Training teaches a person how to use a tool today. A written record is what lets a completely different person use it correctly in eight months, after the first person is gone and nobody remembers the details that felt too obvious to write down at the time.
What to do next
- Pick the one AI tool your organisation currently depends on the most, and check whether anyone besides its builder could actually run it tomorrow.
- Write the four records (Choice, Boundary, Failure, Owner) for that tool today, even in rough form; a rough written record beats a perfect one that doesn't exist yet.
- Schedule the five-day handover test for a month from now, with an actual successor named in advance.
- Put the next review date on a calendar before you finish reading this, not after.
Conclusion
An AI pilot that only one person can run isn't finished, no matter how well it performed in the demo. The gap between a working demo and a working process is almost always a handover gap, not a capability gap, and it's cheap to close while the person who understands the tool is still around to write down what they know. The expensive version of this lesson is learning it after that person has already left.
FAQ
How long should the four written records actually take to write?
For most small nonprofit AI pilots, half a day is realistic for a rough first version of the Choice, Boundary, Failure, and Owner records. They need to be specific enough for someone who has never used the tool to follow them.
What if the person who built the pilot refuses to document how it works?
Treat an undocumented nonprofit AI pilot as unready for production, regardless of how well it performs. A tool that depends entirely on one person’s knowledge carries the same handover risk whether that person is reluctant to document it or simply busy.
Does the five-day handover test need to happen before the pilot is adopted permanently?
Ideally, run the five-day AI handover test before permanent adoption. Having a successor run real tasks using only the written records catches gaps while the stakes are still low. Waiting until a staff departure means resolving those gaps during an operational crisis.
Is this different from a standard employee handover document?
A standard employee handover document usually describes a role. The Choice, Boundary, Failure, and Owner records describe one specific AI tool’s behavior, permitted inputs, warning signs in its output, and current responsibility. A generic role handover rarely captures that detail.
Related
- How Nonprofits Can Choose the Right AI Tools : Gate 4: Human control
- NIST AI 600-1: Generative Artificial Intelligence Profile: a reference for confabulation risk and human oversight.
About the author
Milan Bogojevic works on practical AI adoption in nonprofits, with a focus on workflows, human accountability and organisational capability.
Last updated: 10 October 2026.