How Nonprofits Can Choose the Right AI Tools?

AI Tools for Nonprofits: How to Choose What Your Team Actually Needs

To choose an AI tool for your nonprofit, test it against five gates in order: the specific task it must perform, its true cost, how it handles your data, who stays in control of its output, and how it performs in a 30-day trial on real work. If a tool fails one gate, stop there. Features, demos and brand names matter only after those five answers are clear.

Why do so many nonprofit AI purchases disappoint?

Because most organisations adopt AI before they decide what it is for. A benchmark study of 346 nonprofits, run by Virtuous and Fundraising.AI in December 2025, found that 92% of nonprofits use AI, but only 7% report major improvements in organisational capability. The same study found that 81% use AI individually without shared workflows, and 47% have no AI governance policy.

Read together, those numbers describe a sector that is buying and trying tools but not deciding how they fit into the work. The study covers mostly US fundraising teams, so treat the exact percentages as indicative for other markets. The pattern is what matters: tools arrive faster than decisions about them.

What does a small nonprofit actually face when buying AI?

A small organisation rarely fails because it picked a weak model. It fails because nobody owns the process around the tool.

On paper, you pick the best-rated tool and roll it out. In an organisation with eight staff and four donor projects, the harder questions are different. Who checks the output before it goes into a donor report? Who pays when usage grows? What happens to the workflow when the enthusiastic colleague who set it up leaves? A tool choice that ignores those questions is a purchase, not a decision.

I cover the risk side of this problem in [The Hidden Risk in Buying AI Tools]. This guide is the practical sequence that comes before and around it.

What is the Five Gates Test for choosing AI tools?

The Five Gates Test is a decision sequence: Task → Cost → Data → Human control → Trial. A tool must pass each gate before it reaches the next. The order is deliberate, because the cheapest gates to check come first and the most expensive one, a live trial, comes last.

Gate	Question	Pass signal	Fail signal
1. Task	What exact job must this tool do?	You can describe it in one sentence with a named owner	"We want to use AI more"
2. Cost	What is the real monthly cost, including staff time?	Total cost is lower than the value of hours saved	Only the licence price was checked
3. Data	What data will it touch, and where does that data go?	Terms for your plan are written, and no sensitive data is needed	Vendor terms unread, or free tier used with personal data
4. Human control	Who reviews, approves and takes over?	Named reviewer, written stop rule, one-page handover note	Output goes out unchecked
5. Trial	Does it work on our real work?	Pass criteria were set before the trial and were met	Decision made after a demo

Gate 1: What task should the AI tool do?

Define one task, not a category. "Help with fundraising" is a category. "Produce a first draft of the quarterly narrative section of donor reports from our project notes" is a task.

A usable task statement has four parts: the input (what goes in), the output (what comes out), the owner (who is responsible for the result) and the current effort (how long it takes today). If you cannot state the current effort, you cannot later prove the tool helped.

Start with a task that is repetitive, low-risk and checked by a person who knows the subject. Reporting drafts, meeting summaries, first-pass translations and grant call summaries usually fit. Eligibility decisions, safeguarding cases and anything involving individual beneficiaries do not.

Gate 2: What will an AI tool really cost?

The licence is the smallest honest number. The real cost includes everything the tool asks of your team.

Real monthly cost = licences + usage fees + (review hours × hourly staff cost) + admin time + setup cost spread over 12 months

Review time is the line most teams forget. If a tool saves two hours of drafting but adds ninety minutes of careful checking, the saving is thin. If it needs a paid plan for every user, a data agreement and training time, the cost rises again.

Cost item	What to ask
Licence and seats	Is pricing per user, per organisation, or per usage? What happens at the next tier?
Nonprofit pricing	Is a discount available, and for how long is it guaranteed?
Review and correction time	How many minutes of checking per output?
Training and setup	Who learns it, and how many hours does that take?
Exit cost	Can you export your work and prompts if you leave?

Ask for the cancellation terms before the first payment. An annual contract for an untested tool is the most common avoidable mistake.

Gate 3: What happens to your data?

Sort your data before you sort tools. Three tiers are enough for a small team:

• Green: published reports, website text, public call documents.
• Amber: unpublished strategy, draft proposals, internal budgets.
• Red: beneficiary records, donor personal data, safeguarding information, staff and HR files.

Nonprofit organisations should not enter red data into any AI tool unless a data processing agreement is in place and the director has approved that use in writing. Where the personal data of people in the EU or Switzerland is involved, the GDPR and the Swiss Federal Act on Data Protection (nFADP) apply, and the organisation remains responsible even when a vendor processes the data.

Check these five points in the terms of the exact plan you are buying, since free and paid plans often differ:

1. Is your content used to train the vendor's models?
2. How long are inputs and outputs retained?
3. Where is data processed, and by which sub-processors?
4. Can you delete your data, and is that confirmed in writing?
5. Is a data processing agreement available?

If a task can be done with green data only, Gate 3 becomes simple. That is a good reason to begin there.

Gate 4: Who stays in control of the output?

A person must stay accountable for anything an AI tool produces and an organisation sends out. That is not a technical preference. It is an institutional one: donors, regulators and beneficiaries hold the organisation responsible, not the software.

Put three things in writing before the trial:

• Owner: the person responsible for the workflow.
• Reviewer: the person who checks output before it leaves the organisation, ideally someone who knows the subject matter.
• Stop rule: the conditions under which the tool is paused, such as a factual error in a donor document or any use with red data.

Then apply what I call the handover test: if the person who set this up left tomorrow, could a colleague run the workflow from a one-page note? If not, the organisation owns a dependency, not a process.

Gate 5: How do you run an AI tool trial?

Run a 30-day trial on one workflow with two or three users, and decide the pass criteria before it starts.

1. Measure the baseline. Record how long the task takes today and how many corrections it usually needs.
2. Set pass criteria. For example: time saved per task, accuracy checked by the reviewer, and whether the team would choose to keep using it.
3. Use real work with the lowest-risk data. Anonymise where possible.
4. Track review time. Log the minutes spent checking and fixing, not only the minutes saved.
5. Set an exit date. On day 30, decide: adopt, adjust or drop. Do not extend by default.

Trial metric	How to record it	Decision rule
Time per task	Before and during trial	Net saving after review must be positive
Correction rate	Reviewer's log	Agree an acceptable level beforehand
Data incidents	Any red data used by mistake	Any incident pauses the trial
User choice	Short end-of-trial question	Keep only if users would choose it again

What does this look like in a real organisation?

Consider a hypothetical organisation with eight staff running four donor-funded projects. The communications officer wants an AI tool to write donor reports.

Problem: reports take too long, and the team tests tools informally with whatever is free.

Wrong approach: the team buys an annual plan after a convincing demo, and everyone starts pasting project data into it. Nobody has checked the data terms, and the first reviewer is whoever has time.

Better approach: the team defines one task (first drafts of the narrative section from project notes), uses only green and anonymised amber data, names the programme manager as reviewer, and runs a monthly plan for 30 days. They record the baseline drafting time and the review time.

Practical consequence: the decision on day 30 rests on logged evidence, and the organisation has a one-page workflow note that survives staff changes. No measurable results are claimed here, because this is an illustration of the process, not a case study.

My view

My rule is simple: never buy an AI tool for an organisation before you can describe the task, the owner and the reviewer in three sentences. In my experience the instinct to start with the tool is the strongest bias in this decision, and it is the one the market rewards, because vendors sell capabilities, not workflows.

I would also be cautious about "AI-first" mandates for small teams. A smaller number of documented, repeatable workflows beats many individual experiments, and the benchmark data above points the same way.

What to do next

1. Write one task statement using the four parts: input, output, owner, current effort.
2. Classify the data that task touches as green, amber or red, and restrict the trial to green or anonymised data.
3. Calculate the real monthly cost for one candidate tool, including review hours.
4. Name an owner and a reviewer, and write a one-line stop rule.
5. Book a 30-day trial with an exit date in the calendar, and record the baseline this week.

Conclusion

Organisations rarely regret a tool they tested properly, and often regret one they bought on enthusiasm. The common misjudgement is treating tool selection as a feature comparison, when it is a decision about work, money, data and accountability. Choose the task first, and the tool becomes the easy part.

FAQ

How do nonprofits choose the right AI tool?

Start with one specific task, then test candidate tools against cost, data handling, human control and a 30-day trial. Choose the tool that passes all four with your real work, not the one with the longest feature list. Write the pass criteria before the trial so the decision is based on evidence.

Is it safe for a nonprofit to use AI tools with donor or beneficiary data?

Only with controls in place. Personal data of donors, beneficiaries or staff should not go into an AI tool without a data processing agreement, clear retention and training terms for your specific plan, and written approval from leadership. Where possible, use anonymised or public data instead.

How much do AI tools really cost a small nonprofit?

The real cost is the licence plus usage fees, staff review time, training, administration and exit costs. Review time is the most underestimated part. A tool is worth buying only if the value of hours saved, after checking, exceeds that total.

How long should a trial of an AI tool last?

Thirty days on one workflow with two or three users is usually enough for a small team. Measure the baseline first, set pass criteria in advance and fix an exit date. Long open-ended trials tend to turn into unplanned adoption.

Does a nonprofit need an AI policy before buying tools?

A short policy helps, and it can be one page: allowed data tiers, who approves new tools, who reviews outputs. Nearly half of the nonprofits in the 2025 benchmark study reported having none, so having one already puts you ahead of many peers.

Sources

• Virtuous and Fundraising.AI, The 2026 Nonprofit AI Adoption Report (survey of 346 nonprofits, December 2025; released 16 February 2026).

Last updated

4 October 2026 - Article first published.