The most challenging aspect of integrating AI into non-profit work isn’t finding AI tools. Most teams already possess them. And many staff members are already using them.

The challenge is in making shared decisions that turn individual experiments into processes the organization can duplicate, teach, and improve. A workflow document, a one-page checklist, or a clear data boundary — these artifacts aren’t the main goal. The true focus is on the agreements needed to create them. A consensus on what AI is allowed to influence and what it isn’t.

Two recent studies shed light on why many organizations struggle with these issues.

The Virtuous + Fundraising.AI benchmark (346 nonprofits surveyed throughout December 2025) reveals what many of us observe: AI use is common, but its impact on organisations remains limited. They report a 92% adoption rate, yet only 7% of organizations say AI has significantly improved their ability to fulfil their mission. Additionally, 81% of AI use is still individual and ad hoc (not integrated into shared workflows), and 47% of organizations lack an AI governance policy.

That’s not primarily a training or budget issue. It’s what happens when an organization never reaches the decision stage that documentation requires. What qualifies as acceptable use in this workflow, what needs to be verified, what shouldn’t be touched, and who approves.

CEP’s AI With Purpose clarifies why those conversations are often delayed. In their survey, 83% of nonprofit leaders say none of their foundation funders have approached them about AI. Another 15% say only “a few” have. CEP also highlights that 98% of foundations currently lack a policy regarding whether grantees can use AI in their work.

When expectations are unclear or absent, organizations tend to keep AI use informal. People experiment quietly, without documenting or standardising their efforts. They avoid being the first to set “rules” in a vacuum.

So, what does the decision-forcing process look like in practice? Select one recurring workflow — appeal writing is a good example — and define four aspects.

1) What AI can help with (in this workflow).
Draft options. Refine language. Generate variants. Convert an outline into a first draft. Summarize existing notes.

2) What AI can help draft, but it can’t be the source, the verifier, or the decision-maker. In practice, that means humans must still own:

  • Program claims (what your work achieves, what you promise, what you’ve delivered)
  • Facts (dates, numbers, outcomes, names, policies, grant requirements — anything that could be wrong in the real world)
  • Donor-specific context (relationship history, preferences, constraints — anything that lives in your CRM or in someone’s head)
  • Judgment calls (what you choose to say, what you choose not to say, and what you’re willing to stand behind)

3) The data boundary.
What staff can paste into a tool for this workflow, and what’s off-limits. Put it in writing so people aren’t guessing under pressure.

4) One or two measures you’ll track for a month.
Time to a usable draft. Revision cycles. Errors caught in review. Turnaround time in a peak week. Something concrete.

If you do this, the checklist becomes a straightforward summary of decisions you’ve already made — not a substitute for making them.

For more context (beyond adoption numbers)

The most challenging aspect of integrating AI into non-profit work isn’t finding AI tools. Most teams already possess them. And many staff members are already using them.

The challenge is in making shared decisions that turn individual experiments into processes the organization can duplicate, teach, and improve. A workflow document, a one-page checklist, or a clear data boundary — these artifacts aren’t the main goal. The true focus is on the agreements needed to create them. A consensus on what AI is allowed to influence and what it isn’t.

Two recent studies shed light on why many organizations struggle with these issues.

The Virtuous + Fundraising.AI benchmark (346 nonprofits surveyed throughout December 2025) reveals what many of us observe: AI use is common, but its impact on organisations remains limited. They report a 92% adoption rate, yet only 7% of organizations say AI has significantly improved their ability to fulfil their mission. Additionally, 81% of AI use is still individual and ad hoc (not integrated into shared workflows), and 47% of organizations lack an AI governance policy.

That’s not primarily a training or budget issue. It’s what happens when an organization never reaches the decision stage that documentation requires. What qualifies as acceptable use in this workflow, what needs to be verified, what shouldn’t be touched, and who approves.

CEP’s AI With Purpose clarifies why those conversations are often delayed. In their survey, 83% of nonprofit leaders say none of their foundation funders have approached them about AI. Another 15% say only “a few” have. CEP also highlights that 98% of foundations currently lack a policy regarding whether grantees can use AI in their work.

When expectations are unclear or absent, organizations tend to keep AI use informal. People experiment quietly, without documenting or standardising their efforts. They avoid being the first to set “rules” in a vacuum.

So, what does the decision-forcing process look like in practice? Select one recurring workflow — appeal writing is a good example — and define four aspects.

1) What AI can help with (in this workflow).
Draft options. Refine language. Generate variants. Convert an outline into a first draft. Summarize existing notes.

2) What AI can help draft, but it can’t be the source, the verifier, or the decision-maker. In practice, that means humans must still own:

  • Program claims (what your work achieves, what you promise, what you’ve delivered)
  • Facts (dates, numbers, outcomes, names, policies, grant requirements — anything that could be wrong in the real world)
  • Donor-specific context (relationship history, preferences, constraints — anything that lives in your CRM or in someone’s head)
  • Judgment calls (what you choose to say, what you choose not to say, and what you’re willing to stand behind)

3) The data boundary.
What staff can paste into a tool for this workflow, and what’s off-limits. Put it in writing so people aren’t guessing under pressure.

4) One or two measures you’ll track for a month.
Time to a usable draft. Revision cycles. Errors caught in review. Turnaround time in a peak week. Something concrete.

If you do this, the checklist becomes a straightforward summary of decisions you’ve already made — not a substitute for making them.

For more context (beyond adoption numbers)