92% use AI. 7% see real impact.

Two years ago the question was whether nonprofits would use AI. That question is settled. A benchmark study of 346 nonprofits published in February 2026 found 92% using AI in some form, up from single digits two years earlier. A separate study of 917 organizations by NTEN and The Bridgespan Group put it at 98%.

The interesting number is the one next to it. In the same benchmark study, only 7% reported major improvements in organizational capability.

The gap isn't the tools

The same study found 81% of organizations using AI individually and ad hoc, and only 4% with documented, repeatable workflows. Nearly half, 47%, had no AI policy at all.

A separate 2026 survey of 75 nonprofits by Coastal Cloud found the same pattern from a different angle: 67% named a lack of strategic direction as a constraint, while only 28% named budget. Just 12% said their most recent AI initiative started with a clearly defined problem. More than a third adopted a vendor's recommended use case, and another third picked a platform first.

Most organizations bought the tool before naming the problem. That's not a failure of discipline. It's what happens under pressure to keep up.

Adoption runs ahead of readiness

The NTEN and Bridgespan research found that only 32% of nonprofit staff have received formal AI training, while nearly half of executives expect AI spending to grow in 2026. Spending is scaling faster than governance and skills.

Trust is the other missing piece. In the Coastal Cloud survey, 79% of leaders said employees were eager to work with AI, but 53% said staff didn't trust the outputs enough to act on them. An output nobody trusts still has to be checked by hand, which means the work didn't actually move.

Where AI is used versus where the pain is

One finding stands out for anyone building software for this sector. The most common uses are low-stakes writing help: spelling and grammar, subject lines, first drafts. The workflows where mistakes are expensive, such as grant applications, applicant communication, and program reporting, are the ones least likely to be supported by a shared system.

That's the difference between a tool and a workflow. A draft written in a chat window doesn't know what your organization already promised a funder, who is responsible for the next step, or whether anyone replied to the family that applied last week.

What this implies for software built for nonprofits

• Sell the workflow, not the assistant. The evidence says individual tool access is already solved and hasn't produced results.
• Design for trust. If staff can't see what an AI output was based on, they won't act on it, and the time saving disappears.
• Assume no AI policy exists. With roughly half of organizations lacking one, software should make safe defaults the easy path rather than assuming governance is in place.
• Start from a named problem. The organizations reporting real gains picked a high-friction workflow first.