What Is AI For in a Mission-Driven Organization?

Here is a scene I have heard described in some version many times, so I will tell it as a hypothetical. The development manager at a 40-person family services agency comes back from a conference convinced that AI can cut the time it takes to write grant reports in half. Licenses are cheap. A board member has been asking about AI for months, and the executive director knows that saying yes would feel like progress.

She traces the reporting process first, and the writing turns out to be the easy part. Program data arrives late from the field because case managers enter it on Friday afternoons, after their visits; two funders define the same outcome in different ways; and the development manager spends most of her report time chasing numbers and reconciling them, with only a few hours left for drafting. Faster drafting would save those few hours and leave the real bottleneck where it is.

I run a nonprofit, so I write from the operator's chair, and in my experience the tool usually arrives first in these conversations while the problem gets defined later, if at all. This article argues for reversing that order. AI is worth considering when you can connect a specific use to a specific improvement for your staff or for the people you serve. It also requires a willingness to check, afterward, whether that improvement happened.

Your staff have already answered the "should we" question

Many executive directors still talk about AI as a decision they have not yet made, but the 2026 State of Nonprofit AI Adoption and Governance report from NTEN and Bridgespan suggests otherwise. Of its 917 respondents, 98% reported some use of AI in their organizations. Only 61% described that use as official (organized pilots, team-wide use, or full integration), while about 37% said it was informal or individual only. More than half reported "shadow" use, meaning staff using tools the organization has not approved. Written guidance on responsible use was fully in place at just 38% of organizations, and 57% of executives said they had no AI budget at all (NTEN and Bridgespan, 2026).

Funders have not filled the gap. Ninety percent of the foundations surveyed by the Center for Effective Philanthropy provide no AI implementation support to any grantee, even though nonprofits in the same study named staff education (68%) and funding for tools (63%) as the help they wanted most (CEP, AI With Purpose, 2025).

Both surveys drew on organizations that chose to respond, and NTEN's respondents are probably more comfortable with technology than the sector as a whole, so these numbers describe a pattern rather than an exact rate. The pattern is still hard to miss. Your case managers and development staff are likely experimenting already, often with personal accounts and no guidance, which means the leadership question has moved. What is AI for in your organization, and under what rules?

Saved minutes are not the same as changed work

Research on AI at work shows real gains on specific tasks. It also shows how rarely those gains become something an organization can see.

The strongest evidence of gains comes from customer service. When economists Erik Brynjolfsson, Danielle Li, and Lindsey Raymond followed 5,172 support agents at one company through the introduction of an AI assistant, they found that agents resolved 15% more issues per hour on average. Newer and less-skilled agents gained the most. The most experienced agents barely sped up, and the quality of their work dipped slightly (Quarterly Journal of Economics, 2025). That is a meaningful result for one well-defined workflow, and it says nothing about your case notes or donor letters.

Broader studies are less encouraging. In an experiment across 66 firms, 7,137 workers received an AI assistant built into their email and documents; the roughly 80% who used it spent about two fewer hours a week on email and worked less after hours. What did not change was the amount or mix of tasks people did (Dillon et al., NBER Working Paper 33795, revised November 2025). Some of the authors worked for Microsoft, whose product was being tested.

Denmark offers a view at the scale of a whole country. Anders Humlum and Emilie Vestergaard linked surveys of about 25,000 workers in AI-exposed occupations to government payroll and hours records. Users reported saving about 3% of their work hours, yet earnings and recorded hours did not measurably move (Humlum and Vestergaard, 2025; later NBER version).

British government trials found the same gap, with weaker designs. Staff in a cross-government trial of about 20,000 Copilot licenses estimated they saved 26 minutes a day, though the trial had no comparison group and no record of where the time went (Government Digital Service, 2025). HMRC, the tax agency, assigned licenses at random, and its staff reported a much smaller saving: 2–3% of the working week, or about an hour (HM Revenue & Customs, 2026). At the Department for Business and Trade, evaluators saw small time savings and no evidence of higher productivity (Department for Business and Trade, 2025). Business leaders report something similar: close to nine in ten of almost 6,000 executives surveyed in the US, the UK, Germany, and Australia said AI had not affected their firm's productivity over the previous three years (Yotzov et al., Firm Data on AI, 2026).

None of this makes the time savings imaginary. Left undirected, though, saved minutes dissolve into the workday; fifteen minutes here and twenty there will not add up to an extra home visit unless someone decides they should and rearranges the schedule to make room.

Four kinds of improvement, and when the smallest one is enough

"AI will help us" can mean very different things, and the difference matters because each kind of help calls for its own decisions and its own evidence. I find it useful to separate four levels:

1. Task performance. A single piece of work gets faster or better, such as a draft that takes ten minutes instead of forty. Nearly all the research sits here, and it is the easiest level to test.

2. Working conditions. Staff carry less tedious work and stay late less often. After-hours work fell in the multi-firm experiment, and a review prepared for England's National Workload Action Group expects social workers' AI time savings to go toward more manageable weeks rather than extra cases (Research in Practice, 2025).

3. Organizational capacity. The organization serves more people or responds faster, which happens only when management decides where recovered time goes.

4. Mission outcomes. Children are safer and families more stable. This is why we exist, and almost no AI research reaches this level; none of the studies above tried.

Most AI pitches promise the fourth level while offering evidence for the first. Keep them separate, and ask what path, specifically, would carry a task improvement up to the level you care about.

A fair objection is that this sets the bar too high. Small organizations cannot run randomized trials, and demanding proof of mission impact before trying anything would freeze them in place. I agree. In a field with our burnout and turnover, reducing the load on overstretched staff is a worthy aim by itself, and a modest, reversible change to an internal workflow does not need a research design. It needs honesty about the goal. If the goal is "our program assistants spend less time reformatting reports," say that, check whether it happened, and stop there. Trouble starts when a burden-reduction project is described to a board or a funder as a mission-impact project, because then nobody checks for the thing that was promised.

Design around the purpose, then count the costs

Houston Endowment, a foundation, offers one of the more thoughtful examples I have found of putting purpose first. It ran a pilot that replaced written grant reports with recorded conversations between grantees and program staff, usually 45 to 90 minutes long, and used AI for transcription and summarization so staff could listen instead of taking notes (Houston Endowment, 2025). What makes it useful is the order of decisions: the foundation began with a goal (lighten grantees' reporting burden and hear from them directly), and only then asked where AI could help. It has not published an evaluation, named its tools, or said how many grantees took part, so treat it as an illustration of design thinking rather than proof that the approach works.

Once the purpose is clear, look hard at what speed can hide. A fast draft with errors creates review work for someone else. Staff who write fluently in English may find a tool helpful while colleagues who do not find it frustrating, and a polished summary of a family's situation that misses context can damage trust with the family and with the caseworker who has to fix it. Licenses are cheap compared with the hours that go into setup, training, checking, and correction. Decide before you start which of these costs you are willing to accept.

A worksheet for deciding what AI is for

I developed this worksheet for the series. It has not been tested as a formal assessment, so use it as a structured conversation with your team rather than a score.

Two rows deserve extra attention. Asking who should help define "better" is my own recommendation, not a research finding; in community work, the people affected by a change tend to see costs that managers miss, and involving them early is cheaper than repairing trust later. The alternatives row matters because the honest answer is sometimes a process fix. For the agency in the opening scene, a mid-week data entry deadline and one shared outcome definition would cost nothing and address the real bottleneck.

To start, pick one recurring problem your staff complain about and can describe in a sentence, then fill out the worksheet with the people who do that work. If AI still looks like the right contribution afterward, you have a candidate for a first project, which is the subject of the next article in this series. Otherwise, you leave the meeting with a clearer description of the problem and, often, a cheaper fix to try first.

References

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.

Department for Business and Trade. (2025, August 28). Microsoft 365 Copilot pilot: DBT evaluation report. GOV.UK.

Dillon, E. W., Jaffe, S., Immorlica, N., & Stanton, C. T. (2025). Shifting work patterns with generative AI (NBER Working Paper No. 33795, revised November 2025). National Bureau of Economic Research.

Figueroa, D., & Dwelle, J. (2026). 2026 state of nonprofit AI adoption and governance report. NTEN and The Bridgespan Group.

Government Digital Service. (2025, June 2). Microsoft 365 Copilot experiment: Cross-government findings report. GOV.UK.

HM Revenue & Customs. (2026, July 9). Evaluating the impact of Microsoft Copilot in HMRC. GOV.UK.

Houston Endowment. (2025, June 30). Pilot project centers grantee voice through oral reporting and AI.

Humlum, A., & Vestergaard, E. (2025). Large language models, small labor market effects (BFI Working Paper No. 2025-56). Becker Friedman Institute, University of Chicago. Also issued as NBER Working Paper No. 33777.

Rothera, S. (2025). Artificial intelligence (AI) in case recording: Reducing unnecessary social worker workload. Research in Practice, for the Department for Education.

Smith Arrillaga, E., Grundhoefer, S., & Im, C. (2025). AI with purpose: How foundations and nonprofits are thinking about and using artificial intelligence. The Center for Effective Philanthropy.

Yotzov, I., Barrero, J. M., Bloom, N., Bunn, P., Davis, S. J., Foster, K. M., Jalca, A., Meyer, B. H., Mizen, P., Navarrete, M. A., Smietanka, P., Thwaites, G., & Wang, B. Z. (2026). Firm data on AI (NBER Working Paper No. 34836). National Bureau of Economic Research.