(image from pexels.com)
By Richard Geiger
Fundraisers have been told the same thing about their data for 20 years: Keep it clean, keep it current, keep it complete. It was good advice then, offered mostly in the context of avoiding a bounced appeal letter or a red-faced apology to a major donor whose name got misspelled for the third year running.
It is essential advice now for a much larger reason. Artificial Intelligence (AI) does not fix bad data. It amplifies it.
A recent nonprofit AI benchmark study from Virtuous found that 92% of organizations are already using AI in some form, yet only 7% report the kind of major, strategy-level impact that changes what a team can actually get done. Most are stuck in what researchers called an “efficiency plateau” — a faster first draft here, a quicker research pull there, nothing that adds up to real transformation. A separate 2026 survey of nonprofits with AI already in production by Coastal put a finer point on why: 72% or organizations are struggling with data accuracy or availability once the tools go live. Ambition outran infrastructure.
There is not a lot of love for the term data hygiene. It suggests a once-a-year deep clean, something you schedule and then forget about. What AI requires is closer to data integrity as an ongoing discipline — accurate contact and giving records, consistent formatting across every system that touches a constituent, de-duplicated households, current appended demographic and wealth indicators, and one trustworthy version of who someone is and what your relationship with them looks like.
A 2026 survey on data integrity and AI readiness by the LeBow College of Business at Drexel University found organizations with real data governance programs report high trust in their data at a rate more than 20 points greater than those without governance at all, which tells you the gap is not really about the AI at all.
A predictive model built on duplicate records will double-count your best donors and dilute your worst ones. A generative tool drafting outreach from an incomplete constituent history will write confidently to someone who no longer wants to hear from you or address them by the wrong name entirely. AI does not know it is working from bad inputs. It just produces bad outputs with excellent grammar and total conviction, which is a more dangerous failure mode than the systems it replaced, because it looks credible.
This work starts in an unglamorous place: Standardizing how records enter your systems: Consistent name and address formats; required fields at the point of entry; and, a defined process for merging duplicates instead of letting them pile up. This process accounts for most of the outcome. Once that foundation holds, the next step is that development, marketing, programs, and finance data usually live in systems that never talk to each other. Resolving identity across the whole constituent base before layering AI onto any single one of them means the model you build reflects one donor, not four fragmented versions of the same person.
Organizations extending existing data governance to cover AI, rather than standing up a separate AI policy, are those with notably differentiated performances. The piece often missed is measuring the output and confirming its accuracy and authenticity afterward. Using AI is not the goal. Match and deliverability rates holding steady, fewer manual corrections flowing back from staff, model recommendations that gift officers can act on instead of quietly overriding — that is what success looks like.
If your team spends more time fixing AI outputs than trusting them, the data underneath needs to be adjusted before the tool does.
Many conversations land on the same objection: We do not have a data team or the budget for one. That’s fair. But data integrity work scales down better than most technology investments do. Start with the segment that matters most financially (recurring donors, major gift prospects, whatever drives the bulk of revenue), and get that population’s records correct, updated and organized before repeating across the whole file.
Lean on shared and cooperative data resources rather than building identity resolution capabilities from scratch. The sector has spent decades building that infrastructure, so individual organizations do not have to solve it alone. Resist the pressure to buy an AI tool before looking hard at what you would feed it, practical use cases, desired outcomes, etc. The cheapest AI project is the one you never have to redo a year later because it was built on a shaky foundation.
Funders have started asking nonprofit leaders how they manage data as part of due diligence on AI use, not as an afterthought. Sector research by AI and You this year found that nearly 40% of nonprofits report having no staff formally trained in AI, and a similar group report no formal AI governance policy at all.
Closing that gap in the next year or two will matter more than any single AI pilot a team rolls out because none of those pilots will be worth much if the data feeding them cannot be trusted. That is the quieter competition underway right now, well below the headlines about AI.
*****
Richard Geiger is senior vice president at Data Axle Nonprofit. His email is Richard.Geiger@data-axle.com
