Grant Compliance Is Mostly Documentation. That's Exactly What AI Is Good At.

Two sections of the Uniform Guidance tell you what grant accounting really is on a normal Tuesday. Section 200.413(c) lets you direct-charge administrative and clerical salaries only when the services are integral to the award, the individuals are specifically identifiable, and the costs aren't also recovered as indirect. Section 200.414(f) lets you elect a de minimis indirect rate of up to 15 percent of modified total direct costs. Both of those are documentation jobs - you're producing defensible evidence that a rule got followed - and producing evidence is exactly the category of work that's gotten cheap in the last two years.

We track more than 80,000 accounting and finance job postings, and nonprofit finance reads like its own dialect inside that data: restricted net assets, fund accounting, Uniform Guidance, single audit prep, SEFA. What none of those postings say out loud is that the person in the seat spends most of the week generating paper to support decisions that were already made. That's the part worth handing off.

What does fund accounting actually take, most days?

Restriction tracking and evidence. You're keeping donor-restricted and unrestricted net assets separate, releasing restrictions as conditions get met, tying each federal dollar spent back to a budget line in an award document, and keeping enough support that an auditor a year from now can walk the same path you walked. The judgment calls are real but they're rare, and the paper is constant.

And the paper compounds, because every grant carries its own period, its own budget categories, its own reporting calendar, and its own idea of what counts as allowable. Run six federal awards and two foundation grants and you're maintaining eight parallel sets of rules on top of one general ledger. That's the grind, and it's a big damn reason nonprofit finance burns good people out.

Why the functional expense split is really a mapping problem

Since FASB's ASU 2016-14, every not-for-profit has to present expenses by both natural classification and function - program services against management and general against fundraising - and disclose the methods used to allocate costs across more than one function. Before that update, only voluntary health and welfare organizations had to produce it.

Mechanically, you're building a matrix. Your GL gives you the natural rows: salaries, rent, software, travel. You need the functional columns, plus a written, consistent rationale for how a salary that spans three functions got split. If your executive director spends 15% of her time on fundraising, that allocation needs a basis you can point to - time records, a time study, square footage, headcount - and you need to describe the method in the notes.

A mapping with a stated basis and a written rationale is a machine-shaped task. The inputs already sit in your payroll system, your time tracking, and your lease. The work is assembling them into a defensible narrative every single period, forever, which is why it slides.

Where AI is genuinely good enough right now

Four places I'd hand off today, roughly in this order:

Allocation memos. Feed it the time data and the allocation basis you've already chosen and let it draft the memo explaining the split for each cost pool. You end up editing a first draft instead of staring at a blank page on day four of close.

GL-to-function mapping. Give it your chart of accounts, your program list, and last period's mapping, and have it propose the functional split for the current period with the unusual accounts flagged. It gets the routine 90% right and surfaces the weird ones, which is the whole point of the exercise.

Allowability screening. The selected items of cost in Subpart E run from 200.420 through 200.476 - alcohol at 200.423, entertainment at 200.438, fundraising at 200.442, lobbying at 200.450, travel at 200.475 - each with its own rule. Having a model read a month of transactions against that list and flag the ten it wants you to look at beats the realistic alternative at most organizations, which is nobody reading them at all.

Audit binders. Grant audits and single audits run on request lists, and assembling support against a request list is retrieval plus formatting. Point a model at the request list and your document store and let it build the first draft of the binder while you work on the answers that need you.

It's the same logic as rolling reconciliations: continuous, well-documented grant accounting stayed theoretical because doing it by hand meant doing the annoying part more often. When the annoying part gets drafted for you on a random Tuesday, continuous quietly becomes the cheap option.

Where it will absolutely burn you

Allowability judgment at the edges, cost transfers, and anything ending in a signature. A model will produce a confident, well-written rationale for an allocation that is flatly wrong, and it produces it in exactly the same tone it uses when it's right, which is the part that should keep you careful. Grant auditors ask you to defend a basis, and a paragraph and a basis look identical until somebody pushes.

Read 200.415 before you get comfortable. The certification on your federal financial reports gets signed by an official authorized to legally bind the organization, attesting that the report is true, complete, and accurate, with the false-statement and false-claims statutes named right there in the certification language. No model signs that. A person does, and that person carries it. So treat every AI output in this workflow as a draft a human reviews against source documents, and KEEP the source documents.

The other honest limit is your input data. If your time tracking is three people guessing at the end of the month, AI just gives you faster bullshit, at scale, with better grammar. Fix the input first - unglamorous, and it's the whole ballgame.

What I'd do this month

Pick one grant. Preferably a mid-sized federal award with the messy allocation that pisses you off every quarter, the one you quietly dread. Take last month's actuals and build the functional mapping and the allocation memo with whatever AI tool you already have access to, then diff the result against what you produced by hand. You'll learn in an afternoon where it's reliable and where it's confidently wrong, and there's no way to get that from reading about it.

Then do the next grant. If your accounting system fights you the entire way, and plenty of nonprofit stacks do, our software directory and our breakdown of accounting software for nonprofits are the right place to start, because no amount of AI rescues a general ledger that can't hold a restriction properly.

Worth staying level-headed about the tooling while you're at it - we looked at what Klarna's CEO actually said after everyone cited him as proof you can rip out your stack, and the honest version is a lot more boring than the headline. Nonprofit finance has been underserved by software and underpaid by everyone for about thirty years, which is a hell of a thing when you look at how much federal money runs through these organizations every year. The people who change that will be the ones who built something. Start with one grant, and go find out what it can do.

Frequently asked questions

Can AI do nonprofit fund accounting?

AI handles the documentation layer of fund accounting well - drafting allocation memos, mapping GL accounts to functional categories, screening transactions against the cost principles, and assembling audit support. The judgment calls on allowability and the certifications on federal reports stay with a person.

What is the de minimis indirect cost rate under the Uniform Guidance?

Under 2 CFR 200.414(f), recipients and subrecipients without a current federally negotiated indirect cost rate may elect a de minimis rate of up to 15 percent of modified total direct costs. It requires no documentation to justify its use and may be used indefinitely, but once elected it must be applied consistently across all federal awards until you move to a negotiated rate.

Does ASU 2016-14 require a statement of functional expenses?

It requires all not-for-profits to present the information ordinarily found in a statement of functional expenses - expenses by both natural and functional classification - either as a separate statement, within the statement of activities, or in the notes, along with disclosure of the methods used to allocate costs across functions.

Is it safe to use AI on federal grant compliance work?

It's safe as a drafting and retrieval layer with human review against source documents. It's risky as an unreviewed decision-maker, because 2 CFR 200.415 puts a binding certification on financial reports and a confidently wrong allocation rationale reads exactly like a correct one.

What should a nonprofit finance team automate first?

Start with functional expense mapping and allocation memos, since both are repetitive, rule-driven, and already supported by data in systems you own. Move on to allowability screening and audit binder assembly once you trust the review loop.

The 80,000+ postings figure comes from Audit Friendly's own database of accounting and finance jobs; regulatory citations are from the eCFR and FASB. Hiring for a nonprofit finance seat, or looking for one? Browse the job board.