AI for Nonprofits: How to Reduce Cost Without Creating Audit Risk

The Overhead Problem Nonprofits Actually Have

Every nonprofit finance leader has had a version of the same conversation with a funder or a board member: why is so much money going to administration instead of programs? It is a fair question asked in an unfair way, and in 2013 the three largest charity evaluators — GuideStar, Charity Navigator, and the BBB Wise Giving Alliance — jointly published an open letter asking donors to stop using overhead ratio as the primary measure of a nonprofit's worth. Their argument was simple. Starving your finance function does not make you efficient. It makes you fragile.

But the pressure is real, and it lands somewhere. In most organizations we audit, it lands on the back office. One person doing bookkeeping in evenings around a program job. A part-time contractor who codes transactions from memory. A finance director spending the first week of every month on data entry instead of analysis.

That is the actual overhead problem: not that nonprofits spend too much on administration, but that they spend it on the lowest-value part of administration. Categorizing transactions is not judgment work. It is pattern-matching. And pattern-matching is the one thing current AI does genuinely well.

Where AI Actually Reduces Cost — and Where It Does Not

Be specific about this, because the gap between the two lists is where organizations get into trouble.

What AI handles well

• Transaction categorization. Assigning the coffee shop charge, the airline ticket, and the office supply order to the right accounts. High volume, repetitive, rule-shaped.
• Receipt matching. Connecting a photographed receipt to the card charge it belongs to, without a human comparing two screens.
• Duplicate and anomaly detection. Catching the invoice paid twice and the charge that does not fit the pattern — the kind of thing a tired human misses at 9pm.
• Backlog cleanup. Months or years of uncategorized transactions. This is the worst job in nonprofit finance and the one AI is best suited to.

What AI does not handle

• Functional expense allocation methodology. Deciding how the executive director's salary splits between program, management, and fundraising is a judgment your board owns. AI can apply the split once you have decided it. It cannot decide it for you, and it should not appear to.
• Grant restriction and release decisions. Whether a restriction has been satisfied is a reading of the grant agreement, not a transaction pattern.
• Revenue recognition on complex contributions. Conditional contributions, multi-year pledges, and in-kind valuations require professional judgment.
• Anything that requires knowing what your organization intended. AI reads the record. It does not know the conversation the record came from.

Four Rules That Keep an Automated System Defensible

We audit nonprofits for a living, which means we see what happens when automation is implemented carelessly. These four rules are not compliance theater. Each one exists because we have seen its absence produce a finding.

1. Name a human reviewer, in writing

AI output can be wrong. Not occasionally — predictably, in the cases that are ambiguous. Every automated categorization must pass through a named human before it becomes part of your books, and that person's name should appear in your written procedures. When an auditor asks who reviews the AI's work, "the system handles it" is a finding waiting to happen. "Our staff accountant reviews the exception queue weekly, per section 4 of our accounting procedures" is not.

2. No automation gets authority to move money

Categorizing a payment and initiating one are entirely different risk categories. An AI system should read your accounting records and propose changes to them. It should never hold payment authority, banking credentials with transfer rights, or the ability to originate an ACH. Keep your disbursement controls exactly where they were — dual authorization, segregation of duties, board-approved signing limits. Automation changes who does the bookkeeping. It should not change who can spend.

3. Your allocation methodology stays a written human policy

Form 990 Part IX requires you to present expenses by functional classification, and your financial statements require the same. The methodology behind those allocations — how you split shared salaries, occupancy, and technology across program, management and general, and fundraising — must be documented, board-approved, and applied consistently year over year. Let the system execute the policy. Do not let the system become the policy. If your only answer to "how did you arrive at this allocation" is "that is how the software did it," you do not have a methodology.

4. Document the system itself as part of your control environment

Auditors evaluate internal controls, and an AI-assisted bookkeeping process is now part of yours. That means it needs the same documentation as any other control: what the system does, what it does not do, who reviews its output, how exceptions are escalated, and what happens when it is wrong. Organizations that implement AI and update their written procedures at the same time have a stronger control environment than they started with. Organizations that implement AI and tell nobody have a weaker one.

What This Actually Saves

Honest answer: it depends on how much of your finance cost is currently going to repetitive work, and we cannot tell you that without looking at your books.

The structure of the saving is easier to describe than the size of it. If a bookkeeper spends the bulk of their hours categorizing transactions and matching receipts, and most of that work moves to a system that asks a human only about the cases it cannot resolve, the hours spent on the remainder are the ones that need paying for. That is not a headcount argument. In most nonprofits we work with, the better outcome is the same person doing higher-value work — grant reporting, budget-versus-actual analysis, board packages — rather than one fewer person.

There is also a second-order saving that nonprofits consistently underestimate: clean books make audits cheaper. A meaningful portion of audit fieldwork cost is auditors reconciling records that should already reconcile. When the general ledger is current, transactions are categorized consistently, and reconciliations are complete before fieldwork begins, the engagement takes less time. That is true regardless of which firm you hire.

The Program We Offer

Summit Advisory builds AI-assisted accounting systems for nonprofits. The system runs on COUNT paired with Claude, connected to your bank and card feeds, your chart of accounts, and your invoicing. You own the software and hold the subscriptions directly — those are billed by the vendors and are separate from our fee.

The build works in three steps. We set up and connect the system. The AI does the repetitive work — categorizing, matching receipts, catching duplicates and odd charges. Then it asks you about the rest, in plain language: is this meal a client meal or personal? You answer in one click. No accounting vocabulary required.

• Done-For-You — $7,500 one-time. We build the whole system and clean up to five years of back books. You get written procedures, live training, and a system that is current on day one.
• Guided — $2,500 one-time. We design the system and walk you through building it yourself, then check your work. Current books only, no historical cleanup. Best if you are already caught up.

If you are behind on your books, the first option is what that is for. Years of uncategorized transactions is the worst job in nonprofit finance, and it is the job this technology handles best. We point it at the backlog, it works through the pile, and it brings you only the decisions it cannot make alone.

One Thing We Will Not Do

We will not build or run your accounting system and audit you.

Auditor independence is not a formality. If our firm sets up your books, categorizes your transactions, or maintains your accounting records, we are no longer independent with respect to your financial statements — and an audit opinion from a firm that is not independent is worth nothing to your board, your funders, or a federal agency reviewing your Single Audit.

So the choice is yours, and it is a real one. If you are a current or prospective audit client, we will help you find a qualified implementation partner and we will tell you exactly what to ask for. If you would rather we build the system, we will do that well — and we will refer you to another firm for your audit. What we will not do is quietly take both engagements and let someone else discover the conflict later.

We think that is the right way to run a firm. It also happens to be the reason you can trust the four rules above: we wrote them from the auditor's side of the table.

Where to Start

Before you evaluate any system, get honest about two things. First, how far behind are your books — not how far behind you think, but the actual date of your last complete reconciliation. Second, how many hours a month currently go to categorization and receipt matching. Those two numbers determine whether this is worth doing and which version fits.

We will do that assessment with you in thirty minutes at no charge, and we will tell you straight if the answer is that you do not need us.