The $2,400 Charge Your Board Treasurer Can't Map to a Program: Capping AI Spend Before the 990 Asks

You already know the objection because you have made it. Nobody on your finance committee is going to approve a budget line that reads "AI, amount to be determined," and the last thing an executive director needs is a usage-based charge that arrives at $340 in February and $2,900 in June with no way to explain which programme spent it. In a sector where every dollar has to land in one of three columns on the Form 990, an expense you cannot allocate is worse than an expense you cannot afford.

That objection was correct through most of 2025. It is much weaker now, and the reason is specific.

The 2 July governance update, in plain budgeting terms

On 2 July 2026, Claude Enterprise added the controls that make this a normal line item: a dashboard showing usage and cost by person and by team, spending limits you can set for the whole organisation and for individual users, automatic alerts when spend crosses 75% and again at 90% of the limit, defaults for which model people get, and entitlements controlling who can use what. There are reporting feeds a finance office can pull into a spreadsheet on the same day of the month it pulls everything else.

Read that as a nonprofit CFO would. A cap means the bill cannot run away. Per-user visibility means the charge can be split across programmes with something defensible behind it. Alerts at 75% mean you find out mid-month, not when the statement lands. Entitlements mean the grants manager and the development director can have the good tools while the twelve people who use it twice a month cannot accidentally spend the budget.

Budgeting AI in a nonprofit means treating it exactly like your phone system or your database subscription: a capped monthly amount, allocated across programme, management and fundraising on a stated basis, reviewed by the finance committee on the same cycle as everything else.

The three columns problem, and why per-person data solves it

Form 990 Part IX makes you split functional expenses into programme services, management and general, and fundraising. Your auditor wants the allocation method written down and applied consistently. A single organisation-wide software subscription with no usage detail gets shoved into management and general, which is the column that makes your overhead ratio look worse on Charity Navigator and Candid — and every development director in America has sat through a donor meeting about that ratio.

With per-user spend data, the split is defensible. If the grants manager and the two programme directors account for 62% of the spend, 62% of it is programme. If the development director and the database administrator account for 24%, that is fundraising. The finance director writes the method into the cost allocation plan once and applies it monthly. That is not a trick; it is exactly what you already do with the phone bill and shared occupancy costs.

If you carry federal awards, this matters more. Under Uniform Guidance, 2 CFR 200, a cost charged to an award has to be allowable, allocable and consistently treated. Software charged directly to a grant needs a basis showing it benefited that grant. Usage by person, mapped to the people who work on that award, is that basis. If you are using the de minimis indirect rate — raised to 15% in the 2024 revision — you may simply keep it in the indirect pool, but decide which and be consistent, because inconsistency is the finding, not the treatment you chose.

flowchart TD
  A["Month opens with an organisation cap set"] --> B["Staff use their assigned seats"]
  B --> C{"Spend crosses 75%?"}
  C -->|No| D["Month closes normally"]
  C -->|Yes| E["Alert reaches Finance Director and ED"]
  E --> F["Check which team drove it and why"]
  F --> G{"Legitimate programme surge?"}
  G -->|Yes| H["Raise that team's cap, note the reason"]
  G -->|No| I["Trim entitlements or set a cheaper default model"]
  D --> J["Allocate across programme, management, fundraising in Sage Intacct"]
  H --> J
  I --> J
  J --> A

A number to put in the budget for a 22-person organisation

Start with who actually needs it rather than a per-head licence for everyone. Assume a $4.8 million organisation: an executive director, a finance director, a grants manager, three programme directors, a development director, a database administrator, a communications manager, and fourteen frontline staff.

Those seat amounts are illustrative and you should set yours from a first month of actual usage rather than from this table. The point is the shape: a hard ceiling of $7,800 that the board approves once, split roughly 60/25/15 across programme, fundraising and management using the per-person data, and reviewed quarterly.

Then put a denominator next to it. If the grants manager's share is $1,080 a year and it takes measurable time off eleven funder reports and six proposals, that is $63 per report or proposal. Say that in the finance committee meeting and the conversation ends, because the same committee just approved a grant writer at $85 an hour. Prices for capable models have fallen roughly tenfold from 2025, which is precisely why a number like this is now small enough to be boring — and boring is what you want in a budget line.

What the caps will not save you from

A cap controls the bill. It does not control four other things, and it is worth being blunt about them.

A cap is not a data policy. Spend limits say nothing about whether someone pasted a client's case notes, a donor's giving history, or a Schedule B into a chat window. Write the rule separately: no client-identifying information, no donor financial detail, no personnel matters, use the organisation's account and never a personal one. Then check that the account you are paying for does not train on your content, and keep that in writing with your other vendor agreements.

Alerts still need an owner. A notification at 75% that lands in a shared inbox nobody reads is decoration. Name the finance director, put the monthly review on the same checklist as the bank reconciliation, and give the board treasurer the quarterly summary.

Caps can quietly break the thing that was working. The most common failure is not overspend, it is a cap set too tight in the month a large federal proposal is due, so the grants manager runs out on the 22nd and goes back to doing it by hand at midnight. Build a documented exception path — the executive director can raise a single team's cap up to a stated amount, with a note, and it gets reported to the finance committee after the fact.

State rules still apply. Texas TRAIGA and California SB 53 both took effect on 1 January 2026, and Colorado, New York, Utah, Nevada, Maine and Illinois have their own AI statutes. Whether federal law will override any of them is unsettled as of July 2026, so if you run programmes in those states, ask your counsel one narrow question: does anything we use to screen, prioritise or deny services to clients fall under our state's law? Budget controls are not compliance and no dashboard will tell you the answer.

The first finance committee meeting

Bring one page. Line one: the organisation cap you are proposing and what it buys. Line two: who has seats and why those people. Line three: the allocation method you will use across the three functional columns, in one sentence. Line four: who gets the alerts and when the review happens. Ask for approval of the cap, not of "an AI initiative." Nobody has to vote on the future of technology; they have to vote on $650 a month with a ceiling and a named owner, which is a motion a treasurer knows how to make.

Frequently asked questions

Can we charge this to a federal grant?

Sometimes, if it is allowable under your award terms, allocable to that award, and treated consistently with how you treat similar costs. Usage detail by person is what makes the allocation defensible. If it is easier and defensible, keep it in the indirect pool under the de minimis rate instead. Ask your auditor before the fiscal year starts, not during fieldwork.

What if a funder asks whether we used AI to write the proposal?

Answer honestly and have a written internal rule, because more funders are asking. A reasonable rule: staff may use it for drafting, editing and pulling together material from our own documents; a named human is responsible for every factual claim, every budget figure and every outcome number in the submission. Some funders now require disclosure. Read the guidelines before, not after.

How do we stop one person from spending the whole budget?

Per-user limits, which is precisely what the July update added. Set the individual caps lower than you think you need for the first two months, watch who hits them, then adjust. It is far easier to raise a cap for someone doing good work than to explain a surprise charge to a board treasurer.

Do very small nonprofits need any of this?

If you have three staff and one shared account, the dashboard is more than you need — but set an organisation spend limit anyway, because the whole reason it exists is that usage-based charges surprise people. Five minutes now prevents the conversation where your entire technology budget went somewhere nobody can explain.

One place the spending question comes up quickly is the phone. CallSphere builds AI voice and chat agents that answer nonprofit phone lines and web chat around the clock, answer routine donor and programme questions, book appointments, and capture who called and why — and because that call volume is countable, it is one of the easier lines to budget, cap and report to a finance committee alongside everything else.