Setting AI spending limits when a family board holds the purse
Approval thresholds, pilot budgets and review points for AI spending, and the points where the public sources we checked stop short of helping.

A family board that controls the purse often treats AI spending the way it treats a new roof: one line in the annual budget, approved once and then forgotten until the invoice arrives. The trouble is that AI costs do not behave like a roof. Licences renew monthly, a pilot quietly acquires a second team, and a consultant's retainer becomes the item nobody can explain. We could not find a published benchmark for how much a family company should put into AI, so no figure in this guide is a recommendation. Every threshold below is one you set, and the reasons for setting it should be written down where the next chair can read them.
Before you start
The public sources we checked cover three things: a voluntary risk framework for AI from the US National Institute of Standards and Technology, a warning from the European Central Bank about exposure to a few large technology issuers, and a preliminary field study of when entrepreneurs decide to stop. None of them addresses family boards, board charters or spending limits. What follows is therefore editorial judgement, labelled as such, and each threshold should be checked against your own charter and your own accounts before it reaches a vote.
Before you write a single threshold, read the board charter. If it says nothing about technology commitments above a set size, fix that first. Confirm who can sign a contract, who can stop one, and whether the chair's authority differs from the full board's. If a threshold touches tax treatment, lending covenants or employment terms, put the question to your adviser, a lawyer or an accountant. This guide does not settle those points for you.
- A written definition of AI spend that includes licences, cloud usage, outside consultants and the salary share of staff who work on AI.
- A register of every current AI-related contract, with its renewal date and notice period.
- One named executive owner for each AI commitment, who can be asked to explain it to the board.
- A list of decisions the board has already delegated, so that the same item is not approved twice.
- Step 1
Define what counts
Start with a definition a finance lead could apply to a supplier invoice without making a phone call. It should cover software licences, cloud usage, outside consultants and the staff time spent on AI work. What usually goes wrong: the definition covers only capital purchases, which leaves the largest recurring costs outside any review.
- Step 2
Tier by reversibility
Set three approval tiers, and place each commitment by how hard it is to undo as well as by its amount. A monthly licence that can be cancelled on thirty days' notice belongs in a lower tier than a three-year contract of the same value. Check the notice period in the signed contract, not in the sales deck. What usually goes wrong: tiers are set by amount alone, so a cheap but irreversible commitment slips through.
- Step 3
Fix the pilot ceiling
Give every pilot a fixed budget, a named owner and one question it must answer by its end date. The ceiling should be small enough that a failed pilot is an acceptable lesson, and large enough that the answer is not noise. What usually goes wrong: a pilot without an end date drifts into production because nobody is required to decide.
- Step 4
Write the stop rule first
Before any money moves, write down the evidence that would end the pilot and who must see it. Settle this at the start, while no one is yet attached to the project. The NBER working paper on entrepreneurship training suggests that recognising when to stop can be taught, though the study is preliminary and its excerpt gives no sample size or effect size. What usually goes wrong: stop conditions are written after the spend, shaped to fit what has already happened.
- Step 5
Book the review points
Put review dates into the board calendar at the moment of approval, not 'when needed'. Tie each date to a decision: renew, expand, reduce or end. What usually goes wrong: reviews happen only after a complaint, and by then the contract has already renewed on its old terms.
- Step 6
Set a concentration limit
Check how much of the AI spend, and of the family's wider exposure, depends on one supplier or one group of large issuers. The European Central Bank says the growing hyperscaler presence, though early stage, could amplify portfolio losses if AI-related revenues disappoint. Its page gives no threshold, so any limit you choose is yours to justify in writing. What usually goes wrong: several products from one vendor are counted as diversification.
- Step 7
Cite the framework with care
If the board wants an external reference, use the NIST AI Risk Management Framework as a list of risks to consider, not as a rulebook. Check the current version first, because NIST says AI RMF 1.0 is being revised as part of the White House AI Action Plan. What usually goes wrong: a board cites the 1.0 text as settled and later finds that the reference has moved.
The mistakes that recur
Three mistakes come up again and again when boards set limits on new technology. We offer them as editorial judgement, not as findings from a survey. The first is setting thresholds by amount alone. A gate that opens for anything under a set figure, without asking whether the spend can be undone, gives the board comfort it has not earned. The second is approving a pilot and then approving its expansion as a formality. The second approval should be harder than the first, because by then the project has a team, a supplier and a sponsor who wants it to succeed.
The third mistake is treating review as an event rather than a calendar entry. Boards that review on a fixed date tend to ask the same three questions each time: what did it cost, what did it deliver, and what would end it. Boards that review only when something goes wrong ask none of them until the damage is already visible. A review date is cheap to book and expensive to skip.
The NIST description of its framework is useful here, because it sets the scope of any external reference a board might cite. A voluntary framework is something a board chooses to use, not a rule that binds its approval process. Your thresholds still have to come from your charter and your own figures.
intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI
- Should we set one threshold for all AI spending?
Usually not. A single figure treats a cancellable licence and an irreversible multi-year contract the same way. Use at least two or three tiers, and place each commitment by reversibility as well as by amount. Record the reasoning for each tier in the charter or in a board minute so that a later chair can see why it was set.
- Do we need an outside adviser to design this?
For any question about tax treatment, lending covenants or employment terms, yes. Put those questions to a lawyer or an accountant before the thresholds are voted on. The rest of the design, including tiers, pilot ceilings and review dates, is a governance task the board can do itself, provided the charter is clear about who decides.
- Which AI framework should we cite in the board pack?
NIST's AI Risk Management Framework 1.0 was released on January 26, 2023, and NIST describes it as intended for voluntary use. NIST also says the 1.0 version is being revised as part of the White House AI Action Plan, so check the current version before you cite it. Treat any reference as a checklist of risks to consider, not as a requirement.
- How long should a pilot run?
None of the sources we checked gives a duration, so there is no published guideline to quote. Set the length from the question the pilot must answer. If the question can be answered in one quarter, a longer pilot wastes money. If it needs a full billing cycle to show results, do not end it early for reasons that are only about cash.
- Does the research on stopping mean we should cancel pilots more often?
Not necessarily. The NBER working paper, number 34755, is a multi-year randomised field experiment with innovation-driven entrepreneurs. It reports that training may help founders recognise when a venture no longer warrants further investment, and that some startups shut down sooner as a result. It also finds that training can improve outcomes for others even as it shortens some ventures. The excerpt we reviewed gives no sample size, effect size or completion date, so treat it as preliminary.
This week, do three things. Pull the register of AI contracts and read every notice period in it. Ask the chair to put one proposed threshold on the next board agenda, with its reasoning written beside it. Send the tax, covenant and employment questions to your adviser before any threshold reaches a vote. Most of the work is in the first two; the third simply keeps the board from deciding something it cannot later undo.