How to Build an Enterprise AI Buying Committee

The instinct when AI purchasing gets messy is to put everyone in a room. That instinct produces a committee that cannot decide anything, because its members have different jobs and the meeting treats them as equals with a shared vote.

A buying committee is not a consensus body. Its function is to take a single conversation that currently contains four separate decisions and pull them apart, so that each one has a named owner who can be held to it afterwards.

This is a different question from which executive title should own AI overall, which I have written about in Chief AI Officer vs CIO vs CTO. Ownership is about the org chart. This is about the decision.

Why the single conversation fails

Our September research measures something we call the Self-Attribution Effect. Ask each operating seat who signs off on AI purchases and every seat names itself more often than any other seat names it. The CEO seat names the CEO at 70 percent. The finance seat names the CFO at 63 percent. The technology seat is the exception that proves the point: it names the CIO or CTO at only 26 percent, which is less often than it names the CEO or the CFO. It is the one operating chair that does not put itself first.

Meanwhile, 11 percent of respondents overall report that no single seat owns AI purchasing, and that rises to 30 percent at the technology seat against 2 percent at the CEO seat.

Put those together and you get the practical situation inside most companies. Several people believe they decide. At least one part of the organisation believes nobody does. Both beliefs are held sincerely, and neither is tested until a purchase goes wrong or a renewal arrives.

A committee that produces one more group opinion does not fix this. A committee that assigns four decisions to four names does.

The four decisions

Decision one: is this a problem worth solving? Owner: the function that has the problem. Not IT, not the AI team, not the sponsor. If marketing wants a content system, marketing owns the statement of what is currently broken, what it costs, and what it would look like fixed.

The output is one page written before any vendor is seen. If that page cannot be written, the process stops here, and stopping here is the committee's highest-value outcome.

Decision two: where does the money come from? Owner: finance. Four sources, and they are not equivalent. Net-new money, money reallocated from other software, money that would have gone to headcount, and no clear budget at all. In our September finance data those run at 37, 18, 21 and 25 percent on a base of 290.

The headcount source is the one that needs to be named out loud. When 21 percent of AI funding across the finance lane comes from money earmarked for hiring, and 34 percent at the CEO seat, the company is making a workforce decision. That decision belongs in the minutes, not implied in a budget line.

Decision three: what can it reach? Owner: security. Not a veto and not a sign-off at the end. An access specification written at the same time as the requirements: what data, what systems, under whose credentials, and who revokes it.

Agent access is now the largest AI problem named by senior security leaders in our September data, ahead of data leakage, shadow AI and adversary attacks, and barely a third of those teams have a dedicated budget line to work with. If security joins the committee only at the end, it arrives with an objection and no money, which is the worst position from which to be useful.

Decision four: who signs? Owner: named in advance, in writing, by threshold.

This is the decision the Self-Attribution data says companies get wrong. The point is not that the CEO should or should not sign. It is that the answer should exist before the purchase rather than being reconstructed afterwards. Write two thresholds: below X, the function head signs; above X, a named executive signs. Then check it against reality by pulling the last three AI purchases and seeing who actually approved them.

Who is allowed to start a process

There is a fourth structural problem the committee has to handle, and it comes from the supply side.

We ask AI founders which seat they sell into. The answers split three ways: business-unit leader 39 percent, CIO or CTO 36 percent, CFO or finance 28 percent, on a base of 148. Sellers are not converging on one door, because there is no consistent door to converge on. We call the distance between the door the seller enters and the seat that eventually signs the Seat Split.

The practical consequence is that a process can begin in three places, and two of them will not tell the committee it has begun. So the committee needs one rule: any AI evaluation above a defined size gets registered when it starts, not when it needs approval. Registration is not permission. It is visibility, and it costs the requester five minutes.

The two meetings

Meeting one, before vendors: the problem review. Attendees are the requesting function, finance and security. The agenda is decisions one, two and three. The output is either a written brief or a stop. This meeting should be short and most of them should end in a stop.

Meeting two, before signature: the commitment review. Same attendees plus the signer for that threshold. The agenda is the payback date, the access specification, and confirmation that the money source has not quietly changed between meetings. It frequently has.

A committee that needs a third standing meeting has stopped being a decision structure and become a queue. If evaluations are backing up, the thresholds are set too low, not the meeting cadence too light.

What good looks like after two quarters

You should be able to answer four questions in under an hour: who asked for the last significant AI purchase, where the money came from, what the system can reach, and who signed. If any of those takes longer than an hour to establish, the committee is producing meetings rather than decisions.

Limitations

Two different populations are quoted above and they should not be blended. The sign-off, ownership and funding figures come from operators, on bases of 290 for the finance lane and 80, 52 and 23 for the CEO, finance and technology chairs. That last one falls short of our 40-response floor, so the 30 percent unowned figure is a direction rather than a measurement. The three-way buying doorway comes from a separate base of 148 YC founders and describes how sellers experience entering a company, which is not the same thing as how any one company organises itself.

All of it was collected between March and August 2026 from executives applying to AI-focused sessions, a group further along than most. Multi-select questions push some totals past 100 percent.

The committee design itself is not a research finding. It is drawn from how executives in these rooms describe what worked, and it should be adapted to your own approval structure.

Where this gets worked out in practice

Open Future Forum convenes CEOs, CFOs, CISOs and AI leaders in private rooms where purchasing structure is a recurring subject. The September Executive AI Leverage Report carries the full sign-off data with response bases on every figure, and the calendar of executive events is here.

Last updated: September 8, 2026