The Model in Your Close Came With a Release Note

A founder I work with runs a 34-person B2B services company at $19M. Last month an enterprise customer sent over a vendor review. Question 11 asked her to name every system used in her financial reporting process that incorporates AI, and to describe the assurance she obtains over each one. She forwarded it to me with a single line: "do we?"

She did. Three of them. Nobody at her company had decided on any of them.

The decision got made at your vendor

Her AP platform shipped GL coding suggestions in a February release. Her close tool moved reconciliation matching from a rules engine to a model sometime in Q1 and mentioned it in a changelog. No contract amendment, no evaluation, no security review. Somebody's inbox got a product email, and by spring a model was proposing account codes that her bookkeeper approved with two clicks.

The suggestions are good. That's not the issue. What changed is that a judgment step moved out of her team and into a system she can't inspect, and the paperwork she'd point to if anyone asked was written before that system existed.

Your vendor's report covers the wrong layer

A SOC 1 report tells you about controls over the platform. Access provisioning, change management, whether the coded logic processes transactions the way the documentation says. What it usually doesn't reach is the model itself: training data, version control, or how behavior drifted between the report period and the invoice you coded on Tuesday.

Auditors know this. The AICPA now has an AI-specific examination framework covering data provenance and model lifecycle governance, and vendors are starting to adopt it. Adoption is early. Meanwhile the SEC's deputy chief accountant, Michal Dusza, said the quiet part at a reporting conference in June: "Does management even have the right handle on what these systems are?" You're not an SEC registrant. Your acquirer's diligence team reads those speeches anyway, and so does the bank underwriting your next line.

Nobody is asking whether the model is any good. They are asking whether you can name where it sits and show what you check.

The gap has a name, and it belongs to you

Complementary user entity controls. Ugly phrase for a simple idea: the checks a service provider's report openly assumes you're performing on your end. When the vendor's assurance stops at the platform boundary, those controls are the only thing standing between you and an unverifiable close.

Start with an inventory. One page, every system touching the close, whether it suggests or decides, and who reviews the output before it posts. Write it while the people who turned these features on still remember doing it. Then sample. Twenty AI-coded invoices a month, checked against source documents, with the exceptions logged somewhere durable. That's an hour of work and it's the difference between a claim and a control.

Last, ask each vendor two questions in writing. Does your product use AI in the functions we use. What assurance do you provide over the model, and can we see it. Save the replies. A vendor that answers in a paragraph is telling you something. A vendor that routes you to sales is telling you more.

You didn't buy a model. You got one anyway, and the only file proving you knew is the one you write.