Field Note #4: Vendor Questions That Need Answers?
You evaluate AI vendors on accuracy, speed, and cost. Yet, almost never ask the question that determines whether the tool works for your entire workforce.
Governance Without a Body | Field Notes | AI Procurement | 15 min read
This is the fourth field note in Governance Without a Body. Last week I argued that the regulatory floor dropping in 2026 makes voluntary governance the only floor left. This week I want to get specific about where that voluntary floor actually gets poured — not in a policy document, not in a values statement, but in the procurement process, in the questions you ask before you sign. Because the uncomfortable reality of this summer is that the questions you were counting on a regulator to ask, a regulator may no longer ask. Which means if you do not ask them, no one will.
THE THREE-QUESTION PULSE
Two minutes. It keeps the calendar pointed at what you actually need.
Did the 2026 EU AI Act and Colorado rollbacks change your organization’s AI governance plans? (Sped us up / Slowed us down / No change / We weren’t tracking them)
Where does AI governance currently live in your organization: legal/compliance, IT/security, a dedicated AI function, DEI, or nowhere clearly?
Which would help most next: a vendor-evaluation question bank, a model-card reading guide, a one-page regulatory-status tracker, or contract language you can adapt?
When your organization evaluates a new AI vendor, the procurement questions are predictable, and they are not wrong. What is the accuracy rate? What is the latency? What does the pricing model look like? Will it integrate with our existing stack? These are necessary questions, and they get asked rigorously, because the people asking them will be held accountable if the tool is slow, expensive, or incompatible.
Here is the question that almost never gets asked: for whom does this tool underperform, and by how much? Consider how many AI systems your organization has already deployed without anyone in the room asking that question even once.
The reason it goes unasked is not malice and not laziness. It’s because the question disturbs a comfortable assumption — the assumption that a tool’s published accuracy rate applies more or less equally to everyone who will use it. That assumption is false for any tool that processes faces, voices, names, or bodies, and I have documented exactly how false it is on my own face and my own voice. But the assumption is comfortable, and comfortable assumptions survive until something forces them into the open.
For a while, it looked like regulation would be that forcing function. That is the part that changed in 2026, and it changed the calculus for every organization that had been quietly planning to let the law do its governance for it.
