A buyer asked us for our SOC 2 report last week. Fair question, wrong place. He wanted assurance on the platform: where the data sits, how it is protected, who can reach it. Microsoft answers that, many times over, in its own trust centre. What he didn't ask is the question that carries the real risk: who checks the firm building on that platform?
In the mid-market, the AI platform is the most audited thing in the building. The agent sitting on top of it, built by a partner, by one of your own staff, or by one of the many vendors pitching you this month, usually isn't covered by anything. That agent can read your data, act in your systems and make decisions for your people. If nothing covers it, nothing guarantees it was tested by anyone but its maker, logs what it does, has an owner or can be switched off.
Building an agent has stopped being the hard part. Anyone can do it now: a partner, a vendor, your own staff. The sharpest buyers have already worked that out. A technology leader at a ten-hospital veterinary emergency network was blunt with us: "I've seen many demos where the outcome is excellent." He said he could write the ROI case for his board himself. What he wanted to see was "how do you actually build it out?" and whether it was "a technology that I can control." The AI lead at a community services organisation kept returning to one question in our session: "What's the design intent?"
These buyers are not short of options. One of our team reports that a mid-market prospect is speaking to around ten AI vendors at once.
Nor does a certificate settle it. ISO/IEC 42001 is becoming a procurement checkpoint, but it attests to a management system, not to any specific model or agent. A vendor can hold a valid certificate whose scope excludes the system you are buying.
They ask the provider for the platform's paperwork, and they assume the platform's paperwork covers the provider. Neither holds. Platform certifications tell you the foundations are sound. They say nothing about whether the agent built last week was tested as anyone other than its maker, logs what it does, or has a named owner.
And they judge on the demo. A demo proves an agent can do something once, on clean data, with a presenter who knows what to type. Your production environment offers none of those.
I'll own our part of this. Recently, one of our own team built an AI experiment that read security alerts and raised tickets. It ran in production and created about sixty incidents that weren't incidents. Nothing was harmed. The build worked. The operating rules around it didn't exist yet. That gap, between an agent that works and an agent that is governed, is exactly where the buyer carries the risk.
We have just been certified to ISO/IEC 42001. The certificate isn't the point. The point is that it forces us to answer, for every agent we build or run, the questions most buyers aren't asking yet: what it is for, what it can reach, how it was tested, who owns it, and how it gets switched off. A managed service provider points you to the platform's trust centre. A managed intelligence provider has to show you its own method.