Every accountability chain in your organization rests on one load-bearing assumption: somewhere in it, a human made the decision. So that’s who the RACI points at. And that’s who signs off. Run an agent on what would take a human team twelve hours, and that check quietly disappears. The work happens in the gap between the prompt and the output, with nobody there.
And the safety architecture you inherit isn’t an industry-wide standard. It’s authored contract by contract, vendor by vendor. The Pentagon’s recent vendor list made the dynamic visible: Anthropic refused to alter its safety stance and was left off; everyone else signed contracts carrying the same restrictions Anthropic was excluded for. The buyer wanted the world’s sharpest chef’s knife. Then they asked the maker to remove the handle.
The Governance Gap
Once you accept those two things, the harder edge of the question lands. Now, Medvi is a two-person telehealth company on track for $1.8 billion in revenue, running almost entirely on agents. It’s a genuine demonstration of what the parallel agent economy makes possible. Also, it’s a case study in what happens when capability outruns accountability. FDA scrutiny over compounded GLP-1 sales. Deepfake before-and-after photos on the website. Doctors not in state licensing databases. The same architecture that can scale a legitimate business can scale exploitation just as efficiently.
The enterprise leaders I respect most aren’t moving fastest. They’re the ones who asked the harder questions first: what could go wrong, who is accountable when it does, and is our governance model actually keeping pace with our ambition?
AI capability is no longer the bottleneck. Governance is. The board question is the one nobody is asking out loud: at what point does fiduciary duty force you to deploy autonomous systems you fundamentally cannot control?
References: CNBC, “Pentagon AI Vendor Roster” (May 2026); Forrester, “Beware The Magical Two-Person, $1 Billion AI-Driven Startup” (April 2026).
