AI pilots do not die at creation. They die at approval: provenance with no answer, PII with no record, access with no limit. The way out is to treat AI governance as infrastructure executed at runtime, with a structural access limit, masking by default, human approval for writes and a receipt per access.
In recent months, building an artificial intelligence agent stopped being a project and became a click. The big platforms put "create agent" on the toolbar, and any minimally technical team can assemble in an afternoon an assistant that reads spreadsheets, answers questions and produces reports. That part of the problem is solved, and whoever still sells only that is selling the past.
What is not solved is what happens after the demo.
Pilots do not die at creation. They die at approval.
In 2025, in the study The GenAI Divide, MIT published a figure that became a catchphrase: most generative AI pilots never generate real impact in production. The reflex is to blame the model, "we needed a better LLM". But anyone who follows these projects from the inside knows the model is rarely the defendant. Pilots do not die at creation. They die at approval.
They die when the director asks "where did this number come from?" and the answer is a shrug. They die when the DPO asks which personal data the agent saw and nobody has the record. They die when legal realizes that the sales assistant can, technically, reach the payroll. None of these deaths is technical. All of them are about governance.
Governance that executes, not governance that documents
Governance is not bureaucracy. It is what stops an automated decision from being made on wrong information. And there is one detail that changes everything in practice: that protection has to execute, not document.
Most companies treat AI governance as paper: a usage policy, a committee, an eighteen-page PDF no agent has ever read. Agents act in real time, along any path the data allows. Governance that is not applied at the moment of access is decorative, and the company is stuck between two bad extremes: locking AI down out of fear, or opening up and hoping.
The third path: governance as infrastructure
There is a third path, and it is less glamorous than it sounds: treating governance as infrastructure. In practice, it means four things. The access limit stops being a polite request and becomes structure: the sales agent does not see the payroll because it cannot, not because it promised. Personal data is born masked, and it is the exception that needs a justification. Every write to a system goes through human approval before executing. And every access, every single one, generates a receipt: who read, what, for which purpose, what was hidden.
When it works like this, the conversation with compliance changes in nature. It stops being "may we?" and becomes "it is on record". The audit that would take three days of digging comes out in seconds. And the pilot that was stuck on legal's desk becomes production, not because someone gave in, but because the hard question finally has an answer.
The bonus: whoever governs is whoever gets it right
There is a bonus few people notice: the same infrastructure that governs is the one that makes AI get it right. The context that tells the agent what it can see is the same one that says what "revenue" means in that company: the house definition, not the internet average.