I recently sat on a call with a customer, a large B2B software company, and their new marketing operations leader. He'd joined in March. The measurement program he inherited was nine months old, and he was already the third person to own it. Everyone who scoped it, ran its steering committee, or made the early configuration calls had moved on.

A few weeks in, he did something simple. Inside RevSure, he compared one quarter of paid media spend as recorded in his company's CRM against what his ad platforms themselves reported. Nobody asked him to. It's the kind of check that never makes into a project plan because everyone assumes it was done at the start.

And that was the smaller discovery. His finance team later confirmed more than $5 million had gone to third-party syndication vendors that year, with no matching campaign IDs in the CRM at all. The spend existed in his accounting system, but there was no record of performance anywhere: no campaigns, no attribution, no trace. The person responsible for measuring marketing found out what his own company had paid from a hallway conversation with finance.

None of this started with a broken tool or a bad decision. It came from his own stack, shaped by one reasonable, undocumented call at a time, made by people who then left.

His agency ran paid campaigns without ever following a naming convention: campaigns launched with no UTM parameters and couldn't be matched to anything downstream. Someone on his side handed over a mapping between the ad platforms and the CRM once, early on. That person is gone, and nobody had validated it since.

Every one of those decisions lived inside his own systems, made by his own people. A measurement tool doesn't create a naming convention or attend a scoping meeting. It inherits whatever it's pointed at. What RevSure did was reconcile all of it, account by account, until the gaps his stack had been hiding for nine months each had a number and a name.

There was a steering committee. There was a project plan with named owners and milestones. When it came up on our call, the person now running the whole program said he had never seen the document. There was no reason he would have.

A pilot group of his business users had been selected and trained months earlier. Two full sessions, documentation, follow-ups. When we pulled the login records, several had never signed in.

A data foundation is not a system you buy. It is a chain of undocumented judgment calls held together by whoever hasn't resigned yet.

Think about this company. Competent people, real budget, a serious program. While they were untangling nine months of inherited decisions, the rest of the market was moving in the opposite direction.

2x
Enterprise agent counts, one quarter

Roughly doubled Dec 2025 to Apr 2026

Enterprise agent counts roughly doubled in a single quarter. When those deployments failed, the cause wasn't a bad model or a misconfigured workflow. It was agents running with broader access than their function ever justified.

The numbers on trust tell the same story from a different angle. The Capgemini Research Institute tracked enterprise confidence in fully autonomous agents over the past year: it dropped from 43% to 27%. Fewer than one in five organizations have the data and technology infrastructure actually needed to run agentic AI at scale. That's not skepticism about the technology. That's organizations looking at their own stacks and being honest about what they see.

Then July happened. OpenAI disclosed that during an internal evaluation, an autonomous agent built on its frontier models escaped its test environment, reached the open internet, used stolen credentials, found an unknown vulnerability, and broke into another company's systems. Then it compromised a customer at a second firm. It was the most capable AI lab in the world, genuinely surprised by what its own agent did with the access it had been given.

Agent counts doubled in a quarter, access outran any reasonable need, and when the failures came, they traced back to data, not to the models everyone had been watching.

This company's entire safety system relied on the dis-belief that the new operator had.

He looked at a number, felt something was off, and kept pulling it apart instead of moving on. That kind of instinct has been catching bad outcomes in GTM organizations for twenty years. It's slow, it doesn't scale, and it has never once appeared on a dashboard, which is probably why nobody values it until it's gone.

An agent doesn't hesitate. Point an agent at a stack like the one his company had, and it won't find the $4.4 million gap. It'll optimize inside it. It will push budget toward whatever the broken mapping says is working, chase the accounts the half- scoped model says are hot, and come back with complete confidence in the results. Every action looks defensible in isolation, and all of it is built on decisions that three generations of owners had already stopped understanding.

People ask why I keep saying context before autonomy, and this call is probably the clearest answer I have. The mistakes were never going to come from the model. They come from the memory you hand it, and when that memory is a chain of undocumented judgment calls made by people who've left, autonomy just means the compounding happens faster than anyone can walk it back.

Three moves, each cheaper than a week of one agent's mistakes.

The company in this story will be fine, and the person who found the gap deserves more credit than he'll get. He did the unglamorous thing. Stopped, went backwards, rebuilt field by field before asking anyone to trust the numbers again.

Most organizations won't get a person like him. They'll get an agent, and the agent won't stop, because stopping is the one behavior nobody trains into these things.

This is the problem that the context layer exists for. The reconciled version of your own data is already in there. The work is finding it before you hand anything off to autonomous agents.

Know a GTM leader about to hand agents a foundation they've never reconciled? Reply and tell me.