Which rule applied?
Policy changes. What was allowed last quarter isn't what's allowed now. A trace shows the steps the agent took, not which version of the rules was live when it took them, so nobody can say whether it was permitted at the time.
Getting to production
Why agent pilots stall
You've seen this meeting. The demo goes well, everyone's impressed, and then the agent sits there for five months. Here's what's usually going on.
The demo was never the hard part
Someone builds an agent that drafts the refund, or preps the claim, or pulls the filing together. It works. They show it around and people are excited about it.
Then it goes in front of the people who are accountable for what the company does, and everything stops.
The objection is almost never "it gets things wrong too often." It's quieter than that. Someone in risk or compliance asks a few questions, and the room goes silent.
It shows up in the surveys, too. In Gravitee's State of AI Agent Security 2026 report, only 14.4% of organizations said all of their AI agents went live with full security or IT approval. Everyone else has agents still waiting for it, or agents that went live without it.
The four questions
They're not gotchas. They're the questions that person is paid to ask.
Who handled the exception?
Something will fall outside policy, and a person will decide. If that decision happened in a chat thread, it isn't attached to anything. Six months later someone is scrolling a channel trying to work out who said yes.
Which agents does this cover?
The control wraps the agent your team wrote. It doesn't see the one someone in operations assembled in Copilot Studio last month, or the one clicking through a browser.
What actually happened?
Policy said proceed. Did it proceed, and for the amount policy approved? A record of the decision isn't a record of the outcome unless the application reports what it did.
Why teams fix the wrong thing
The natural response to a stalled pilot is to make the agent better. Tighter prompts, more evals, a stronger model, another layer of guardrails.
All worth doing. None of it moves this blocker, because the blocker was never accuracy.
Look at it from the other side of the table. If you're the person signing off, and you approve an autonomous system that can't answer those four questions, you've taken on an obligation you have no way to meet. You'd stall too. That isn't obstruction. It's doing the job properly.
No amount of eval coverage produces a record of who approved an exception. Those are different things, built by different parts of the system, and one of them is what the risk owner is asking for.
Where does your agent stand?
The go-live check asks eight questions about one agent, including who decides an exception, who changes a rule, and what an auditor would get, then shows which of those are already in place. You don't need to give us your email to see the result. Check your agent's go-live readiness.
What the conversation turns into
Once you can answer them, the argument stops being "should we deploy this at all" and becomes "how much should it decide on its own." A much better argument to be having, especially if you're the person who wants the agent doing more of the work.
Forrester's advice matches what works in practice: scale with bounded tasks behind approval gates and rollback paths, and widen autonomy only when the controls earn it.
In practice that looks like a small set of agents with clear decisions and an approval gate on anything outside policy, small enough that a person can review every exception. Then you watch what gets escalated. The volume and the pattern tell you whether your policy matches how the business runs. It usually doesn't the first time, and that's fine. As the controls hold up, you move decisions out of the escalation path and into the automatic one, and each move is a call someone made on evidence rather than a default nobody chose.
The pilot stops being a pilot the day you can write down what the agent is allowed to do and show that it stayed inside the lines.
Test the four questions on the agent that matters most
Start with the go-live check, then evaluate the agent with Developer, free for 60 days, and examine its recorded results.