Deny by Default: Deciding What your Automation Is Allowed to Think

Right now the default answer to most automation problems is: make it an agent. The Tutero team nearly did that with churn — and are glad they stopped.

Churn at Tutero is mission critical. It touches revenue directly. The original plan was an agent to handle it end-to-end. Then they looked at what that agent would actually be deciding. Customers already answer a long-form questionnaire, so every detail needed is already there, structured, before anything runs. There was nothing left to interpret — the rule was writable as a set of conditions. So they built it as a deterministic workflow instead. No model in the path.

That's the real question this talk is about: not "can a model do this," but "is there anything left to decide?" Support and recruitment are different — judgment is real there, so those are agents. The second half covers containing them: least-privilege tool access across nine internal systems, and a deny-by-default rule for what agents can hand to each other.

Attendees will leave with the test used to make the call, and the patterns used to stop an agent reaching further than intended.

• Learn a test to run on any automation before building it: is there anything actually left to decide?
• See a worked example from a revenue-critical system where the obvious answer was an agent, and the right answer wasn't
• Understand the two cases where agents were the right call, and what made them genuinely different from churn
• Discover the tool-access and delegation patterns that stop an agent from reaching further than intended