"Probably a lot of people in this room have had this experience of taking your prompt, rewording things, putting things in all caps at the end of the instructions, adding some exclamation points. Maybe after the third exclamation point, you start to feel like this isn't the right control surface for hard control."
Our CTO
@profdanklein has been on a lot of stages this year, most recently at
@stepconference and the question he keeps getting asked is why agentic AI stalls between the demo and production.
His answer starts with the control surface. A prompt turns every instruction into a suggestion. "Be reassuring" and "loan quotes are only authorized through this API" get exactly the same treatment, and capitalization doesn't change that.
Plenty of CX agents take actions today. The question is what holds them inside policy while they do it. If the answer is a prompt, the rule is soft. If the answer is a second model checking the first, the rule is harder, and you're paying for it in latency, cost, and complexity. A disputed charge or a rebooking mid-cancellation needs a rule the model can't override in the first place.
Our model, APT, puts those constraints inside the model itself and verifies them there. The hard requests get handled with guarantees.