Juan Piaggio · 2026-07-31 · 7 min read · ai · product-development · trust
Every demo of an AI agent shows it acting. It labels the ticket, links the duplicate, drafts the reply. What almost no demo shows is the more valuable behavior: the moment the agent looks at a case, decides it isn't sure, and does nothing. That restraint is not a gap in the product. It is the product.
An agent that always acts is trivial to build and impossible to trust. Give a model a task and it will produce an answer for every input, including the inputs where it has no business answering. Language models don't naturally say "I don't know" — they interpolate, confidently, right up to the edge of nonsense. Left unchecked, that means an agent will mislabel the ambiguous ticket with exactly the same certainty it labels the obvious one.
The teams that struggle with AI automation almost always have this failure mode. Their agent is not too weak; it is too eager. It acts on the 60%-confidence cases the same way it acts on the 95%-confidence ones, and every shaky action erodes a little more of the trust the strong actions earned. The fix is not a smarter model. It is teaching the agent when to hold back.
Abstention has to be engineered, because it will not emerge on its own. At Meshworq the mechanism is a confidence floor: a threshold below which the agent will not act on a case, even when the action type is one it's otherwise allowed to perform. Being permitted to add a label does not mean adding a label the model is only 40% sure about. Below the floor, the agent abstains and the case flows to a human instead.
This turns uncertainty into a routing decision rather than a gamble. Two independent gates sit in front of every agent action: an allowlist that says which action types are even permitted, and the confidence floor that says how sure the agent must be to use one. The allowlist bounds what the agent can do; the floor bounds when it does it. Together they make "do nothing" a first-class outcome the system reaches for deliberately, not an accident.
An agent that correctly declines to act on a case it doesn't understand is doing its job perfectly. Measure that as a win, not a miss.
Abstention looks like lost automation until you count what a wrong action actually costs. A confident automation that lands saves a few minutes. A confident automation that's wrong costs the reversal, the loss of trust, and — most expensively — the reviewer who now double-checks every future action because they got burned once. The cost of a bad action is not symmetric with the value of a good one. It's much larger, and most of it is paid in trust rather than time.
Once you see the asymmetry, the floor stops looking conservative and starts looking correct. You are trading a small number of automations you might have gotten right for the elimination of the automations you'd have gotten wrong — the exact ones that poison confidence. A narrower agent that people believe outperforms a broader one they've learned to second-guess, every time.
The risk with a quiet behavior is that it stays quiet. If abstention is invisible, it looks like the agent did nothing useful, and someone will eventually "fix" it by lowering the floor. So make it legible:
Surface those signals and abstention becomes something a team can reason about and tune, instead of a silent behavior nobody notices until it's gone.
The right floor is not a philosophical constant; it's an empirical one. Start it high — higher than feels comfortable — and let data pull it down. Watch the acceptance rate of the actions that clear the floor. If humans keep everything the agent does, the agent has earned room and you can lower the floor to let it handle more. If reversals climb, the floor is too permissive and the agent is acting on cases it shouldn't. The floor becomes a dial you turn against real outcomes, not an argument you have in a planning meeting.
The instinct to maximize what an AI does is exactly the instinct to resist. The agents people trust are the ones that know their limits and act only inside them. Engineer abstention as a deliberate behavior — a confidence floor, a clean handoff, a visible rate — and you convert the model's uncertainty from a liability into a routing signal. The best AI move is frequently no move at all, and a product that knows the difference is one your team will actually rely on.