Two unresolved turns. That is the default Dr. Jonah Tebaa returns to most often when he is asked how long a customer-facing AI should be left to keep trying before a human takes over. Not zero, which throws away the resolutions the system would have reached on its own. Not five, which lets a customer repeat the same complaint until someone notices the dashboard looks wrong. Two. But the number is the least interesting part of his argument. What matters is that it is a number at all — written down, enforced by something outside the conversation, and not left to the AI's own sense of whether it has succeeded.
Dr. Jonah Tebaa's position starts from a problem he has written about separately: containment and deflection rates measure whether a human was avoided, not whether the customer's problem was solved. An AI can post excellent numbers on that metric while a customer sits through four increasingly circular exchanges and never gets an answer. If the measurement itself cannot be trusted to flag failure, something else has to. That something, in his framework, is a handoff rule, and he argues it only qualifies as a rule if it has two specific, deliberately chosen parameters: a turn threshold and an intent scope.
Why a Guideline Is Not the Same Thing as a Rule
Most operations Dr. Jonah Tebaa has examined already have a version of a handoff policy. It typically reads something like: "if the AI cannot resolve the issue, escalate to a human." He treats this kind of language as advisory rather than operational, and the distinction is central to his argument. Advisory language of that kind quietly assigns the detection job to the model itself, live, with nothing else watching. A system that is confidently wrong has no mechanism for flagging itself as confidently wrong — that is precisely what being confidently wrong means. Asking software to certify the boundary of its own competence yields an intention, never a control.
What makes something a rule, in his usage, is that it does not depend on the conversation to police itself. A separate counter tracks turns regardless of what the model claims to believe, and crossing the threshold triggers a handoff without asking the AI's permission first. That is, in his view, the entire difference between a policy that sits in a document and a control that changes what actually happens to a customer at 2am on the fourth unresolved exchange.
The First Parameter: How Many Turns Is Too Many
Dr. Jonah Tebaa frames the turn threshold as a trade-off between two costs rather than a fixed setting to be copied from a vendor default. A generous threshold buys a healthier containment figure with the customer's patience: nothing registers on the dashboard while someone restates the same request for a third or fourth time. Stopping at a single turn errs the other way, discarding the cases the AI would have closed on one more clarifying answer and pushing conversations that never needed a person into the human queue, which slows the ones that genuinely did.
The cost of a wasted turn, measured in patience and a marginally worse repeat-contact figure, is not the same size as the cost of a wasted escalation, which consumes a human agent's time and queue capacity that a genuinely urgent case might have needed instead. Because those two costs scale differently depending on what is actually being asked, Dr. Jonah Tebaa argues the threshold should never be applied as a single number across every conversation type. It should vary by category, which leads directly to the second parameter.
The Second Parameter: Where the Rule Should Apply at All
In Dr. Jonah Tebaa's account, this is the parameter most operations skip, and the more consequential of the two. Nobody is harmed by a sluggish exchange about store hours or the whereabouts of an order, since the system settles those correctly on its own, and forcing a person into them just burns staff time. His test is simpler: does an extra unresolved turn make the eventual fix harder, costlier, or riskier to leave alone? Four categories fail it.
- Fraud exposure grows with every minute a customer stays locked out, which is why account access failures cannot be allowed to run past the second unresolved turn.
- A shipment that is already late, lost, or damaged cannot be un-broken by more conversation, so each extra turn on a delivery exception just delays a fix only a person can actually authorize.
- Contractual deadlines do not wait for a chatbot to catch up, so a stalled reply to a cancellation inside that window can turn a routine request into an avoidable dispute.
- Money and trust both ride on a billing dispute, so leaving one unresolved risks a chargeback, an outcome that typically costs several times more than simply routing the conversation to a person sooner.
Categories that survive the four-part test need no forced handoff attached to them; a looser threshold, or none, does the job. The narrower scope exists to keep the human queue open for moments where a delay is genuinely expensive, not for questions the AI was always going to answer correctly.
Defining "Unresolved" Before the Rule Can Work
None of the above functions, in Dr. Jonah Tebaa's view, unless a team has explicitly defined what counts as an unresolved turn. A counter with no working definition of failure will either fire on almost everything or almost nothing. His own working definition fires on three signals: the customer restates one underlying request in fresh words, they voice explicit dissatisfaction, or the preceding AI reply moved the case nowhere. One exclusion matters more than any of them. A turn spent collecting something the model actually requires is not a failed turn. Asking for an order number advances the case rather than stalling it, and scoring that as failure produces a trigger that fires nonstop on exchanges which were proceeding perfectly well. Missing that distinction, in his experience, is the most frequent design mistake, producing a rule that either stays silent through real failures or interrupts conversations that were never in trouble.
Among the mid-size retail and e-commerce operations Dr. Jonah Tebaa has observed around the region (composite patterns rather than a single audited case), enforcing the two-turn threshold across a narrow band of high-risk categories has been associated with repeat-contact rates on AI-touched conversations dropping from roughly a third of cases to closer to one in seven. He is careful to frame that figure as a pattern rather than a guarantee; the number moves or does not depending on the operation. The two-parameter structure underneath it, a defensible turn threshold paired with a deliberately narrow intent scope, both enforced outside the model's own judgment, is, in his account, the part actually worth building.
