What does Setting the Rule That Decides When AI Support Hands Off to a Human mean in practice?
To set the rule that decides when AI support hands off to a human, Dr. Jonah Tebaa defines two operational parameters: a turn threshold and intent scope. Rather than using a single global setting, organizations assign explicit thresholds, such as two unresolved turns, specifically to high-stakes intents like billing disputes, delivery exceptions, and account access. System triggers count customer repetitions, rephrasings, or dissatisfaction signals as unresolved turns, routing the conversation to human agents automatically.
A retail contact center's AI assistant fields a billing dispute. The customer says the charge is wrong. The AI apologizes, asks for the order number, apologizes again, asks for the invoice date, and by the third turn still has not resolved anything. No human has been looped in. The conversation is still open, still counted as "contained," and still going nowhere. This is the scenario Dr. Jonah Tebaa returns to when he talks about handoff rules — not whether AI support works, but what happens in the seconds before it stops working.
Tebaa has already made the measurement argument elsewhere: that deflection rate and containment rate reward keeping a customer away from a human rather than resolving what they came for. That case is settled ground for him now. His current focus is narrower and more operational — the specific rule that determines when an AI conversation escalates, and the two decisions that rule requires before it can be written down at all.
Two Parameters, Not One Slider
In Tebaa's framing, a handoff rule is not a vague commitment to "escalate when needed." It is a pair of explicit parameters that someone in the organization has to decide in advance, and defend.
The first is the turn threshold: how many unresolved back-and-forth exchanges the AI is allowed before it must hand the conversation to a person. The second is intent scope: which categories of request the rule applies to at all. Tebaa's position is that most teams either skip the second parameter entirely, applying one global threshold to every conversation, or never formalize the first, leaving it to whatever a vendor ships as a default.
Setting the Turn Threshold
Two unresolved turns, in Tebaa's account, is a defensible default rather than a magic number. The reasoning is a trade-off, and he is explicit that it cuts both ways. Set the threshold too high and the containment number looks good on a dashboard while the customer repeats the same request for the fourth time. Set it to one and the AI never gets the chance to reach the resolutions it is genuinely capable of on turn two, and the human queue absorbs volume that did not need to be there.
Tebaa describes this as weighing the cost of a wasted turn against the cost of a wasted escalation, and argues that weighing should not produce a single number for the whole business. A wasted turn on a store-hours question costs almost nothing. A wasted turn on a locked account costs a customer who cannot get into their own money. Those are not comparable costs, so in his view they should not share a threshold.
Scoping the Rule by Intent
This is the part of Tebaa's argument that most distinguishes it from generic advice to add a human fallback. Not every intent category needs a hard trigger, and treating all of them as if they do wastes the rule's authority. Some questions the AI resolves reliably enough that a forced handoff would only slow the customer down. Others carry enough downside — financial, legal, or reputational — that even a single unresolved turn is worth acting on immediately.
- Order status and store hours — no hard trigger needed; these are categories where the AI genuinely resolves the request, and a forced handoff here would degrade service rather than protect it.
- Billing disputes — earns a hard trigger; money and trust are both at stake, and an unresolved dispute risks becoming a chargeback the longer it sits with the AI.
- Delivery exceptions — earns a hard trigger; a late or missing package is a real-world event the AI cannot change, only report on, so extra turns rarely add value.
- Account access — earns a hard trigger; a locked-out customer cannot self-serve, and every additional turn compounds fraud-risk exposure rather than reducing it.
The dividing line, in Tebaa's telling, is not how frequently an intent occurs but what an unresolved turn costs when it happens. High-frequency, low-stakes intents can run without a trigger at all. Lower-frequency, high-stakes intents earn one even when they represent a small share of total volume.
Defining "Unresolved"
Tebaa argues this is the piece teams most often get wrong, precisely because it looks like a technical detail rather than a design decision. A turn counter is only as good as what it counts. In his framing, three signals should count against the threshold: the customer repeating the same intent, the customer rephrasing the same request in different words, and any explicit signal of dissatisfaction. All three indicate the conversation is not moving.
What should not count, in his account, is the AI legitimately gathering information it needs to resolve the request — asking for an order number, a shipping address, a preferred resolution. That is progress, even though it can look identical to stalling on a raw turn count. A rule that cannot tell the difference will either escalate too eagerly on ordinary information-gathering or fail to escalate the conversations that are actually stuck. Getting this distinction wrong, in Tebaa's view, undermines the rule more than getting the threshold number wrong.
A Rule Only Works If It Fires
The last piece of Tebaa's argument concerns enforcement rather than design. A handoff guideline that lives in a training document or gets mentioned to the support team once is not, in his terms, a rule — it is a hope. The trigger has to be built into the system itself, so that hitting the threshold on an in-scope intent routes the conversation to a human automatically, without depending on anyone noticing and acting on it in the moment.
He points to an illustrative pattern rather than a specific audited study: across mid-size regional retail and e-commerce operations he describes observing, repeat-contact rate on AI-touched conversations has moved from roughly 34% to roughly 14% after an intent-scoped, enforced handoff rule replaced an informal one. The figure is presented as a composite description of a pattern, not a client result, and Tebaa is careful to frame it that way.
His argument stops short of taking a position on AI support itself. He is not claiming AI should handle less, nor that human agents should handle more by default. The claim is narrower: the point at which a conversation moves from AI to human should be a designed, enforced rule with named parameters — not a fallback that happens whenever the AI runs out of things to say.
