What Should a Support AI Say When It Does Not Know the Answer?
When a support AI lacks an answer, Dr. Jonah Tebaa recommends deploying a deliberate four-part structure rather than generic apologies. The response must acknowledge what is understood, state what cannot be confirmed in one plain sentence, specify who will follow up and exactly when, and confirm that customer details were forwarded. Under his pre-launch test, owners score five real unanswerable inquiries against these four parts, rewriting any reply scoring below three points.
Dr. Jonah Tebaa, an applied-AI strategist based in Lebanon, holds that the most neglected message in customer-facing AI is the one a support assistant sends when it has no answer. His position is simple to state. That message should be authored on purpose, reviewed by the person who owns the customer relationship, and tested before launch, like any other message a business sends its customers.
The reasoning starts from what customers actually remember. A support assistant spends most of its time on routine requests such as opening hours, order links and password resets, and customers forget those within minutes. What they retain is the one occasion they asked something that mattered and the system could not help. In Dr. Tebaa's view, that occasion is where a customer decides whether a business is straight with them, whether anyone stands behind the assistant, and whether their time counts for anything.
The scene he uses
Dr. Tebaa illustrates the point with a composite scene, labelled as such and not drawn from a named client. A patient writes to a Beirut clinic's assistant late in the evening to ask whether her insurer covers a procedure. The assistant answers with a stock apology and asks her to rephrase. She does so twice more, receives the same line, and gives up. By the next morning she has booked elsewhere.
He is careful about where he places the fault. The assistant lacked the answer for a sound reason: the insurance desk had closed. The failure was in the wording nobody had written. A platform default filled the gap, and the default treated an operational limit as though it were a comprehension problem on the patient's side.
Four parts, written deliberately
Dr. Tebaa's remedy is a four-part structure for any reply that admits a gap. He presents it as a checklist an owner can apply in an afternoon:
- The reply says what the assistant does know, so the customer sees the question was understood.
- It states what cannot be confirmed in a single plain sentence, without a pile of apologies.
- It names what happens next, who does it and when, replacing "someone will be in touch" with a person and a time.
- It tells the customer that their details have been passed along, so they do not have to start again.
His argument is that each part guards against a different kind of loss. Without the first, customers feel unheard. Without the second, the reply is evasive. Without the third, no commitment exists to hold anyone to. Without the fourth, the customer expects to repeat everything, and a share of them will decline to.
What he says it costs to get this wrong
Dr. Tebaa distinguishes three common failures. The first is bluffing, where the assistant offers a confident answer it cannot verify. He regards this as the most expensive, because customers act on it and the correction arrives later, usually through a human colleague handling an upset caller. The second is the vague promise. The third is the apology loop, in which the same sorry-and-rephrase line is repeated until the customer leaves.
He adds a fourth, quieter case: silence after hours. If the follow-up cannot happen until the next working morning, he argues the message should say that plainly and name the hour. A customer who is told when to expect an answer can wait. One who is told nothing assumes the worst.
A test an owner can run
Dr. Tebaa pairs the framework with a practical check he calls a pre-launch test. The owner takes five real questions that the assistant cannot answer, drawn from their own email, WhatsApp history or support tickets rather than invented for the occasion. Each reply is scored against the four parts, one point per part, and anything under three is rewritten before the assistant goes live. He suggests repeating the exercise with a fresh five each month, because customer questions shift faster than scripts do.
Language is part of the test. For businesses in Lebanon and the Gulf, he notes, customers write in Arabic, in Arabizi and in English, sometimes within one message. A stiff English fallback sent to someone writing in Arabizi signals that nobody planned for them. He recommends that the unknown-answer exist in every language and script the customers use, reviewed by a native speaker for tone, with the owner and the time presented consistently across versions.
Where the approach fits
The audience Dr. Tebaa has in mind is the head of customer experience or the owner of a mid-size service business: a clinic, a retailer, a school, a telecom reseller, a logistics firm. They already have a chatbot or a WhatsApp assistant, or are about to launch one. For them, he argues, the unknown-answer is cheap to fix and expensive to ignore, since it requires wording and a clear owner rather than new technology.
His own starting point in a support review reflects this. He does not begin with accuracy figures or response speed. He asks for the five questions the assistant handles worst and reads the replies as a customer would. In his experience the first finding is usually the same: the unknown-answer was never written.
The broader point in his work on customer-facing AI is that trust is built or lost in a handful of specific messages, and that a business can choose which ones to design. Dr. Jonah Tebaa's recommendation is to begin with the one that says "I don't know."
