How Should a Lebanese Business Sort the Voice Notes Its AI Receives?
To sort AI-received voice notes, Dr. Jonah Tebaa recommends classifying an inbox into a four-permission framework: Kind A clean orders for automatic AI action; Kind B conditional orders requiring human confirmation; Kind C questions answered strictly from approved facts; and Kind D personal or upset messages routed untouched to staff. In an afternoon exercise, businesses audit forty sample voice notes against a human answer key to set predetermined pass lines, defining tool permissions by note type rather than adopting an uncalibrated system.
Picture a food supplier near Beirut opening WhatsApp at twenty to eight. Most of the overnight messages are spoken, not typed. One says, over the drone of a generator, that the sender wants the usual, with the labneh doubled, and nothing on Thursday. Later that morning the supplier will sit through a software demo in which an assistant answers typed questions flawlessly. No one will play a voice note. Dr. Jonah Tebaa builds his argument on that mismatch.
The scene is an illustration, assembled from patterns rather than taken from one client. Its point is general: for many Lebanese and wider MENA businesses, the spoken message is the real order channel, and tools are routinely judged on something else.
Why the voice note became the order form
According to Dr. Tebaa, nobody designed this behaviour. Customers speak because speaking is quicker than typing Arabic on a small keyboard, and because it feels like a relationship rather than a transaction. A typed order resembles a form. A voice note says that sender and receiver already know each other.
That familiarity is exactly what makes spoken orders hard for software. In his work, he points to four recurring features:
- Assumed history. "The usual" is meaningful only to someone who remembers it.
- Buried conditions. The exception often arrives as an afterthought, in the same breath as the order itself.
- Language switching. Arabic, French and English can alternate within a single clause, frequently on the very words that matter.
- Poor audio. Traffic, kitchens, children and generators are the normal recording studio.
A customer who sends a short note believes the job is done. If the system catches the product but misses the condition, the customer will not blame the technology. They will blame the business.
One inbox, four different permissions
The common error, as Dr. Tebaa describes it, is to ask whether an AI "can handle voice notes" as though they were a single thing. He proposes a more useful question: which kinds of voice note does this inbox contain, and what should the AI be allowed to do with each one? He sorts them into four.
- Kind A, the clean order. A product, a quantity, a delivery detail, nothing odd. Here the AI may act: it records the order and reads it back for confirmation.
- Kind B, the order with a condition. The request is genuine but bent by an exception, such as a day to avoid or a substitution. The AI drafts the order and flags what it heard, and a person confirms before anything is committed.
- Kind C, the question or complaint. Someone asks where a delivery is, or reports a problem. The AI may answer only from facts the business has approved, such as opening hours or delivery windows, and otherwise passes the note to a person with a short summary.
- Kind D, not for a machine. A personal message, an angry customer, a money dispute, a negotiation. The AI does not respond at all. The note goes, untouched, to someone with the standing to reply.
The effect of this framework is to replace a vague argument with a precise one. Instead of debating whether the AI is good enough, the owner asks whether it is good enough for a given kind. A system that excels at Kind A and stumbles on Kind B remains valuable, provided it is never allowed to treat B as though it were A. The danger lies in the permission, not in the tool.
Finding your own mix by sorting
Dr. Tebaa is careful not to hand out a universal ratio. A bakery, a clinic and a parts distributor will all see different proportions, and he declines to guess at any particular company's. Discovering the mix is the purpose of a simple exercise that can be done in an afternoon by one patient member of staff.
First, gather around forty recent voice notes, with the sender's agreement or as re-recordings of real patterns in staff members' own words, and include the noisy ones. Second, have a person write a one-line meaning for each, conditions included; that becomes the answer key. Third, label every note A, B, C or D and count them. Fourth, and before looking at any results, write down what counts as a pass, for example no missed condition on a Kind B note and every Kind D note routed without a reply. Fifth, run the AI on the same untrimmed audio. Finally, score it separately for each kind.
Forty notes will not produce statistics, and the exercise does not pretend to. What it produces is a picture: where the system is dependable, where it is guessing, and which kinds it mistakes for each other. Fixing the pass line in advance matters because a standard chosen after seeing the scores tends to drift toward whatever the scores turned out to be.
Turning the result into decisions
Suppose the sorted notes lean toward clean orders, with a fair share of conditional ones and a handful of questions and sensitive messages. Suppose also that the AI performs well on the clean orders but loses the exception in several conditional ones, and replies to a note that should have gone to a human. The sensible response, in Dr. Tebaa's view, is neither to buy nor to reject. It is to grant permissions kind by kind: automatic action on A, human confirmation on every B for an initial period, approved facts only on C, and silence on D until the failure is understood and retested.
The same result also sharpens the conversation with a supplier of the technology. Owners can ask, in writing, whether audio is processed directly or first converted to text and what is lost in that step, how the system copes with noise and mixed languages, and what it does when it is unsure. The last question reveals the most. A system that guesses silently is a different proposition from one that raises a flag a person can see.
Where to start
His practical advice is modest. Before any vendor meeting, ask whoever manages the WhatsApp line to save forty recent voice notes and write a line on what each one means. Sort them. Only then decide what the machine may touch. In the opening scene, the supplier would replay the note and hear the words about Thursday, and the question for the meeting would change from whether the assistant is clever to which messages it is permitted to handle.
The full argument, including the sorting steps in his own words, appears in the original article: The Order Arrived as a 23-Second Voice Note.
