What does Dr. Jonah Tebaa recommend a customer AI has in writing before a sale?
Before a sale, Dr. Jonah Tebaa advises firms to provide customer-facing AI with an Offer Sheet containing five lines: In, Out, Combine, Ends, and After. It is the single operational source for inclusions, exclusions, stacking rules, expiration times, delivery cut-offs, and return policies. To validate readiness, teams execute the Twenty-Question Rehearsal across five inquiry buckets. In a hypothetical homeware shop, pre-launch testing cut the expected wrong or unsupported answers from 220 to 30 out of 600 terms questions, and the one remaining miss was a gap in the sheet itself.
One page of five lines
Dr. Jonah Tebaa's advice for any firm about to run a sale starts with a single sheet of paper. He calls it the Offer Sheet. It has five lines: In, Out, Combine, Ends and After. In names exactly what the offer covers. Out lists what is excluded, one item at a time, including whatever customers tend to assume is included. Combine says yes or no for each code, loyalty scheme or bundle price. Ends gives the date and time in the customer's own clock, and what happens if stock runs out first. After gives the delivery cut-off and the return terms for items bought on the offer, where those differ from normal policy.
He presents the sheet as a working method, not as research. It is meant to be written by whoever owns the offer, in the words a customer would use, and then treated by every person and every automated assistant as the single source for that sale.
Why a sale needs its own source
Dr. Tebaa starts from an observation about where information sits. The terms of an offer are usually found in the banner, the launch email and the marketing brief. A customer-facing AI assistant, and any seasonal staff, are usually set up on standing policy. Normal returns and normal delivery times are in their briefing. The sale's exclusions and its stacking rules often are not. He is careful to say this is not a failing of the marketing team: a banner is made to attract, and it does. The gap is only that nobody has put the fine print in one place that service can read.
His point is that the gap matters because a sale changes the mix of what customers ask. In his hypothetical homeware shop, a three-day sale lifts total chats 2.5 times, from 600 to 1,500, while questions about the offer's terms rise from 30 to 600, which is 20 times. Everything else a customer might ask rises from 570 to 900. All of these figures are invented for the example and he does not offer them as statistics about any market. The lesson he draws is that a team planned around total volume prepares for the smaller number.
The rehearsal, and why he weights it
Before the banner goes up, Dr. Tebaa asks for what he calls the Twenty-Question Rehearsal. The terms questions fall into five buckets: does it apply to this item, can it be combined, when does it end, will it arrive in time, and can it be returned. Four questions go into each bucket, 20 in all, and each is answered from the sheet alone, by the assistant and by a colleague who was not at the briefing.
The step he stresses most is the weighting. In the example, 7 of the 20 rehearsal answers fail when no sheet exists, which is 35 percent. Taken flat, that suggests 210 failures across 600 real terms questions. But the buckets differ in size, and the two that fail most carry a larger share of real questions than of rehearsal questions: about 47 percent of the 600, against 40 percent of the 20. Weighting each bucket by how many customers would ask gives 220 expected failures, about 36.7 percent. The difference is small in this example, and he suggests it grows where the buckets are less even.
What the retest found
After the sheet was written and the same questions rerun, one answer in 20 was still wrong, in the Combine bucket. The sheet said discount codes do not combine and was silent on loyalty points. Dr. Tebaa uses this detail deliberately. The remaining failure was in the document, not in the assistant, so he tells owners to fix the sheet first, rerun the questions, and only then look at the assistant itself. Weighted, the expected failures fall from 220 to 30, and after the loyalty-points line was added the Combine questions all passed.
He also gives a rough staff estimate, labelled hypothetical. If one wrong answer in five becomes a ten-minute dispute, 220 failures mean 44 disputes and 7 hours 20 minutes of staff time, against 6 disputes and 1 hour with the first sheet. He insists the point is the method, forecast by type, weight by volume, rehearse before launch, and that a 20-question sample is a planning estimate and not a measurement.
Retiring the sheet
The last part of his advice is about housekeeping. The sheet carries an end date and time, and when the offer ends the sheet is retired, so that a finished offer is not quoted to customers by accident. For firms selling into several markets, he writes that end time in each customer's own clock, the same discipline as his calendar grid for Lebanese teams serving the Gulf and the diaspora. If the offer changes in the middle, such as a late extension or a new exclusion, the rehearsal is run again. He also recommends that owners check the rules for advertised prices and terms in their own market, which is a separate matter from the sheet.
Dr. Tebaa sets out the full worked example, the table of buckets and the pre-launch checklist in his own post on jonahtebaa.com: A Sale Lifts Chats 2.5 Times and Terms Questions 20 Times.