What does Dr. Jonah Tebaa say should be tied to a passed run on real work in an AI build?
Dr. Jonah Tebaa ties the last 40 percent of an AI build fee, the Proof Tranche, to a passed acceptance run on 50 real items the buyer draws. In his hypothetical USD 24,000 example, the bar is at least 44 of 50 items accepted without rework, which is 88 percent, and correction time of 3 minutes an item or less. Every re-run uses a fresh 50 items.
Dr. Jonah Tebaa starts from a small piece of arithmetic that is easy to skip when buying an AI build. Compare what a supplier is still owed at the moment a tool goes live with what it will cost that supplier to make the tool work on real business. If the second number is larger, he says, finishing is a loss, and a sensible supplier will drift toward the exit. All figures in his example are hypothetical and illustrative.
The arithmetic of a back-loaded schedule
His invented case is a 40-person building-materials distributor in Beirut that commissions AI-assisted quoting and order entry for a fee of USD 24,000. In a familiar-looking payment pattern, which he calls Schedule A, the supplier invoices 50 percent on signing (12,000), 40 percent on delivery of the build (9,600) and 10 percent on handover and training (2,400). By the time the tool exists, 21,600 has been paid, which is 90 percent of the fee, and it has not yet been run on real work.
Dr. Tebaa then prices the remaining work. Making the tool cope with real quote requests, odd formats and cautious staff takes, in his example, 60 hours at a cost of USD 80 an hour, or 4,800. The unpaid 2,400 less that cost leaves a gap of minus 2,400. He is careful to say this is not an accusation. In his account the schedule, not the supplier, creates the incentive, and an honest firm will feel the same pull as any other.
Paying the last tranche for proof
His remedy is to change what the final payment is for. Under Schedule B, the supplier receives 30 percent on signing (7,200), 30 percent on delivery (7,200) and 40 percent (9,600) only when the tool passes an acceptance run on real items. The supplier has then been paid 14,400, or 60 percent, before the run. With 9,600 outstanding and 4,800 of work to do, finishing leaves the supplier 4,800 better off than leaving. The tool, the people and the fee are unchanged. Only the trigger for the last invoice has moved.
He calls the instrument the Proof Tranche and reduces it to five lines agreed before signing: the sample, the bar, the window, the rounds and the release. The buyer draws 50 real items from the previous two weeks of ordinary work. The bar in his example is at least 44 accepted with no staff rework, which is 88 percent, and average staff correction time of no more than 3 minutes an item. The run takes place in the five working days after go-live, with up to two fix rounds, and payment follows within seven days of a pass.
Why every re-run needs a fresh sample
The detail Dr. Tebaa returns to most is the sample. If a failed run is repeated on the same 50 items, the supplier can correct those particular cases and pass, having learned the test rather than the work. He therefore requires a new 50 for each attempt, with the answer key sealed until the run is complete.
In his example the first run accepts 39 of 50, or 78 percent, below the 88 percent bar, so the tranche is withheld. After one fix round a fresh 50 gives 45 of 50, or 90 percent, and staff spend 130 minutes correcting them, which is 2.6 minutes an item. Both thresholds are met and USD 9,600 is released. He picked 50 items because that is enough to tell a result near 78 percent from one near 90, and still small enough for one person to review in an afternoon. He describes this as a planning size, not a statistical finding.
Protections for the supplier
Dr. Tebaa insists the method must be fair to both sides, or suppliers will refuse it. If the buyer fails to supply the sample within three working days of being asked, the tranche falls due by default. Fix rounds are capped at two, so the work cannot be reopened indefinitely. The bar is written down and known in advance, and a referee from each side is named before any money is at risk. He adds that the wording belongs with a lawyer in the buyer's market, since the method is for planning and is not legal advice.
What the method leaves open
He is open about the limits. A pass on 50 items from two ordinary weeks says little about a month-end rush or a seasonal range. A vague brief cannot be rescued by a fair bar. A very small supplier may need more than 30 percent on signing to cover real setup costs, in which case the percentages should move. The same reasoning applies to internal projects, where he suggests holding back the last slice of the budget or the team's project bonus until the run passes.
For the full first-person argument, including the side-by-side table and the action he suggests before signing, read his article on jonahtebaa.com: The Invoice Schedule Decides Who Finishes the Last Mile.