Back to Blog

The Invoice Schedule Decides Who Finishes the Last Mile

In a hypothetical USD 24,000 AI build, 90% is paid before real work and finishing costs more than the last invoice. Five lines fix the schedule.

A navy wool jacket on a cream dress form, one sleeve only tacked on with short rust stitches round an open armhole, a needle on rust thread at the raw cuff. Text: Pay for proof, not delivery.
Direct answer

What payment schedule makes an AI supplier want to finish the last mile?

One that keeps real money behind a passed run on real work. In my hypothetical USD 24,000 build, a 50 / 40 / 10 schedule pays 90 percent, 21,600, before the tool touches real work, so the last 2,400 is worth less than 4,800 of finishing. A 30 / 30 / 40 schedule, with the last 40 percent tied to a passed run on 50 items the buyer draws, leaves 9,600 unpaid and finishing 4,800 ahead.

A contractor has been paid USD 21,600 of a USD 24,000 AI build. Finishing it properly, on real quotes and real customers, will take him 60 hours, which cost him USD 4,800. The invoice still unpaid is USD 2,400. Every figure in this post is hypothetical, built to show a method, not taken from a client or from any market.

The buyer in my example is an invented 40-person building-materials distributor in Beirut. It has hired an outside implementer to build AI-assisted quoting and order entry. The demo went well and the tool has been delivered. The question is what happens next, and my answer is that the payment schedule has already decided it.

What the remaining invoice is worth

Call this Schedule A, the familiar-looking one in my example: 50 percent on signing, 40 percent when the build is delivered, 10 percent on handover and training. On a USD 24,000 fee that is 0.5 x 24,000 = 12,000, then 0.4 x 24,000 = 9,600, then 0.1 x 24,000 = 2,400. Check: 12,000 + 9,600 + 2,400 = 24,000.

Now look at the moment the tool is delivered. The supplier has been paid 12,000 + 9,600 = 21,600, which is 90 percent of the fee, and the tool has not yet run on a single real piece of work. Still unpaid: 2,400.

Finishing properly means 60 hours at a cost of USD 80 an hour (hypothetical): 60 x 80 = 4,800. Set that against what remains to be invoiced: 2,400 - 4,800 = -2,400. By finishing, the supplier loses USD 2,400.

I want to be fair to suppliers here. Nobody in this story is dishonest. A good supplier may finish anyway, for reputation or out of pride in the work. But a schedule that pays more for walking away than for finishing is asking for the wrong outcome, and it asks it of every supplier, including the good ones. When the last mile gets cut, I look at the invoice before I look at the character.

Why the last mile is the expensive mile

The demo used clean examples chosen by someone who knew they would work. The real work arrives as a photographed price list, a quote request typed with one thumb, a customer who names a product by a nickname, a line item that exists in two units. Each odd case is small. Together they are most of the 60 hours.

There is also a human part. Staff who have done quoting for years will check the tool's output for a while, and the supplier has to tune around what they correct. Little of this appears in a proposal, because it is hard to price before the real items show up. That makes it the part of the work most likely to be squeezed when the money has already been paid.

The Proof Tranche, five lines

The Proof Tranche is my working method, not research. It moves the last slice of the fee from "delivered" to "proven on real work", and it is written on one page before anyone signs.

  1. Sample. The buyer draws 50 real items from the last two weeks of normal work, items the supplier has never seen. The answer key stays sealed until the run.
  2. Bar. The pass mark in numbers, agreed before signing: at least 44 of 50 items accepted with no staff rework, and average staff correction time at or under 3 minutes an item.
  3. Window. The run happens in the 5 working days after go-live. The buyer must supply the sample within 3 working days of being asked. If the buyer is late, the tranche falls due by default.
  4. Rounds. At most two fix rounds. Every re-run uses a fresh 50 items, because re-running the same 50 trains the supplier to the test, not to the work.
  5. Release. On a pass, the tranche is paid within 7 days. On a second miss, a named person on each side decides, and the referee is agreed before any money is at stake.

Why 50 items? It is large enough that a result near 78 percent looks different from one near 90 percent, and small enough for one person to review in an afternoon. It is a planning size, not a statistical claim. Ask a lawyer in your market to draft the wording; this is a planning method, not legal advice.

The same contract, rewritten

Schedule B, the Proof Tranche version: 30 percent on signing, 30 percent when the build is delivered, 40 percent on the passed acceptance run. That is 0.3 x 24,000 = 7,200, then 7,200, then 0.4 x 24,000 = 9,600. Check: 7,200 + 7,200 + 9,600 = 24,000. Before the run, the supplier has been paid 7,200 + 7,200 = 14,400, which is 60 percent.

Hypothetical, USDSchedule A (50 / 40 / 10)Schedule B (30 / 30 / 40)
Payments12,000 / 9,600 / 2,4007,200 / 7,200 / 9,600
Paid before any real work21,600 (90%)14,400 (60%)
Still unpaid2,4009,600
Cost to finish (60 h x 80)4,8004,800
Unpaid minus cost to finish-2,400+4,800

Under B, 9,600 - 4,800 = 4,800. The supplier now earns USD 4,800 above cost by finishing, and loses the 9,600 by leaving. The incentive has flipped without anyone becoming more honest or more demanding.

Here is how the run might go. The distributor's operations lead draws 50 real quote requests. On run 1, 39 of 50 are accepted without rework: 39 / 50 = 78 percent, below the bar of 44 / 50 = 88 percent. The tranche is withheld. The supplier spends fix round 1 on the misses, and run 2 uses a fresh 50. Now 45 of 50 are accepted, 45 / 50 = 90 percent, and staff correction time totals 130 minutes over 50 items: 130 / 50 = 2.6 minutes an item, under the limit of 3. It passes, and USD 9,600 is released.

The fresh sample matters. Had run 2 reused the first 50, the supplier could have fixed those exact items and passed on memory. A new 50 tests whether the tool handles the work, and that is the only thing the buyer is paying for.

What this does not fix

  • A vague brief. If nobody can say what a correct quote looks like, no bar will be fair. The five lines assume the work has been defined.
  • A thin early payment. A supplier needs enough on signing to start. If 30 percent does not cover real setup costs, the schedule hurts a small supplier, and you should adjust it rather than squeeze it.
  • Rare cases. Fifty items from two ordinary weeks may miss month-end or a seasonal product. A pass is evidence for normal work, not a guarantee for every case.
  • A biased sample. If the buyer picks only its ugliest items, the run is a trap, not a test. Draw from normal work and keep the answer key sealed.
  • The contract itself. The numbers are invented, fees and rates vary, and the wording needs a lawyer in your market.

What to do before you sign

Open the quote you are about to approve. Mark how much of the fee is paid before the tool has run on real work. If it is 90 percent, ask for 60 or 70, and move the difference behind a run. Write the five lines on one page: sample, bar, window, rounds, release. Name the referee on each side, and make sure the supplier has read the clauses that protect them: the default release if your sample is late, the cap of two rounds, the fresh sample.

The same logic holds for a build done in-house. Release the last slice of the project budget, or the team's project bonus, after the run passes rather than at go-live.

If you want a second pair of eyes on a quote or a contract schedule, write to me.

Frequently asked questions

What if my supplier refuses to hold back 40 percent?

Offer the clauses that protect them: the tranche releases by default if my sample is late, there are at most two fix rounds, and a referee is named in advance. If they still refuse, I would move to a smaller final tranche, say 25 or 30 percent, with a written bar. A supplier who will not agree to any bar is telling me something.

Who should draw the 50 items?

Someone on my side who does that work every day, not the project sponsor and not the supplier. They pick from the last two weeks of normal work, not from the hardest cases. The answer key stays sealed until after the run, so nobody can tune the tool to the answers beforehand.

Does this work for a subscription tool with no build fee?

Partly. There is no build tranche to move, so I ask for a paid pilot month run with the same five lines before the annual term begins, or for a written right to exit if the bar is missed after the trial. The principle is the same: the long commitment follows proof on real items.

How do I use this for an in-house build?

Hold back the last slice of the budget, or the team's project bonus, until the acceptance run passes. Use the same five lines and the same fresh sample for any re-run. The incentive problem is the same inside a firm: if everyone is rewarded at go-live, nobody is paid to finish the last mile.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.