Tuesday. 8:57 a.m. Manolo has bolted a taxi meter to his computer.
—Are you driving someone to the airport? I ask.
—I'm showing customers how much work our AI does. Look: every token is recorded. They'll know exactly what they're paying for.
There's a six-page invoice on the desk. The first page shows the total. The other five explain how the machine arrived at it. Manolo runs his finger down a column of numbers like a man showing off the fuel economy of a new car.
His product compares suppliers' invoices with purchase orders for a restaurant chain. When it finds a difference, it prepares the case for someone to review before payment. Until now, he charged a monthly subscription. This month, he has added AI usage plus a markup.
—Total transparency, he says.
Isabel, the chain's finance manager, walks in holding an invoice. His invoice.
—I need you to explain this line, she says.
—Input tokens or output tokens?
—The one called “retries.”
Manolo turns the screen towards her. The system read a blurry delivery note, hesitated over the number of boxes, read it again, checked the order, and repeated the comparison four times. Every step has a cost and a timestamp.
—That's why we show it to you, he says. No black box.
—And how many boxes were there in the end?
—It doesn't know. It left the case pending review.
—So you're charging me for the four times it failed to find out.
—It used resources.
—A taxi uses fuel when it circles the block three times. I don't applaud the route.
Manolo looks at the meter. He brought it in as a metaphor and has just discovered that Isabel knows how to use one too.
They open the monthly report. Some invoices are matched to orders, some discrepancies are flagged, and some cases remain unresolved. In one of the latter, the supplier billed for four hundred boxes while the purchase order said forty. Someone should make a call before that bill is paid. The report includes the difference, buried among thousands of lines explaining what it cost to process.
—This is useful, Isabel says, pointing to the case. I need to know what I can pay, what I need to check, and what could go wrong if I make a mistake. I didn't come here to buy your attempts.
—But if I stop showing you the tokens, how do I justify the price?
—By explaining the work you do for me.
The conversation has a timely echo. Scott Woody, Metronome's founder, wrote on Stripe's blog that counting tokens helps an AI company control its costs but is usually a poor way to present a customer price. He argues for charging for a verifiable unit of work, such as an image generated or a case resolved. He sells billing tools, so this is a vendor's point of view rather than a universal law. Still, his question applies to Manolo's invoice: does it show the value delivered, or the cost of producing it?
—Fine. I'll charge per invoice reviewed, Manolo says. Problem solved.
Isabel goes back to the report.
—Does this one count as reviewed?
It's the invoice with the boxes. The system marked it “processed” and left the discrepancy open.
—The AI reviewed it.
—And I still have to review its review. If you charge me for clicking “process,” you've just renamed the tokens.
Manolo puts down his pen. The new unit is no better unless they agree on what it means to finish the job.
Together they separate three things: invoices that match their orders and pass the agreed checks; documented discrepancies, with the missing information and a named person to decide what happens next; and failed attempts, which Manolo needs to measure to improve his system but cannot present as completed work.
—And if your team gets a discrepancy wrong? Isabel asks.
—We'll have to fix it.
—Then put that in the proposal: what's checked, what's out of scope, and who pays to correct a mistake.
Manolo sets aside the six-page invoice. He keeps the usage logs to watch costs, spot loops, and see when the product needs help. The new proposal for Isabel has a much shorter table: accepted work, open cases, and a correction rule. They still need to agree on a price. At least, for the first time, they know what they're pricing.
—The meter cost me sixty euros, Manolo says, unscrewing it.
—Put it under internal costs, Isabel replies.
On the screen, the counter is still running. Manolo disconnects it before the bill grows again while they argue about the bill.
Thanks for reading.
We all know a Manolo. Sometimes we are Manolo. A new one every Tuesday.