Where the last six weeks went, and what I decided to build instead.
This is the first issue in six weeks, longer than this newsletter has ever gone quiet. Enough of you noticed to write and ask, and the honest answer is that the time went into two books. One of them, The Human Check, is out next month. The other is further off. Between them they took every writing hour I had.
The quiet did something a fortnightly rhythm never does. It gave me room to ask what this newsletter is for. Looking back over twelve issues, the ones I would stand behind in a board meeting are the ones with a number in them. The Operating Tax. The Inversion Point. The Headcount Chain. Each gave you a sentence to say to a CFO and a reason the CFO would listen.
So while the books were being written, the research did not stop. For most of this year I have been working on the economics of AI in production: what agents cost once they are running, which numbers a business case has to carry, and the tools that let a finance team see those numbers before the budget review finds them. That work is now big enough to need a shape of its own, and I decided the shape should be a case note rather than more issues like the last twelve.
A case note takes one company that has put its AI economics on the public record and reads the record properly. Not the interview. Not the article. The filings, the investor presentations, the letters to the regulator, dated and laid side by side. No access to the company, no inside information, and no estimates of anything the company has not disclosed. Where the record stops, the note says so.
The reading is the work, and it is slow. The first note took weeks, and most of that was not writing. It was deriving one quarter’s cost from a six-month figure because the quarter was never filed on its own. It was finding the letter the company wrote to the SEC that nobody who tells the story has read. It was chasing a figure half the commentary repeats, 81 per cent of chats handled by AI, back to the investor presentation it is attributed to, not finding it there or in the call transcript, and leaving it out. A case note is as much about what it refuses to use as what it uses.
Each note has the same five parts. The record: every figure the company has put on the record, where it said it, and what it says now. How the story got told: the company’s statements against the coverage, in order, because the gap between those two is usually where the lesson is. My reading, marked as mine. A one-page tool you can fill in for your own program. And every source, linked, with a list of what the note does not know.
You know the story, because everyone tells it the same way. The AI assistant that did the work of seven hundred agents, then the admission a year later that the company had cut too deep and was rehiring people. I told it that way myself, in Issue 10. It comes from one interview and one article in May 2025. The version Klarna is accountable for is in a letter to the SEC two months later, during its IPO review, and it says the remarks were about a pilot to raise the quality of human service, did not relate to AI quality, and were misreported. The assistant’s own numbers held or improved. What the filings add is the part nobody put next to the saving: cost per contact fell 40 per cent in two years while the customer service and operations bill rose year on year in every filed quarter the note compares. Both true at once. A business case that carries only the first looks like a success in month one and a mystery by month nine, when the CFO asks why the line went up. The Klarna story is not a warning about AI. It is a warning about which number you let into the business case.
What you can do with it. The note closes with the complete AI business case on one page: everything Klarna’s 2024 case had, plus the four things it did not have to carry, with a signature line for the finance partner, the process owner and the agent owner. Open your own case beside it. Find the number it counts. Then find the number it doesn’t. The note is free to read, the template is free to use and share with credit, and every source is linked so you can check me.
Case Note 01. Klarna. The AI business case wasn’t wrong. It was incomplete.
https://aideliverydiscipline.com/economics/case-notes/01-klarna/
From here, one case note a month. Case Note 02 arrives in December. Each one tests an assumption enterprises are making right now, which means the company is the evidence, not the target. The series has its own home at aideliverydiscipline.com/economics, next to the two economics papers it tests.
One thing before you close this issue.
Does your current AI business case report cost per unit and total cost side by side? Reply (vijayan@aideliverydiscipline.com) with one word. Both, one, or neither. I read every reply, and the answers tell me which of these assumptions matters most to the people who sign.
The number that goes down will always win the budget. The job is to make sure it is not the only number in the room.
New issues arrive by email first. Subscribe here.
