What to remember from this issue
If the specification keeps every old habit, you are not buying a new system, you are paying to copy the old one.
The field note
In 2010 I joined a medical society in Lausanne to run its website. The society had just spent three years and about two million francs moving everything onto SAP: the congress, the courses and exams, the scientific programme and the accounts. The system went live late and full of bugs. I said it would not work, my boss agreed, and neither of us had a say.
The failure was written into the requirements. Nobody wanted to change the way they worked, so the document listed every way that every team had ever done every task, and the vendor built all of it. Each team then tested the new system against its own habits. The result was the old company copied into software, with bugs, for two million francs.
A year later, a blackout hit the city centre. The server was plugged into a wall socket instead of the backup power, the backups had never been tested, and a consultant had deleted the table that says where the data lives, in order to save disk space. We could see the data, but we could not get it back. The congress was one month away and there was no member system.
So we built one. We recovered the member list from old exports and from people’s mailboxes, and a very simple application, with one path from registration to badge, ran the congress. It then took over membership, and then the rest. SAP kept the accounts, where it worked, and lost everything else.
In one month, the disaster did what three years of meetings had refused to do: it forced people to choose. That happened not because our tools were better, but because the deadline cut what nobody had wanted to cut.
Why does this matter in 2026? The consultants took three years to build every variant. An AI agent would build them in three weeks, and it would never ask whether a society needs twelve ways to register. Building got cheap, but deciding what to build did not. The slow part was never the building. It was the choosing, and it still is.
Try this week
Pick one workflow your team is about to automate with AI, and count the ways it is done today. Any way used by a single team is a habit, not a requirement. Write the number down, because that number is the specification work nobody has done yet, and no AI will do it for you.
Seen this week
Two people wrote a specification, and an AI built an entire chip design from it in under two weeks.
On 27 August, Architect Labs announced Redwood, a chip designed to run AI. By the company’s own account, two human architects wrote a high-level specification and there was “no human intervention below the specification”. Everything underneath, from the performance model to the drivers, was generated. It runs today on programmable test hardware (an FPGA), not yet on a manufactured chip. Two people wrote one document, and the whole human contribution sat above the line they drew. Architect Labs, “Introducing Redwood”
Someone said it better
“If you took AI out of your organization today, would any workflows change? And the answer for most people is no, which tells you that they’re tinkering with AI but they’re not really making structural change.”
Salim Ismail, on Moonshots with Peter Diamandis, 5 September 2026, at 19:48. Ask that question about the society’s SAP: unplug it, and everyone keeps working exactly as before. A blackout eventually ran the experiment.