The Unread Loan: What Your AI Assistant Forgot to Mention
There is a particular smell in a small British software team the week after a bank holiday. It is not coffee. It is the faint ozone of a dozen AI assistants being asked to 'just scaffold the billing module' while everyone nods along. By 2026, most of us have stopped pretending this is a novelty. The assistant writes, the tests pass, the pull request goes green, and somewhere in the corner of the room a small unpaid invoice begins to curl at the edges.
The invoice arrives quietly
Prompt debt is not a line item on your management accounts. It is the interest you pay on code you did not type, cannot fully explain, and have not yet had the courage to delete. The principal was borrowed in seconds. The interest compounds every time someone opens the file, squints at a regular expression that looks like a cat walked across the keyboard, and decides to leave it alone because it is probably fine.
In a large organisation this is merely annoying. In a small British team — three engineers, a product person, and a founder who still answers the support inbox — it is a solvency risk with a nice user interface. You shipped faster than you could understand. That is not a crime. It is a loan.
What the interest actually looks like
The first payment is comprehension. A developer joins, opens a module, and asks why the tax calculation branches on a value called legacy_flag. Nobody knows. The assistant knew, briefly, in a previous conversation that has since been deleted by a data retention policy. Now the team has a folklore problem disguised as a codebase.
The second payment is review. Reviewers look at AI-generated code the way Brits look at a neighbour's new extension: politely, briefly, and without asking the question that would make everyone uncomfortable. The tests pass. The linter is happy. The logic is plausible. Someone approves. The interest ticks up.
The third payment is change. You want to alter a feature. The code was generated in one go, in a style nobody on the team writes, with abstractions that exist only because the model was in a mood. You spend an afternoon tracing a function that calls a service that calls a helper that returns a promise wrapped in a try/catch that catches nothing. You have not been debugging. You have been translating.
The statement: a five-column audit
You do not need a task force. You need a spreadsheet, an hour, and the sort of grim satisfaction that comes from opening post you have been avoiding. Pick the ten modules that worry you most. For each one, fill in five columns.
- What is it? A plain-English sentence. If you cannot write one, that is the finding.
- Who prompted it? A person, not a tool. Accountability is not a vibe.
- Why does it exist? The ticket, the customer, the deadline, the bad idea. Context is the collateral.
- What does it actually do? Not what the ticket says. What the code does when the input is empty, the API is down, and the clock crosses midnight.
- Confidence. One to five. Be honest. A three is a two wearing a tie.
The point is not to shame anyone. The point is to turn a vague sense of unease into a list you can act on. Prompt debt thrives in the fog. A statement brings foglights.
Paying it down without banning the robot
The wrong response is a dramatic ban followed by a quiet reintroduction three weeks later. The assistant is useful. It is also a lending facility, and you would not let the finance team borrow without a ledger.
So add a small ritual. When AI-generated code lands, the author writes one paragraph explaining the logic in their own words. Not a comment block. A paragraph, in the pull request, for a human. If they cannot write it, the code is not ready. That single habit catches more debt than any policy.
Second, mark the strange bits. A short comment that says 'generated, not fully understood, revisit before scaling' is not an admission of failure. It is a warning label on a borrowed ladder.
Third, schedule a paydown. Once a quarter, pick one module from the statement and either explain it, simplify it, or delete it. You will be amazed how often the correct answer is delete it. The assistant was very keen on that feature. The customers were not.
The bit where we admit we use it too
None of this is an argument for artisanal, hand-carved code written by a person in a shed with a fountain pen. We like shipping. We like the assistant. We also like being able to change our own software without holding a seance.
Prompt debt is not a moral failing. It is a normal consequence of borrowing speed from the future. The trick is to read the statement before the bailiffs arrive. In 2026, the bill is arriving for a lot of small British teams. The ones who will be fine are not the ones who never borrowed. They are the ones who kept a ledger.