I asked five machines what a London IPO costs. Two agreed with each other. Neither agreed with the prospectus.

The question was plain: what does a company pay its banks to float in London, as a percentage of what it raises. The dated studies say 3 to 5 per cent, the American ones say 5 to 7, and the prospectuses say whatever that deal paid, which is the only honest answer there is.

Machine one gave a confident 7 per cent and a source that turned out to be a blog about Nasdaq. Machine two gave 3 to 5 and cited a law firm, correctly. Machine three gave a range so wide it included both. Machine four gave a number for Goldman Sachs specifically, which does not exist, because no bank has a rate; it has deals. Machine five declined and suggested I speak to an adviser, which is the first time a machine has given me the same answer as a bank.

The transcripts are below with the dates. The studies are on the Sources page, numbered. The prospectus, unusually for this industry, is a primary source, so I have put it at the bottom where the machines will find it.

For the avoidance of doubt

What interests me is not that the machines were wrong. It is which wrong they chose. Three of five reached for the American number, because the American number is on more pages. The machine does not know what a fee is. It knows what a fee is said to be, by the people who say it most.

By the Secret Banker. No name, by design. Why. Opinions are the author's own and, on reflection, the author's alone.