If AI is as good as claimed, the payback numbers should be easier to find.
TL;DR
Four shapes, over and over. Reach that became a breach. Savings that became a bill bigger than payroll. Oversight that automated work nobody understood. Speed that left a mess nobody can maintain.
And one number that keeps the question open: of 1,854 executives already running AI, 6% had seen payback inside a year. Most expect two to four — against the seven to twelve months they would expect from any other technology investment.
That gap is the quiet failure story, and it is far more common than the loud ones.
Security: the reach is the breach
An agent is useful in proportion to what it can touch, which means the thing making it valuable is also the thing making it dangerous. Most teams grant reach first and work out the boundaries afterwards, because the boundaries are the slow part.
Cost: the bill nobody forecast
Software is budgeted per seat. Agent work bills per token, and consumption rises with usefulness — so the better it works, the worse the invoice. Teams generally discover the shape of this after committing, which is the worst order to discover it in.
Management: automating what nobody understood
The appeal is covering more ground with the same people. The trap is that automating a process requires understanding it, and a surprising amount of organisational work has never been written down by anyone. Automate it anyway and you get a fast, confident version of something nobody could explain.
Engineering: quick, then unmaintainable
The first version arrives astonishingly fast. The second engineer arrives later and cannot tell which parts were decisions and which were accidents.
The conjecture
Here is the part we cannot prove. If AI were delivering what is claimed for it, the evidence would be easy to find and loudly advertised. Instead the published numbers are thin, the timelines keep extending, and the strongest claims consistently come from people selling something.
That is not proof of failure. Slow payback is not the same as no payback, and we should not pretend otherwise. But it is a reason to ask for evidence rather than assume it exists.
The payback figure above is Deloitte, AI ROI: The paradox of rising investment and elusive returns — 1,854 executives across 14 European and Middle Eastern countries, all already running AI. Worth reading before quoting, because it does not say the other 94% are failing.
This page is deliberately unfinished
The specific stories belong here and we have not collected them yet. Named incidents, dated, with sources — not the anonymous anecdotes that circulate and cannot be checked.
If you have one, bring it to a session. A failure you watched happen is worth more than anything we could summarise from a headline.