AI failures never announce themselves
Nothing threw an error yet the system was failing to act as expected...
I said the bots were costing me $50 a day. I said it publicly, mid-incident, before I had read the thing properly, because the traffic was right there and it looked guilty: 5 IP addresses out of Singapore, straight at the sitemap, no browsing behaviour, hammering a long tail of malformed URLs.
About 2/3 of that bill turned out to be mine.
3 defects, none of which produced an error:
New pages shipping with no caching instruction, because doing nothing is a valid choice that happens to be the most expensive one.
A release that changed tens of thousands of addresses in a single step and discarded the whole cache at once.
And the one I liked least: pages that did request caching were being silently downgraded by an unrelated line elsewhere in the build. The code read correctly. The build did not do what the code said.
No alert. No failed deploy. No red anything. The money just left.
I have been reading this year’s research on AI at work with that incident still fresh, and the same shape keeps showing up at a scale much larger than my server bill.
The working day filed no complaint
ActivTrak’s Productivity Lab published a study in March that almost nobody picked up. They compared telemetry from 180 days before and after AI adoption across 376 companies and 10,584 users. Not a survey. Measured activity.
After adoption, email went up 104%. Chat and messaging up 145%. Business management tools up 94%. Every single measured category of work rose, somewhere between 27% and 346%. Daily focused time fell by 23 minutes for AI users while barely moving for everyone else.
Not one category went down.
The obvious reading is burnout, and the obvious reading is wrong. The same dataset has AI users spending 9% more days in healthy utilisation patterns than non-users, and burnout risk across the whole population falling 22%. People were not suffering. They were producing.
That is the part worth slowing down on. AI made everybody faster at generating the things that create work for everybody else. More drafts, more decks, more threads, more thoughtful replies. Each one lands in somebody’s queue, and that somebody now answers with AI too. The volume is the output and the volume is the cost, and no dashboard anywhere separates the two, because from the tooling’s point of view nothing has gone wrong. Usage is up. Adoption is up. Engagement is up.
Buried in the same report: employees who spend between 7% and 10% of their working hours in AI tools sit in the highest productivity tier measured. Only 3% of users are in that band. Almost the entire population is either barely touching it or swimming in it, and neither state generates a warning.
Neither did the P&L
In March, nearly 750 corporate executives were surveyed by Duke with the Atlanta and Richmond Feds. The paper went up at NBER. It asks 2 things that are usually asked separately: what did AI do to your productivity, and what did AI do to your revenue and your headcount. The first is a feeling. The second lets you calculate the answer instead of asking for it.
For 2025, executives reported 1.8%. Their own revenue and employment numbers imply 0.6%.
For 2026, they report 3.0%. Their numbers imply 1.8%.
Read those 4 figures again in order. What executives reported feeling in 2025 is exactly what their own 2026 forecast implies. The same 1.8%, 12 months apart.
The authors are careful about this and I will be too. They call it a possible resolution to the productivity paradox, and both sets of figures come from the same people, so what we have is a feeling running about a year ahead of a forecast, not yet a feeling vindicated by audited accounts.
But it does settle an argument I keep watching people have. When your team says AI is working and the P&L does not agree, nobody in that room is lying. One of them is reading a signal that has not reached the other one yet. And in the meantime, headcount decisions are being made on the strength of a gain that has not landed, in a year when the same paper expects total employment to fall less than 0.4% because of AI, with small firms expecting to grow.
Nothing errors. The forecast is just quietly wrong for 12 months, and by the time it is right the people are gone.
And then the review went quiet
The third one is the one that made me sit up, and it is the one every write-up I have seen has reported incorrectly.
Boston University and BCG ran a randomised experiment with 813 HR and finance managers. Identical documents. Identical planted errors. The only variable was the label on the work: output from an AI tool the manager had used, or output from a named “AI employee”.
Across the whole sample the effect was small. That is the finding most coverage skipped straight past.
Among managers whose organisations already list AI agents on the org chart, and that is 23% of them, the “AI employee” label produced 18% fewer errors caught, 16% worse review performance, 9 points less accountability assigned to themselves, 8 points more assigned to the AI, and 22 points more escalation for further review.
The last 2 belong together. Those managers did not become more confident in the work. They stopped checking it personally and pushed it to somebody else. Oversight did not improve and it did not vanish. It moved, and on the way it thinned out.
So the danger is not the pilot. It is the promotion. Trying an agent changes very little. Giving it a name, a box on the chart and a job description is what switches off the human review layer, and it does so without a single meeting in which anyone decides to review less.
The researchers also found no increase in adoption intent, no more excitement, no greater willingness to learn the tool. You take the oversight cost and get none of the enthusiasm you thought you were buying.
What a check looks like when it works
What eventually fixed my invoice was not care. I was already being careful. 2 of the 3 defects were fixes I had written and that were quietly doing nothing, which is a particular kind of humbling.
What fixed it was making the build fail. The caching rules stopped being something I was trusted to remember and became something the pipeline enforces, so the failure mode now breaks a deploy instead of costing money in silence. Cache misses went from 63% to 26%. Database queries per hour fell 56%. Rows returned per hour fell 78%.
The management version of that is not a policy document. It is one number you agree to watch that would move if the thing you fear were happening, chosen before you have any reason to fear it.
If you adopt AI across a team, the number is not usage. It is what happened to the volume everyone else now has to absorb. If you claim a productivity gain, name the quarter it should appear in revenue and the figure that would confirm it, and write it down while you are still optimistic. If you put an agent on the org chart, measure your review catch rate first, because afterwards you will have nothing to compare against and 18% is not a gap anyone notices from the inside.
A caveat I would rather state than have someone find: that experiment is roughly 85% American, and the 18% is a pre-registered subgroup rather than an average. It tells you what happens where the practice is already normal, not what happens everywhere.
I still do not think the crawlers meant to find my seam. They hit the sitemap the way crawlers do, and the sitemap happened to point at a site that had just gone cold. What made it expensive was everything before that, which was working exactly as badly as it had been configured to and reporting nothing at all.
The usage dashboard was the first honest signal I got, and by then it was a description of the past. That is what these 3 studies have in common with a hosting bill. All of them are telling you what already happened, in a system that had no way of telling you while it was happening.