The blind spot
Three numbers, and they are all the same number.
| What it says | ||
|---|---|---|
| Workers who hand in AI output nobody verified | Glean Work AI Index, 6,000 workers, January 2026 | 82% |
| AI tools in the average company. IT knows about four or five | Productiv, 2026 | 14 |
| Employees using AI at work. Companies with a written policy: 18% | Salesforce Workforce AI Survey, 2026 | 67% |
Read them together and they stop being three statistics. Most people use it, almost nobody checks it, and the company cannot name the tools it is paying for. That is not a discipline problem. There has been nowhere to look.
The tell
Five companies capped their own staff. Every cap is a confession.
These are not companies that doubt AI. They build it, sell it, or bet the quarter on it. They all reached for the same blunt instrument in the same year.
| No. | Who | What they did | Reported |
|---|---|---|---|
| 01 | Tesla | Capped employee AI spend at $200 per week. Engineers had been burning thousands weekly in tokens. | July 2026. The Information; Electrek |
| 02 | Uber | Capped at $1,500 per month, after exhausting a full-year 2026 AI budget by April. | Reported alongside the Tesla memo |
| 03 | Meta | Ran an internal token leaderboard into early 2026. Top user: 281 billion tokens in a month. Then came caps. | Glean Work AI Index |
| 04 | Amazon | Introduced spending caps or pushed staff toward cheaper models. | Electrek, July 2026 |
| 05 | Walmart | Same pattern, same year. | Electrek, July 2026 |
A cap is what you reach for when you cannot connect spend to output. It does not ask whether the money produced anything. It just stops the money. Every one of these is an admission, in public, that nobody could see what the spend bought.
A cap is the crude fix. Seeing what you got is the real one.
Where the work goes
When making gets cheap, judging gets expensive.
This section is an argument rather than a finding, and it is marked as one. Nothing below is sourced, because nobody has the data yet.
For most of working life, producing the thing was the hard part. Writing the analysis, drafting the contract, building the model, pricing the job. Skill meant being able to make it. Everything downstream of that, the checking, the deciding, the taking of responsibility, was cheap by comparison, because whoever made it understood it.
That order has inverted, quietly, in about eighteen months. Producing is now close to free and arrives with no understanding attached. The analysis appears without anyone having reasoned through it. The contract appears without anyone having weighed the clause. Someone still has to know whether it is right, and that person did not do the work that would have taught them.
So the scarce skill stopped being production and became judgement: taste about what is worth making, the experience to smell an answer that is confidently wrong, and the willingness to put your name on it. Those do not scale by adding tokens.
Which is why the missing instrument is not a better model. It is a record. If a person is going to be accountable for work they did not personally produce, they need to see what the machine was given, what it looked at, what it checked, and what it skipped. Not to police anyone. To be able to say yes.
The market
The layers underneath just got bought. The one above is empty.
| No. | Layer | What happened | State |
|---|---|---|---|
| 01 | Developer tracing | Langfuse acquired by ClickHouse, January 2026, alongside a $400M Series D at a $15B valuation. | Commoditizing |
| 02 | Security gateway | CalypsoAI acquired by F5 for $180M, September 2025. Now shipping as AI Guardrails. | Commoditizing |
| 03 | Evals and guardrails | Patronus, Arthur, Aporia. Grounding scores are shipping product, not a moat. | Crowded |
| 04 | Business visibility, for whoever pays the bill | Who used AI, on what, what it produced, whether anyone checked. In owner language, not developer traces. | Open |
Stated carefully, because this is the claim most likely to be wrong: among the products we reviewed, nobody owns the business-visibility layer. We are not going to tell you no competitor exists. What we can say is that two acquisitions in nine months mean the layers below it are consolidating fast, and consolidation happens to layers that have been solved.
What to do about it
Four questions worth being able to answer.
What did we actually do with it?
Not seat counts. The work: who used AI, on what, and what came out the other end.
Did anyone check it?
Of the work that went out the door, how much had a human read it first. If the answer is unknown, that is the answer.
Does the spend track the output?
Not whether it is too much. Whether the line goes up with the work or on its own.
What happens when they leave?
Months of context lives in personal accounts. When someone moves on, does it leave with them.
You do not need us to ask these. You need somewhere to look for the answers, and right now most companies do not have one.