AI Is a Tool. The Market Priced It as a Replacement.

We have seen a bubble before. We have not seen one shaped like this. The dot-com wave spread risk across thousands of small companies. This one rests on a handful of the largest technology firms on earth, all buying the same chips, all building the same data centres, all priced for a revolution that arrives in a specific form or does not arrive at all.

I work with these tools every day. I build software with them. The case below is not that the technology fails. The case is that the economics assume something the technology does not do.

The tool is real

Separate the utility from the story told about the utility.

As an assistant, the technology delivers. Developers accelerate coding, generate boilerplate, and debug faster. Teams query internal knowledge bases and summarise dense documentation in seconds. Scoped, supervised, bounded work produces immediate value. I use it this way and it pays for itself.

The financial danger starts when a leader hears "excellent assistant" and buys "independent worker." Those are different products. Only one of them exists.

What Microsoft found when the human left

In April 2026, Philippe Laban, Tobias Schnabel and Jennifer Neville at Microsoft Research published LLMs Corrupt Your Documents When You Delegate. They built a benchmark called DELEGATE-52 and ran 19 models across 52 professional domains, from Python to crystallography to music notation.

The headline finding sits in the card above. Across long delegated workflows, frontier models corrupted an average of 25% of document content. Weaker models failed more severely.

Three details matter more than the number.

• The errors are sparse and severe. Corruption does not accumulate smoothly. Between 80% and 98% of the total degradation comes from isolated critical failures rather than steady erosion. Stronger models do not avoid those failures. They postpone them.
• The damage does not announce itself. Frontier models tend to modify content where weaker models delete it. Deletion is visible. Modification is a qualifier that changed meaning in a paragraph you were not rereading.
• Agentic tooling made it worse. Giving models direct file access did not help. The four models tested that way performed about 6% worse. Longer interactions, larger documents, and unrelated files in context all degraded the result further.

Structured, programmatic domains held up best, and Python was close to lossless. Natural language and niche formats were where the catastrophic cases sat. Delegate a contract to an unsupervised model across twenty steps and it does not edit your document. It rebuilds your document from statistical probability, and it drops a qualifier somewhere you never look.

The equation that does not balance

Capital expenditure assumes future licensing revenue on a scale nobody has yet demonstrated anyone will pay. Meanwhile the cost of running the technology behaves nothing like the software line items finance teams know.

Uber makes the case better than any projection. The company rolled Claude Code out to roughly 5,000 engineers in December 2025. By April it had consumed its entire 2026 AI budget. Four months into a twelve month plan.

Read the detail carefully, because the obvious conclusion is the wrong one.

• Adoption was a success. By spring, 95% of engineers used AI tools monthly, and roughly 70% of committed code came from them. Worth separating from the headline version of that stat: CEO Dara Khosrowshahi has put the share written by autonomous agents at around 10%. The other sixty points are a person driving a tool.
• The pricing model broke the budget. Average cost landed between $150 and $250 per engineer per month. Heavy users ran $500 to $2,000. CTO Praveen Neppalli Naga said he personally spent $1,200 during a two hour demo.
• The incentives pointed the same way as the spending. Uber ran internal leaderboards ranking teams by AI usage. When consumption is the scoreboard and consumption is the bill, there is nothing in the loop pulling the other way. The CTO confirmed there were no plans to add budget for the rest of the year.

This is the part executives miss. The tools worked. The budget still detonated, because token consumption does not behave like a seat licence. Agentic workflows consume many times what a chat interaction consumes. You are not buying software. You are buying a metered utility with no natural ceiling. The standard mitigations being passed around now, prompt caching, spend caps, routing routine work to cheaper models, are all sensible, and all of them are things you build after the invoice teaches you to.

Where the savings go instead

Replacing payroll with algorithms moves cost rather than removing it.

• Compute and operations. Every advanced query consumes real power at real prices, against constrained chip supply.
• Data preparation. Models need structured, clean inputs. Most organisations must fix their internal data before deployment delivers anything.
• Quality assurance. Because output degrades silently, you must staff human review. The control layer you removed reappears with a different job title.
• Risk. A dropped liability clause or a confidently wrong answer to a customer costs more than the salary it displaced.

When the correction comes

A bubble bursts when investors accept that future revenue does not cover present spending. The dot-com era laid fibre that sat dark for years before it became the internet we use. Today's specialised data centres and processors face a similar question about timing.

The correction arrives when enough organisations complete the same arithmetic. The assistant is worth paying for. The replacement was never on sale. When that conclusion spreads, demand decelerates, and the firms that staked everything on the second story take the damage.

AI does not disappear in that scenario. The pricing does.

There is a version of this argument about the people rather than the money, which I wrote about in the cost of outsourcing our thinking. And the labs themselves have started saying the quiet part in public, which I covered when the sellers began asking for speed limits.

Buy the assistant. Supervise the output. Treat any promise of autonomous knowledge work as a claim awaiting evidence, because the people measuring it carefully keep finding the same thing.