The context nobody wrote down

I keep reading the same story with the serial numbers filed off. A company spends a fortune building an AI agent. It's perfect in the demo. Then it meets real work and starts inventing answers with total confidence.

I know this failure. Until recently, part of my job was finding gaps in developer documentation and getting the facts straight before customers relied on it. Remove that work, and errors in the source material spread downstream at machine speed.

It's easy to create demos

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of escalating costs, unclear value, or inadequate risk controls. Rough headline.

In a June 2026 VentureBeat survey of 101 enterprises, 57% said they had traced a confidently wrong answer to missing or inconsistent business context. A month later, 68% reported the same failure. The model is only one piece. The agent also needs a reliable account of how the company works.

A documentation problem in an engineering costume

An agent is only as good as what it can read. Point it at a knowledge base full of holes and it fills the holes itself, fluently, and wrong.

Technical writers deal with this gap all the time. It's the stuff that lives in one person's head and never reaches a page. Why the refund flow has an exception for a single country. The config the support team tells people to change when a setup breaks. Nobody wrote it down, because it was probably faster to just ask Sofia, who's been here nine years and knows.

AI can already produce serviceable prose. Technical writing earns its value earlier, during discovery. You sit with an engineer who swears the behaviour is "obvious," and keep asking dumb questions until the assumptions they didn't know they were making fall out. Then you write them down so a stranger can act without you in the room. That stranger used to be a new hire or a customer. Now it's also a machine that will follow what the page implies and can't tell when the page is lying by omission.

What disappears with the team

Companies are cutting roles while describing AI as an efficiency lever. The causal story is messier than the announcements make it sound, but the operational bet is clear: fewer people, more automated output.

If you think a docs team's product is text, a chatbot that can generate text makes the team look expendable. The missing work happens before the text: asking basic questions until exceptions and contradictions surface, then working out which source deserves trust.

AI can compare the sources a company has recorded and flag inconsistencies between them. Finding what nobody recorded still requires access to Sofia and enough judgment to know which question is worth asking.

Someone has to find the context and make it usable. Until recently, that was my job. It's the work I want to keep doing.