In August, AI agents made 257 million requests to documentation sites running on Mintlify, compared with 131 million page loads from people. Agents now read company knowledge at nearly twice the rate humans do. But that doesn’t mean that human developers have stopped reading; they’re now just delegating more. Consider a developer asking a coding agent to wire up an API, or a buyer asking a chatbot for vendor comparisons, or a customer pinging a support bot about why a payment failed.

To be sure, my company Mintlify could directly benefit from this argument, but the truth remains that each of those agents goes looking for your company docs, your help center, and your policies, and takes whatever it finds at face value to respond to those requests.

This is a fundamental change in who your documentation is for, and most companies are still writing for an old audience.

Every company has knowledge that looks authoritative, but actually isn’t. Think about a setup guide someone wrote carefully a month ago and never touched again. Or a pricing page that actually now contradicts the updated help center. A person might notice something feels off and ask a colleague for clarification, but an agent won’t. It picks up the version it found, writes it into code or a support answer, and repeats the mistake thousands of times before anyone catches it.

Developers are already feeling this pain. Recent research found that the top frustration named by 66% of developers was AI output that is close to correct but still wrong. Some of that is the models, but in my experience, a lot of it is the material we feed them.

So who owns this problem? At most companies, nobody. Documentation has long been treated like a publishing task, which worked when the reader could fill in the gaps. With the rise of agent readers and writers, this process no longer works at scale.

What’s emerging instead looks closer to infrastructure work. Someone has to decide which source is authoritative when two conflict. Someone has to connect product changes to the content that describes them, so the docs update when the product does. Someone has to structure information so agents can retrieve it cleanly, and track what agents search for and fail to find, because those gaps are now a direct signal of where customers are getting bad answers.

I think of this as knowledge engineering. It’s less about producing pages and more about keeping a company’s knowledge accurate, current and usable by both people and machines.

This is a similar shift in roles to what’s happening in radiology. In 2016, AI godfather Geoffrey Hinton told the world to stop training radiologists because AI would make them obsolete. Instead, Mayo Clinic, one of the most aggressive adopters of AI in medicine, grew its radiology staff 55 percent between 2016 and 2025, and U.S. radiology residency programs offered a record number of positions in 2025. Jensen Huang has argued that the doomsayers mistook the task of reading scans for the whole job. In reality, radiologists now spend less time recognizing images and more time on patients, and demand for those roles is growing.

The practical shift is in how you measure success. Page views and time-on-page tell you how humans engage. They tell you almost nothing about whether an agent got the right answer. The companies getting this right are watching different signals: which queries agents run, where retrieval comes up empty, where sources contradict each other, and how quickly a product change shows up in the knowledge base.

Your documentation used to be a reference people consulted. Now it’s an input that machines act on, often without a human in the loop. The companies that come out ahead will be the ones that treat their knowledge like the infrastructure it has become.

Han Wang is cofounder and CEO of Mintlify.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.