Companies are automating documentation to cut costs. AI agents are turning that same documentation into something their customers now depend on.
In their final weeks, according to accounts from affected employees, technical writers at Snowflake helped train the system built to replace them. In March 2026 the cloud-data company eliminated its entire technical writing and documentation department, a decision first reported by Benzinga, which noted that Snowflake was following Amazon and Canva toward AI-generated content.
The context is what made the move notable. Snowflake was not cutting to survive. The company had just reported roughly 30 percent product revenue growth. The layoffs followed a 200 million dollar partnership with OpenAI, signed in February, that integrates GPT-5.2 into Snowflake's platform. That deal powers Project SnowWork, an agentic system the company announced in mid-March that drafts API documentation and user guides directly from source code in minutes, work that previously took human teams weeks of research and interviews.
It is the most direct bet so far that documentation is a cost to automate rather than an asset to fund. The evidence suggests the timing is wrong.
Why companies are cutting documentation
Technical documentation has rarely been easy to justify on a budget. It is necessary, it is expensive to produce, and it is seldom tied to a revenue number. When companies need to redirect spending toward AI, documentation is an obvious target.
The first quarter of 2026 showed the pattern. Trackers reported tech layoffs up about 51 percent year over year, much of it from profitable companies funding AI investment. Atlassian (News - Alert), the company behind the Confluence wiki used for internal documentation at many enterprises, cut around 1,600 roles, about 10 percent of its staff, to redirect money toward AI. Snowflake's documentation team, reported at roughly 70 people though estimates vary, was a smaller case in the same quarter.
The reasoning is easy to follow. Generative models have lowered the cost of producing usable reference material. If an agent can read a codebase and produce an API reference overnight, paying a department to do it more slowly looks inefficient. On a cost basis, the decision looks reasonable.
The problem is that the cost basis measures the wrong thing. It counts pages produced per dollar at the point when the value of documentation has moved to a different place.
Documentation became a product surface
Documentation now has a second audience, and that audience is not human.
When a developer adopts a tool in 2026, they often do not read the manual first. They ask an AI coding assistant such as Cursor, Claude Code, or GitHub Copilot, and that assistant answers by reading the product's documentation for them. Mintlify, one of the larger documentation platforms, has reported that close to half of the traffic to the documentation sites it hosts now comes from AI agents rather than people. Documentation has become the interface through which software systems determine what a product does and whether to recommend it.
A vocabulary has formed around making content readable by machines. The Model Context Protocol, the connector standard Anthropic released and later transferred to the Linux Foundation (News - Alert), has become a common way for AI agents to reach external tools and data. By early 2026 the foundation counted more than 10,000 public MCP servers. A lighter convention called llms.txt has appeared as well, giving AI systems a curated guide to which pages matter. Some analysts describe this layer as business-to-agent: companies that once needed a website for people and an app for phones now also need content that agents can read accurately.
Adoption is still limited. One study of 300,000 domains found llms.txt on about 10 percent of sites. The direction is clearer than the adoption rate. When an AI agent gives a wrong answer about a product because its documentation was thin or outdated, the result is not a support ticket. It is a developer who picked a competitor the model described more confidently.
That changes the cost calculation. Documentation that drives adoption, lowers support volume, and supplies accurate answers to the AI tools customers use is part of the product, not overhead. Investors have priced this view, valuing standalone documentation companies at hundreds of millions of dollars.
The quality question
This does not mean AI has no role in documentation. It clearly does. The useful question is where human judgment still adds value.
The common objection to fully automated documentation is that quality falls. Snowflake is betting it will not, that a model trained on its writers' methods can match them. Early reports said documentation quality had held steady after the change, which is the part that should concern other knowledge workers. But steady in one quarter and reliable across a year of product changes are different claims. Documentation fails when what the code does and what the docs say no longer match, and that mismatch grows with every feature shipped. A model can regenerate a reference. It cannot yet attend a planning meeting, understand why a confusing feature was built a certain way, or judge what a specific frustrated customer needs. Remove the people who held that context and the cost may not appear until the next major release.
For a company that sells to technical buyers, weak documentation is not an internal issue. It is a weaker product experience, delivered through every AI assistant that now reads those docs to answer questions about Snowflake.
What teams should do instead
The lesson is not that AI has no place in documentation. It is that documentation should be treated as infrastructure a company owns rather than a line item it trims. That changes the questions worth asking.
The first is about platforms. Many teams are finding that the wikis and tools they collected for human readers handle machine readers poorly. Older platforms were built for visual editing and team collaboration, and they lag on structured exports and agent readiness. Teams reviewing those tools are weighing a gitbook alternative or a more developer-focused system that stores documentation as versioned, structured content that agents can parse. The decision depends on whether the platform produces output that both people and software can use, not on which editor feels nicer.
The second is build versus buy, where Snowflake's choice is hardest to defend. Building a custom AI documentation pipeline, as Snowflake did, is costly and fragile, and it diverts engineering attention away from the actual product. Established documentation platforms already include the capabilities that matter now, including API reference generation from specs, structured output that agents can read, semantic search, and maintenance workflows that keep content aligned with the code. Mintlify grew by offering exactly these features, and teams that want the same capabilities on different terms can choose an alternative to mintlify rather than build the pipeline in-house. For most teams, buying that capability costs less and breaks less than firing the team and rebuilding the function internally.
The third question is about people, and it is strategic rather than sentimental. Practitioners increasingly describe the approach as augmentation rather than replacement, with writers moving from drafting text to reviewing AI output, deciding what agents should read, and keeping a fast-moving product's documentation accurate. Teams that keep that judgment while automating the repetitive work tend to produce better documentation at lower cost than either all-human or all-machine approaches.
Snowflake could still be right. If SnowWork holds its quality through a year of feature releases, the decision will read as an early and clear call. The more probable outcome is that Snowflake optimized a function it had already stopped valuing, at the point when AI agents were making that function central to how products get chosen. Cutting the people who maintain documentation to save a small fraction of headcount is not efficiency if the documentation is what customers and their AI tools now rely on.
