In March 2026, Snowflake eliminated its entire technical writing department. Not a round of cuts. The whole team, roughly 70 people, gone, while the company was reporting 30% product revenue growth and courting more than 9,100 customer accounts on its AI push Snowflake’s reported AI-related team cuts. The writers spent their last weeks training the system that replaced them.
Your own leadership has probably watched a quieter version of the same scene: an AI turns a release-note outline into polished copy before the coffee cools, and the budget question follows naturally. If the tool can write this, why keep paying writers?
Here is the reframe worth sitting with before you answer that question. A practical workflow for using AI without surrendering the source is outlined in AI-assisted technical writing. AI will replace part of technical writing. It will compress drafting, formatting, and standardisation, and it will make some team structures hard to defend. But “will AI replace technical writers” is the wrong decision frame, because it asks about the role instead of the work. The work that survives is the work that makes documentation true, not the work that makes it readable.
What the employment data actually says
The US Bureau of Labor Statistics projects technical writer employment to grow 1% from 2024 to 2034, from a base of roughly 56,400 jobs, and says explicitly that AI productivity tools are the reason growth isn’t higher the BLS outlook for technical writers. That’s a flat market, not a vanishing profession. About 4,500 openings a year are still expected, mostly from people moving on rather than being pushed out the BLS outlook for technical writers.
Flat aggregate employment and one company deleting an entire department are both true at once. Neither number tells you what actually decides whether a given writer’s job survives contact with a language model. What does is narrower: does the work carry a cost if it’s wrong?
Where AI genuinely earns its keep
Start with what isn’t in dispute. In a preregistered MIT study of 453 professionals, ChatGPT cut writing task time by 40% and raised output quality by 18%, with the weakest baseline writers gaining the most an MIT study of AI-assisted writing productivity. A 2026 survey of 109 technical communicators found 62% already use AI regularly or daily, mostly for drafting, tone cleanup, and turning raw source material into a usable first pass Cherryleaf’s 2026 technical communication survey.
That’s AI working in its favourite conditions: known inputs, a repeatable format, a low penalty for a missed nuance. A team that refuses assistance on release notes built from clean, approved change logs is losing time for no defensible reason.
The blank page loses. It deserves to.
The danger starts when a company mistakes faster drafting for a complete documentation function. Drafting is visible, so it gets counted. Finding the authoritative source, resolving conflicting inputs, and testing an instruction against reality happen in meetings, test environments, and the moment an engineer says “that’s technically correct, but no customer should do it that way.” Those conversations decide whether a document deserves to ship, and AI cannot sit in on them.
The receipts
Some companies have found that edge the expensive way.
Deloitte’s Australian arm delivered a $290,000 report to the federal government to guide a welfare compliance crackdown. A University of Sydney researcher named Chris Rudge found a footnote crediting a real law professor with a book that doesn’t exist. “I instantaneously knew it was either hallucinated by AI or the world’s best kept secret,” he said, “because I’d never heard of the book and it sounded preposterous” Deloitte’s hallucination-driven refund. The report also contained a quote fabricated and attributed to a federal court judge, plus citations to nonexistent studies from two real universities. Deloitte refunded part of the final payment and disclosed, in the revised version, that Azure OpenAI had produced it Deloitte’s hallucination-driven refund.
CNET ran the same experiment on itself in early 2023. More than half of its 77 AI-assisted finance explainers needed corrections, including one on compound interest that got the actual math wrong CNET’s AI-assisted finance explainers. Sports Illustrated published articles under the bylines of writers who don’t exist, their headshots traced to an AI portrait marketplace Sports Illustrated’s AI-written article scandal. Consulting, journalism, and B2B publishing hit the same wall from three different directions.
The legal system is running the largest live version of this experiment. A live-updated tracker of AI hallucination cases in court filings has documented over 1,300 proceedings worldwide involving fabricated citations or quotes, most of them in the United States documented AI hallucination cases. Courts handed down $145,000 in sanctions for this in Q1 2026 alone, including $15,000 fines against each of two attorneys in one Sixth Circuit case and $15,500 against a lead lawyer in a federal case in Oregon the Q1 AI sanction wave. Every one of those citations resolved to a real-looking case name. None of them existed.
Fluency is cheap now. Evidence is not.
Notice what’s common across a consulting firm, a news outlet, a magazine, and a courtroom: none of these organizations were replaced by AI. They were fined, refunded, or humiliated by it, because nobody with authority checked the output before it shipped.
Why the stakes are higher in medtech and industrial documentation
A misleading SaaS help article creates support tickets. A misleading medical-device or industrial procedure creates a much uglier calendar.
The FDA’s January 2025 draft guidance for AI-enabled device software functions sets out recommendations for the design, development, deployment, and maintenance of devices that include AI, covering the documentation needed to support marketing submissions across the product’s life cycle the FDA’s guidance on AI-enabled device software. NIST’s AI Risk Management Framework warns that generative models can produce what it formally calls “confabulations”: erroneous or false content presented confidently, including internal inconsistencies and content that diverges from what was asked, colloquially known as hallucinations NIST’s generative AI risk profile. That’s a formal description of exactly what happened to Deloitte.
In regulated and physical-product documentation, the writer’s job is closer to source control than prose production: tracing a claim back to risk documentation, validated test results, and the specific decision that changed. O*NET’s own definition of the role includes writing “equipment manuals” and “operating and maintenance instructions” for a reason, not as an afterthought O*NET’s technical-writer role definition. AI can compare document versions, flag inconsistent terminology, or assemble a draft from an approved content library. It cannot become accountable for what it assembled.
The trust gap nobody’s pricing in
This is where the productivity story and the receipts collide.
A Q1 2026 survey found 40% of consumers trust marketing content less when they know AI wrote it, versus only 7% who trust it more research on B2B buyer trust in AI. Marketers are more optimistic about their own output than the evidence supports: 58% believe AI improved their content quality, but only 4% consider AI-generated content highly trustworthy without a human checking it first research on B2B buyer trust in AI. That gap, between what producers believe and what readers actually do, sits directly underneath every “just use AI for docs” decision, and it gets worse, not better, as the stakes of being wrong go up.
The operating model that replaces the old writing team
Tom Johnson, who has covered this industry for close to two decades, calls the pattern that actually survives contact with these tools the cyborg model: prompt, review, adjust, re-prompt, on a loop, with a human catching what the model has no context to see Tom Johnson’s cyborg model for technical writers. That’s not a hedge. It’s the only production pattern in this article that doesn’t show up in a court filing or a refund notice.
Building it means four gates, not four job titles.
Approved sources first. Product requirements, validated test results, and signed-off procedures sit ahead of the prompt. If the source is unclear, the workflow stops there, not after publication.
Assign by consequence. Let AI handle repeatable transformations: release notes from clean change logs, style-guide compliance, reformatting. Put human experts on ambiguous instructions, high-risk warnings, and anything that requires explaining why a user should act a certain way.
Name the reviewer. “Human in the loop” means nothing until you can say which human, with what authority, before or after publication. A loop without ownership is just a circle drawn on a slide.
Keep the evidence trail. A reviewer should be able to see the source, the tool or automation used, the edits made, and the approval decision. Useful in SaaS. Basic self-preservation in regulated work.
The labour signal supports building this now rather than waiting. In its 2025 survey, the World Economic Forum found 77% of employers expect to upskill workers for AI, while 41% expect workforce reductions where AI automates tasks the World Economic Forum’s 2025 jobs survey. Those are employer intentions, not proof that a whole function disappears. Entry-level production work is genuinely exposed. Treating that exposure as proof an entire discipline has become unnecessary is management by demo video.
Five questions to ask before cutting a writer role
Before removing a technical writing role or handing its output to a general AI tool, ask five questions.
- Is there a current, authoritative source for every material statement the tool will produce?
- Who resolves conflicts between requirements, implementation, and customer reality?
- What happens if a user follows an inaccurate instruction?
- Can a named reviewer validate the output faster than they could write the critical parts themselves?
- Can you show an auditor, customer, or engineer why the published version was approved?
If the answer to any of those is vague, you haven’t automated documentation. You’ve moved its risk to whoever discovers the mistake, and that’s rarely you.
Your next documentation hire shouldn’t be judged on how many words they can produce without AI. That’s the least scarce capability in the room now. What’s scarce is someone who can build the source system, interrogate a subject-matter expert, design the review gate, and put their name behind what gets published. The Technical Article Content Engine is built around that kind of source-led review. Give that person good tools and take the mechanical work off their week. Then ask the only question that actually matters: what does it cost you specifically if this is wrong, and who checked it before it shipped?