AI

The industrialization of technical writing

How enterprise economics are compressing the craft.

by
published
updated
read 14 min read

There is a quiet initiative happening inside almost every major enterprise software company right now. You won’t find it on a public roadmap. You will find it in the discussions, the experiments, the agentic workflows, the internal architecture diagrams.

Almost every engineering team is quietly building its own docs-orchestrator to bypass us. Almost every docs team is vibe-coding its own AI assistant to stay useful. Both sides sprinting. The finish line is the same room. Fewer chairs in it. Maybe none.

For years, we believed our process was sacred. We called it the Document Development Life Cycle. DDLC. We thought the sheer friction of navigating it made us irreplaceable.

It didn’t.

If you look at how organizations are actually implementing AI today, the DDLC isn’t a moat. It is a sequence of tasks. Every single stage is being industrialized. Quietly. While you read this. While you slept last night.

The autopsy of the DDLC

Let’s look at what is actually happening on the ground, phase by phase.

1. Requirement gathering

The Human

You stalk Jira boards. You schedule 30-minute discovery calls with SMEs who don't want to talk to you. You dig through fragmented Slack threads to figure out why the API changed. You test, experiment, read legacy docs, do detective work nobody will ever see.

The AI

Organizations are deploying agents using MCP (Model Context Protocol). The agent has direct, autonomous access to the Jira database and the Slack API. The moment a feature is marked "Code Complete," the agent queries the Jira ticket, reads the associated Slack thread, and rapidly synthesizes fragmented data in seconds. No meetings required.

2. Drafting and translation

The Human

You open a blank document. You try to translate the engineer's messy PR comments, scattered discussions, and half-remembered meeting notes into a conceptual overview and a 14-step tutorial. You stare at the cursor for ten minutes before the first sentence comes.

The AI

The engineer uses GitHub Copilot Workspace, Cursor, or Atlassian Rovo. Before the PR is even merged, the AI indexes the entire codebase context, reads the diffs, and generates the technical documentation natively. Are the generated docs perfect? No. They are often shallow, missing intent, and poor at edge cases. But the enterprise has decided they will tolerate that degradation because the first draft is instant. Users will experience more mediocre documentation long before they experience perfect autonomous documentation.

3. Review and validation

The Human

While drafting, you try to remember every style guide rule. Then you enforce them on yourself. Then you ship to three reviewers. You wait four days. You get a comment that says "looks good" from someone who never opened the file.

The AI

Organizations are building custom Claude skills and plugins, with their entire branding and style guide loaded inside. An automated editorial assistant runs deterministic NLP checks for brand compliance, and LLM checks for structural logic. It reviews a 5,000-word architectural spec in 14 seconds. It doesn't get tired. It doesn't miss the Oxford comma. It doesn't ask for a long weekend.

The economic meat grinder

This apocalypse isn’t about AI being “smarter” than us. It’s about AI being free at scale.

The disruption comes down to zero marginal cost. A senior human technical writer can draft 2,000 words a day. An agentic pipeline writes 2,000,000. Even if the human is stylistically superior, the human loses the war of attrition.

This is the CFO logic. If the documentation pipeline is 50x faster and 100x cheaper, the “quality gap” is an irrelevant luxury.

You don’t have to look far for the proof. Look at Zendesk’s AI agent auto-generating support knowledge base articles directly from closed tickets. Look at GitBook’s AI workflows summarizing PRs into user-facing changelogs without human intervention. These aren’t beta tests; they are active, deployed workflows compressing the marginal labor cost of documentation.

The 50% fallacy

When you present this picture to technical writers, the defense mechanism is immediate.

“Well, the agents hallucinate. They don’t have the institutional empathy. They can only do the basic stuff or what we tell them to do. We cannot trust them with critical tasks…” You can argue that right now, across the entire DDLC, an agent can only do 50% of the job.

And you would be right. Today.

But the 50% argument is a trap. It assumes the technology is static. And it assumes perfection is required.

The hallucination argument ignores a brutal truth: the enterprise has decided that a “good enough” 70% accurate doc produced autonomously in two seconds is fundamentally more valuable than a “perfect” 95% accurate doc that takes two weeks to write.

If a multi-agent system can do 50% of your DDLC today, what is the fundamental barrier preventing it from doing 70% next year? Or 90% the year after? What is the guarantee it won’t hit 100%?

Today
Agent does 50%. The human still drafts, decides, polishes. The team is intact.
Next year
Agent does 70%. The human reviews. The team is 70 percent of its size.
Year after
Agent does 90%. A handful of humans sample the high-risk pages. The team is halved.
Then
Agent does 99%. One Director signs off, by exception. The team is the Director.

“The hardest thing for a human brain to calculate is compounding interest applied to artificial capability.”

We treat our domain expertise like an impenetrable fortress. It isn’t. It is just a highly structured dataset. And models are exceptionally good at processing structured datasets.

Why this transition will be messier than evangelists pretend

Real enterprises are chaos machines. The transition will be slower and more dysfunctional than the tech demos imply.

There will be security blockers, compliance reviews, and AI governance committees slowing down every deployment. LLMs will hallucinate critical infrastructure steps. Automated pipelines will choke on terrible internal data hygiene, fragmented permissions, and shadow documentation. Legal teams will freeze deployments because of exposure risk.

But the economics overwhelm the friction. None of this stops the compression. It only slows it.

The segmented grave

We need to be clear: not all documentation dies at the exact same rate.

We are seeing a staggered industrialization. The hierarchy is clear: structured domains compress first, ambiguous domains compress slower, and liability-heavy domains compress slowest.

API references, release notes, and standard operating procedures are already in the coffin because they are predictable and directly tied to codebase changes. Aviation manuals, medical device documentation, and life-critical hardware specs are in the waiting room because they carry heavy legal and compliance weight.

But make no mistake—they are all in the same building.

The “I learned AI” fallacy

This is the cope with a LinkedIn badge.

“I did the prompt engineering course. I read the Anthropic docs end to end. I am Claude-certified. I built a tiny RAG pipeline on a Sunday. I am safe. I am the future of technical writing. The people who don’t learn AI will be replaced. I learned it. I am on the right side of history.”

Oh, good.

Congratulations on the certificate. The agent earned six new capabilities while you were on the coffee break between modules. It does not need a certificate. It is the thing the certificate is about.

You memorised the prompt patterns from May. The model that ships in September does not require those patterns. The patterns were scaffolding for a generation of models that has already been deprecated. The course you finished last week was teaching a workflow that was obsolete the day it was recorded.

The people on the LinkedIn lives shouting learn AI or be replaced are themselves about to be replaced. They have not noticed yet. They will. Around the time their second course stops selling, because the next generation of agents does not need the workflow the course was built on.

The half-life of an AI skill in 2026 is shorter than the half-life of the job it is supposed to protect.

You did not learn AI. You learned this year’s version of how to ride it. Next year’s version arrives without asking your permission.

The “institutional knowledge” fallacy

Then you might also be thinking: fine, but the agent doesn’t know us. Not really. Not the way I do.

This is the senior writer’s last refuge.

“The agent doesn’t know our codebase. It doesn’t know our culture. It doesn’t know the three architects who quit and took the system in their heads. It doesn’t know that our pricing page broke in 2023 and we never speak of it. We are the institutional memory. That is the moat.”

Oh, good.

Connect the MCP server to Confluence. Connect it to the wiki. Connect it to the Slack archive going back four years. Connect it to the GitHub history. Connect it to the Notion the engineering team quietly uses instead of Confluence. Connect it to the customer support ticket corpus. Connect it to the retro docs nobody reads anymore.

Now ask the agent about the 2023 pricing page incident.

It knows.

It knows in a way you do not, because it read every Slack thread, every retro doc, every support ticket from that week, and held all of it in working memory at the same time. You read three threads at the time and remembered the gist. The agent processed far more historical context than any single human realistically could.

Institutional knowledge was never the thing in your head. It was the thing in the systems. We were just the librarians.

The agent does not need a librarian when it can read the library directly.

The illusion of the “human in the loop”

The most dangerous lie circulating in our industry right now is the myth of the permanent editor.

It goes like this: AI will generate the content, but they will always need us to be the reviewers. We will be the quality gatekeepers. The human in the loop.

It’s a beautiful, desperate fiction.

Think about the brutal economics of it: why would an enterprise pay an engineering team to build a $100,000 automated documentation pipeline connecting Jira, Slack, and Cursor, only to bottle-neck it with a human reviewer who reads at 250 words per minute?

Human review scales linearly. Agent generation scales exponentially.

The “human-in-the-loop” isn’t a permanent career destination. It is a temporary speed bump on the way to full autonomy. As agents become more capable of reflecting on their own outputs, the loop closes. The human gets pushed out. First, they automated the drafting. Now, they are automating the reviewing.

Next, the chair.

The disruption is finite

There is one more meeting you have not been invited to.

The CFO is doing one math. Per page. Per writer. Per year. We saw that math. The CTO is doing a different one. The math of disruption.

The CTO does not ask “can we lose the team?” The CTO asks “how long will it hurt if we do, and is that pain less than 12 months of fully loaded salaries?”

Once the answer is yes, the meeting is over. The announcement is just calendar work.

Here is the curve the CTO is looking at. The one nobody puts on a public slide. The internal model often looks something like this:

Q1 · Months 0–3
Severe disruption. Support tickets spike. Docs go stale. Engineers grumble. A customer LinkedIn post goes mildly viral. Nothing breaks the business.
Q2 · Months 3–6
Manageable disruption. Engineering ships the docs-orchestrator. Agent coverage handles the stale pages. Tickets start trending down. Renewal calls go smoothly enough.
Q3 · Months 6–9
Background noise. Pipeline stable. Doc freshness exceeds the 2025 baseline. Cost-per-page collapses by an order of magnitude. The CFO gets a small applause in the board readout.
Q4 · Months 9–12+
Operational baseline. Steady state. Most KPIs net positive. The chart is green. The team is not back.

Look at the shape of that curve.

The pain is real. The pain is bounded. The pain recovers on its own. Payroll does not.

The CTO is not cruel. The CTO is doing arithmetic. Twelve months of disruption costs less than twelve months of fully loaded salaries. Year two is upside on every line of the spreadsheet.

The disruption is finite. That is what makes it survivable. That is also what makes it inevitable.

You can argue with the per-page math. You can argue with the human-in-the-loop math. You cannot argue with a curve that flattens to two percent in four quarters. It is the most boring chart in the deck. It is also the one that closes the meeting.

The “context engineer” paradox

If you are still employed as a technical writer right now, look closely at what you are actually doing.

I spent the last six months building custom Python orchestrators and RAG pipelines. I thought I was “enhancing the documentation experience.” I wasn’t. I was training my replacement.

We used to hold a monopoly on translation. The line between robots, humans, and technical writers used to be clear. But the machine learned to speak human. And the users realized they never actually wanted to read our beautiful, meticulously crafted, 14-step tutorials anyway. They just want an agent to fix their database connection.

You can pivot. You can become an AI Solutions Architect. You can transition from Technical Writer to Context Engineer.

But do not treat “Context Engineer” as a happy new career. It is the final stage of industrialization.

The new job isn’t writing; it’s cleaning the data supply chain so the agents don’t choke on fake “done” tickets. We are the people decommissioning the library. The orchestration layer survives longer than the execution layer—that is the key insight for survival. But what survives longest is not writing. It is organizational arbitration.

The strategic nihilism of being a Context Engineer is that the better you build the Context, the less the Engineer is needed. It is a self-consuming role.

The better you do your job, the faster you take yourself out of the equation.


The “credible gloom” reference table

A framework for understanding the compression of the craft.

Documentation Task Human Value (2024) Agent Value (2026) The “Gloom” Reality
Drafting High (Craft) High (Speed) Craft is now a commodity.
Editing Medium (Style) Near-total at scale (LLM) Style is operationalized.
Discovery High (Meetings) High (MCP/Jira) The agent doesn’t need “meetings.”
Governance High (Trust) Low (Risk) Humans become “Liability Shields.”

The sprint is over. We have finally reached the finish line.

That quiet initiative is no longer quiet. The internal architecture diagrams are now deployed pipelines. The docs-orchestrator is live.

And when you look inside that room, you see exactly what the spreadsheets demanded.

The dashboards are green. The support queue is moving. The organization has stopped caring whether the documentation is beautiful.

And there are no chairs.

— end of essay —

Personal essay. Views are my own and do not represent any current or former employer.