Nearly every company has tried AI for writing, translating or organizing content. Few have made it past the pilot. You know the script: the demo impresses, someone runs a test, the test works, and months later nothing in the operation has changed.
Sanity, a content platform used by thousands of companies, set out to measure why, using behavior rather than opinion. It tracked 1.46 million AI agent tool calls to its MCP server across 12,500 organizations from September 2025 to April 2026, and interviewed 12 enterprise content leaders. Here is our read, and what it teaches any team, not just the people who run content.
Where this comes from
Usage grew 70× in eight months
Monthly calls went from 7,400 in September to 521,000 in April. Part of that is new teams arriving, but part is existing users doing more every month. The activity compounds: early movers are pulling away.
And the gap is wide. The median user made 19 operations in the period. The top 1% made 1,658. Close to 90 times more.
The striking part: the top users run the same tools and the same actions as everyone else. What differs is the organization around them.
The single most intense user in the whole dataset wasn’t an engineer, either. It was a content marketer working in a conversational AI app.
Eight weeks, once the conditions are in place
Teams that reached real operations got there in about eight weeks, following a repeatable sequence:
- Weeks 1–2, inventory. The agent reads the content and maps how it is organized before changing anything. Three in four teams already let it write at this stage.
- Weeks 3–4, experimentation. Sessions get longer and the first bulk operations show up. Every output gets checked.
- Weeks 5–6, routine. The experiment becomes a recurring workflow, and legal, compliance and leadership sign off.
- Weeks 7–8, operations. Scheduled publishing and recurring pipelines running with minimal human review.
Those teams’ first week looked like everyone else’s. The difference showed up later, wherever one of four conditions was missing: context (the agent understands how content is organized), confidence (the team trusts the output), buy-in (someone with authority says yes) and content readiness (structure clean enough for bulk work).
The funnel shows where each stalls. Of 9,472 users who start by mapping, 45% never write, 60% of those who write never build a routine, and 64% of those with a routine never reach sustained operations. Only 758 people, about 8%, make it to the end. They drive 68% of all activity.
Three things no technology fixes
In the interviews, teams that reached operations had three things in common. None of them can be bought:
- Written rules. 11 of 12 enterprises had no AI use policy. The report suggests one page, not a six-month project, answering three questions: what data can go to AI, who reviews what it produces, and which content stays human-only. Legal can’t approve what hasn’t been defined.
- A tidy house. 64% of teams that cleared every other hurdle still couldn’t automate, because their content was too messy. The agent’s first read is a diagnostic. Budgets tend to go to tools, but the bottleneck is how information is organized.
- One owner. Every successful team had one person wiring it together: part engineer, part business expert. They’re usually already inside the company, experimenting on their own.
The blocker is permission, not persuasion. Decision makers expect months of governance work. The practical version takes a week.
The volume is everyday work. The return is in projects
91% of calls are everyday work: querying, auditing, editing and publishing. The other 9% are projects: platform migrations, translation, SEO and AI-answer optimization, image generation.
Reading those percentages as importance would be a mistake. Migration is only 3% of calls, yet each project replaces weeks of engineering. Translation replaces an agency contract. The right question is what does each task replace? Everyday work keeps the operation moving; projects are where the return shows up.
One more point that applies to any business: AI doesn’t replace everything the same way. The report separates strategic content (breaking news, legal copy, brand-defining pages), where AI assists and people create, from high-volume content (product descriptions, translations, ad copy), where agents do the work and people approve.
What leading teams are already building
The report flags four patterns that don’t show in the numbers yet but came up in interviews:
- AI checking AI. A travel company producing 70,000 hotel descriptions can’t read them all, so it built a layer where AI checks data and translations before anything goes live.
- Agents in sequence. A software company runs each piece through agents for tone of voice, copy editing, compliance and answer-engine optimization, with a person reviewing only at the end.
- Writing for the answer engines. Several companies already optimize content to be read by AI assistants, not just search engines.
- Content as an asset. A media company blocks AI models from its library unless they pay for a license.
What this means in practice
The report is about content, but the lesson holds for any function: procurement, finance, customer support, legal. The model isn’t the bottleneck. What separates a pilot from production is written rules, organized information and someone who wires it all together.
- Our AI Analyst does the first read: maps processes and information, separates what stays with people from what an agent can do, and hands over the list of what needs cleaning up.
- Our Forward Deployed Engineer is the person the report calls the owner: embedded in your operation, drafting the one-page policy with your team and building workflows with human sign-off wherever a mistake is expensive.
- Side-by-side training prepares whoever will own that role inside the company afterward.
Content, translation and migration are just examples. If a task repeats, it can be automated, and eight weeks is a sensible horizon to get out of test mode.
AI can write from day one. The eight weeks are how long it takes the company to get out of its way.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.