A 7-Person Marketing Team Made Feeding Its AI a Full-Time Job

Mintlify’s marketing team is 7 people, has no engineers on it, and one of those 7 carries the title Knowledge Engineer. The job is the material the company’s AI reads. That work is already happening on your team too, unassigned and after hours. It’s the likeliest reason your shared AI produces worse output than the demo did.

Emily Kramer of MKT1 published interviews with three teams on September 2: Mintlify, LangChain, and Buffer. She describes the reporting plainly, which matters later: “I chatted with 3 team leaders from 3 different companies.” The line that carries the piece comes from Simon Heaton, who runs growth marketing and data at Buffer:

“The problem is an AI workflow without context is running at maybe 60% of its potential. Your business context is what makes the output relevant and genuinely useful.”

Context here means the written record of how your company works. Positioning, pricing, which customer stories are approved, what the last launch taught you. An agent, the kind of AI that goes off and completes a job unsupervised instead of waiting to be asked the next question, reads that record before it starts. Nobody has to update it for the model to run. Somebody has to update it for the answer to be right.

Feeding a shared AI is maintenance work. Maintenance work gets an owner or it rots.

The role, and how new it is

Mintlify sells documentation hosting, so the docs bias is in the house. Lauren Volpi, its head of marketing, still thinks most marketers have the priority wrong:

“Few marketing teams think about documentation. That’s a mistake.”

Her reasoning is about who reads the pages now, not who wrote them. “Marketers don’t always own docs. Sometimes they do, sometimes they don’t. But agents prioritize docs now, and it’s the first site they review and search.” During one migration Kramer reports the team made 419 contributions to its knowledge base in 66 days.

Ethan Palm holds the Knowledge Engineer title. He maintains the internal knowledge bases and runs the automation that drafts updates for a human to approve. The title is new. In an August 2025 post on Mintlify’s own blog, Palm bylines as Technical Writer and argues there for the opposite of hands-off. “You are smarter than Claude,” Palm wrote then. “If you remember only one thing from this post, remember this: don’t publish something just because Claude suggested it.”

We covered the marketing engineer title and the job board attached to it two days ago. This is the unglamorous half of that job. Whoever holds the title spends a large share of the week on inputs, and the inputs are where the hours go.

What the numbers are, and who counted them

LangChain runs the most centralized version. Danny Lambert, its head of GTM engineering, leads a 4-person team that builds agents for the whole go-to-market org, and he’s candid about how early it is: “We’re still in the early innings of what an AI-native GTM org looks like.”

His numbers are the ones getting quoted this week, and they need a date on them. Lead-to-qualified conversion up 250%, three times the pipeline dollars, follow-up on lower-intent leads up 97%, roughly 40 hours a month back per rep, 86% weekly usage. Those figures are real and public. They’re also not new. LangChain published them itself on March 9, in a post by Vishnu Suresh and Jess Ou, measured “from December 2025 to March 2026.” They describe a sales agent at the company that builds agent software, reported by that company, six months ago.

The same caution applies to the traffic number. Kramer reports 66.36% agent share on Mintlify-powered documentation sites as of September, with Volpi expecting it to pass 90% by year end. Mintlify’s own published measurement, 790 million requests over 30 days, dated April 3, put agents at 45.3% against browsers at 45.8%. We wrote about that dataset in August. The April figure has a method attached. The September one is a number given in an interview by the company that sells the hosting.

One more disclosure belongs here rather than in a footnote. Kramer sells the MKT1 MCP, a paid library of pre-built marketing skills, and the piece names Mintlify’s marketing team as a customer of it: “Everyone on the marketing team has the MKT1 MCP connected too.” The newsletter also carries Framer, Profound, and Closing Media as sponsors. None of that makes the reporting wrong. It does mean every number came from a company the piece flatters, given to a writer who sells into the same audience.

The part that transfers

Buffer’s version costs nothing and is the easiest to copy. Skills live in a shared plugin on GitHub. The written context lives in Notion, where the marketing team already works, with a monthly nudge to doc owners to refresh what they own. One automated skill cut Buffer’s broken internal links by 87%. Hailley Griffis, the communications director, “went in identifying as non-technical and came out having shipped 50 pull requests and 6 internal tools that week” during an internal build week, third on the company’s makers list.

And Buffer still writes by hand. Every blog post, social post, and landing page. “Anything we ship to users has to be written by a human, or human in the loop to a very high degree,” Heaton says. So the 60% he’s describing measures rework. It’s the time a team burns correcting an AI that started from a stale picture of the business.

You’re not hiring a knowledge engineer at 9 people. What transfers is the ownership line, and it’s cheap. Pick the six documents your AI reads, put one name next to each, and set a recurring date to refresh them. Buffer’s version of that is a monthly reminder, not a platform.

Then go find out who last updated those six documents. If the honest answer is nobody since the tool was rolled out, the model isn’t what’s holding the output back, and another seat won’t fix it.