Hire the AI Support Trainer Who Owns What Your Bot Knows

What Does an AI Support Trainer Actually Do on a Tuesday?

A customer asks about your return window. The bot answers thirty days, confidently, with a link. The policy changed to fourteen days in March, and the help center article that says thirty is still the top retrieval hit. Nobody filed a bug. An AI support trainer and knowledge curator is the person whose job it is to find that, fix the source, and prove it stayed fixed.

The work has four repeating parts. Read the conversations the bot handled badly and label why: missing document, contradictory documents, a correct document that retrieval never surfaced, or a question the bot should have handed to a person. Edit or retire the source. Re-run the failing question. Write down what changed, so the next reader can tell a fix from a coincidence.

The org-chart claim deserves care, because the most detailed published version of it comes from a vendor. One support-AI company's write-up of AI-first team structure assigns a trainer and knowledge curator specifically to maintain the sources the AI draws from, adding FAQs derived from real conversations, resolving conflicting information, and updating policies when they change ². That is a supplier describing the shape it sells into, so read it as a description rather than as evidence. The corroboration comes from a different direction and is duller: Microsoft's 2025 index put AI trainer at the top of the new roles leaders said they were considering, named by 32 percent of them ¹. Neither source proves your team needs the seat. The thirty-day article does, because it is the kind of thing nobody owns until somebody is assigned to own it.

Two things are not in the job. The curator does not answer tickets all day, and does not own the model. Give the seat both and it collapses back into queue coverage within a month, which is the most common way teams lose this role after creating it.

Which Backgrounds Produce a Knowledge Curator Who Holds Up?

Tier-two support is the most reliable feeder, because those agents already know which documented answer is wrong and have been routing around it for a year. Technical writers are second. After that the list gets less obvious: reference librarians, paralegals, QA testers, teachers who write their own curriculum, and the operations person who quietly maintains the internal wiki nobody assigned them.

The trait underneath all of those backgrounds is the same. This person is uncomfortable when two documents disagree, and does something about it instead of picking one. In an interview it shows up as unprompted questions: which source wins, who approves a policy change, and what happens to the old article afterward.

The tells that separate real from performed are concrete. A real curator talks about specific documents and specific customers. A performed one talks about taxonomy, governance frameworks and prompt engineering. Ask for the last thing they deleted. People who have done this work remember deletions, because removing a wrong article is the highest-value edit available and the one that takes the most nerve.

Do not screen this role on writing samples alone. A clean sample proves someone can produce a document. It says nothing about whether they can find the one article out of four thousand that is quietly costing you refunds, or whether they will argue with a product manager about a policy page that was never updated.

Ask the AI Support Trainer How They Learned to Distrust a Confident Answer

The skill comes from practice rather than from a course. Ask what they used an assistant for last month, then ask about the time it was wrong and they believed it. Candidates who have done real work with these tools answer immediately and specifically. The ones who have only read about them describe a workflow in the abstract and never name a failure of their own.

The practice that produces a good curator has a recognizable shape. Draft with the assistant, then check the claim that matters against the actual policy document instead of against the assistant's summary of it. Ask the model to argue the opposite. Notice which questions it answers confidently and wrongly, which is a different set from the questions it answers slowly. That habit is the job, pointed at a customer-facing bot instead of at a draft.

An adjacent role is worth knowing about here. A red team engineer attacks a model to find what it will say under pressure. The curator runs a gentler version of the same work every week, on a system that is already live and already talking to paying customers.

A screen that works: hand over twenty real transcripts, four help center articles that partly contradict each other, and a note describing a product change nobody documented. Give it sixty minutes. What you are reading is the ordering. Strong candidates resolve the contradiction before they add anything new, because a fresh article stacked on top of a conflict makes retrieval worse rather than better.

Where Do You Find an AI Knowledge Curator, and What Closes One?

Look inside first. Microsoft's 2025 Work Trend Index put AI trainer at the top of the new roles leaders said they were considering, named by 32 percent of them ¹, and most teams that fill it fill it internally. Your tier-two queue, your escalation writers, and whoever gets tagged when a policy question needs a real answer are the shortlist you already have.

Outside, the venues are unglamorous and specific. The Support Driven community and its regional events, the Society for Technical Communication, knowledge management groups on LinkedIn, and the user communities around help center platforms such as Zendesk and Intercom, where the admins answering other admins' questions are auditioning in public. Feeder companies are support-heavy software firms and any business that has run a serious help center for years.

What closes this person is usually not compensation first. The role is often a promotion out of a queue, and what the candidate wants is what the queue never gave them: authority to change a document without three approvals, a named relationship with product so they hear about changes before customers do, and their edits visible in the bot's behavior within a day. Say all three out loud in the offer conversation.

What kills the offer is equally predictable. Discovering the seat still carries a ticket quota. Learning the knowledge base is owned by marketing and locked. Finding that nobody upstream is required to tell them when a policy changes. Any one of those and a good candidate leaves inside two quarters, usually for a team that got the reporting line right.

If the content touches anything under regulatory or security control, settle the sign-off early. That boundary is real, and it is often shared with an AI security governance officer rather than owned by the curator alone.

What Does an AI Support Trainer Cost, and Where Do They Work?

No published salary series covers this title, so treat any point estimate for it as invented. Price the nearest settled title and adjust for scope: the seat hires against your senior support band and your technical writing band, because it carries the ownership of both. As a proxy anchor rather than a rate for this job, one salary aggregator puts United States median total compensation for technical writer at $117,400 ³.

In practice, teams land between the senior support band and the writing band. Internal promotions out of tier-two commonly move a person one level rather than two, and that is the most frequent reason the promotion fails to hold: the responsibility grew, the title changed, and the pay did not. If the seat now decides what a customer-facing system says, price it against that fact.

Location norms are permissive. The work is asynchronous, document-shaped, reviewable from anywhere, and the tooling is browser-based, so remote is the common default and few teams argue about it. The exception is content held under regulatory or on-premise control: healthcare, finance and government support teams often require the knowledge base to stay inside a network, which puts the curator onsite or on a managed device.

One budget line teams forget. This role needs time inside the product, not only inside the documents. Reserve a standing block for the curator to use new features before customers ask about them, or the knowledge base will keep describing last quarter's behavior and the bot will keep answering from it.

See what gets scored >