AI Engineering Manager Coach & Mentor. For the manager whose title means two jobs, and only one of them exists yet.

The title on your contract says AI Engineering Manager. What it meant when you took it: a hands-on manager over a team of ML engineers and data scientists, shipping NLP and computer vision into a product. That is the job Everpure is hiring for in Prague right now, and it is a real one.

Then the all-hands happened. Now every function head assumes you also own the coding agents in engineering, the chatbot in support, and the model the finance team started paying for last month. Three jobs, one salary, and a CEO who cannot tell the difference between them because nobody has drawn the line.

So we draw it. On one page: what your team builds, what you advise on, what you refuse to own. You take it to your CTO before someone else's incident becomes your performance review. Usually two sessions to write it, one more to rehearse the conversation where you hand it over.

Throughput went up. Nobody argues with that. What went up faster is the pile of pull requests waiting on the three people you trust to read them properly, and one of those three is you.

Anthropic's own study of its engineers puts the shift in one line: people now describe themselves as managers of AI agents and spend most of their day reviewing rather than writing. The consequence lands on the manager, because deciding what gets a deep review, what gets sampled, and what gets trusted is a risk call, and risk calls are your job, not the team's.

We write the review standard for your team. Three tiers, named. Which changes get a human read line by line, which get sampled, which ship on tests alone. One number your director can defend upward: rework and revert rate on agent-written changes, tracked weekly. If it stays flat while volume climbs, the agents are paying for themselves. If it climbs, you are shipping faster and fixing more, and you find out in week three instead of at the incident review.

By week 6 the standard exists as an RFC with your name on it. The queue is shorter because fewer things are waiting on you specifically.

Google cut a third of the managers running teams under three people. Amazon raised its engineer-to-manager ratio by 15 percent and hit the target early. Meta's new applied AI unit runs one manager to fifty engineers. Your company is smaller than all of them and is reading the same memos.

This is the 2027 version of the title, and it has no job postings yet. An AI engineering manager here runs three to five mostly agentic teams. The span in humans stays roughly where it was, around eight. What grows is the fleet under those humans, and the job moves from coordinating people to deciding what the fleet is allowed to do without you.

The shape under pressure is the one most EMs hold today: one human team of six to nine, run the way it was run in 2022. Nobody calls a meeting to tell you that shape stopped being funded. The org chart is redrawn around the people who said which half they wanted.

So you say it. We cost out both routes for you, in writing: wider across teams, or deeper into technical direction as a staff AI engineer with no reports. Then the conversation with your director, rehearsed, with one question in it. Over the next four quarters, is my scope going wider or deeper?

Two of your engineers hit the usage cap on their coding assistant in the second week of the month. One upgraded to the top plan on a company card. The platform team routed a batch job through the expensive model because it was the default. None of that reached you until the finance director forwarded the invoice with a question mark.

This is now a line in the engineering manager's job, whether the job description mentions it or not. The Pragmatic Engineer's 2026 survey of over 900 engineers describes EMs managing token budgets, usage caps, and justifying the cost trajectory to finance, next to the hiring plan.

What we do is short. A budget per team, not per person. A default model routing you can explain. A monthly number you send finance before they ask. It takes one session and saves you the quarterly conversation that starts with why.