Two job titles dominate AI hiring talk in 2026. The AI product manager owns what an AI feature should do and whether it does it. The forward deployed engineer (FDE) sits inside a customer's organisation and makes the AI work there. Both roles exist for good reasons. Both also attract claims that outrun the evidence, and if you are hiring, retraining a team or planning your own next move, you need to tell the two apart.
What the hiring data shows
For product managers, the strongest public data comes from Lenny Rachitsky's state of the product job market, published in March 2026 using TrueUp data. He counted over 7,300 open PM roles at tech companies worldwide, about 75 percent above the low of early 2023. He also reported demand for AI PMs and AI engineers climbing on the same steep curve, while design openings have stayed flat since early 2023. The data leans toward US tech companies, so treat it as a signal of direction for other markets.
For FDEs, the numbers come from one source. TechCrunch's July 2026 report cites executive search firm Christian & Timbers: about 17,000 FDEs in the US, of whom about 2,000 can deliver meaningful enterprise AI returns. The same firm found the share of companies planning FDE hires rose from 5 to 10 percent at the start of 2026 to 70 percent in the second quarter. A search firm benefits from a talent shortage, so read those figures as one firm's view of the market. Recruiter blogs also quote posting growth of several hundred percent. We leave those out because we could not trace them to primary data.
The AI PM skill map, sorted by evidence
Course providers and vendors such as Arize and Product School have converged on a similar skill list. Here is how each item holds up against what practitioners publish.
Substance: defining "good" as evals
The best-supported AI PM skill is turning a vague quality bar into testable criteria. Hamel Husain and Shreya Shankar's evals FAQ recommends that a domain expert or PM lead error analysis and act as the final judge of quality. They argue for binary pass or fail criteria over rating scales. Their guest post on Lenny's Newsletter frames evals as a skill for product builders. Nobody has measured whether evals are the single defining PM skill, as some recruiter posts claim. The direction holds up, and the superlative has no data behind it.
Substance: reading traces
An AI PM who cannot open a trace and follow what the agent did is managing by anecdote. Trace literacy is the working form of observability: finding the tool call that went wrong and exporting a sample to review with your designer. We describe the method in LLM error analysis is user research.
Substance: knowing which layer a problem lives in
PMs need enough fluency in context and harness engineering to tell a prompt problem from a retrieval problem or a missing approval gate. The layer decides who fixes the problem and how long the fix takes.
Substance: designing the level of autonomy
Feng, McDonald and Zhang's paper on levels of autonomy for AI agents defines five levels by the user's role: operator, collaborator, consultant, approver and observer. Their point for PMs is that autonomy is a design choice, separate from capability. An agent that can send customer emails does not have to send them unreviewed. We apply the five roles to real workflows in levels of autonomy for AI agents.
Mixed: prototyping with AI tools
PMs who can build a working prototype with Claude Code or a similar tool test ideas faster and write sharper specs. That part is sound. The claim that PMs will use these tools to replace engineers is hype. A prototype lacks the evals and harness a production agent needs, and building those is engineering work.
Hype: prompt engineering as a career
Prompting remains part of the job. It has also shrunk to the innermost of four layers, and a résumé built on it alone ages fast.
Hype: "LLM-as-judge solves evaluation"
The eval practitioners cited here agree that judges must be checked against human labels. Eugene Yan's essay on eval process makes the case in full: a judge helps after your team has looked at the data and agreed on what good means.
The FDE: substance and hype
Start with the substance. A gap exists between a model API and a deployment that pays off inside a specific company. The data is messy and the workflows are undocumented. The people who will use the agent were absent when someone designed it. An FDE is paid to close that gap on site.
Chris Taylor, CEO of Ode, gave TechCrunch a useful distinction. In his view, many FDEs can help a company roll Claude Code out to its workforce, while few can build a flagship AI feature. Those are two jobs under one title. The first is enablement and change management. The second is product work: research, design, evals and engineering on a feature that customers will touch.
The hype sits in the projections and the title. Christian & Timbers projects FDE demand rising 2,100 percent by year end. Its founder, Jeff Christian, also speculated to TechCrunch that agents may soon automate much of the work FDEs do today. When the recruiters placing a role question how long it will last, think twice before you rebuild a team around it.
What this means outside big tech
Demand looks different outside Silicon Valley. Malaysia, where Produlogi is based, gives a useful view of a mid-size market. An AWS study of Malaysian businesses, reported by Xinhua in August 2026, found that 38 percent use at least one AI tool on a regular basis. Among adopters, 57 percent source their AI from external providers. The share with a formal strategy for scaling AI stands at 19 percent.
A Xero survey of Malaysian MSMEs found 82 percent say they need more education to deploy AI with confidence, and half want advisory or consulting support.
Few mid-size firms anywhere will hire a full-time FDE, and most do not need one. They need two people. The first works the way an FDE works for a defined stretch: embedded with the team, building on the company's own data and workflows. The second sits inside the company and owns the product after the first one leaves. That inside role is the AI PM. Without it, whatever the outside partner builds decays as models, data and users change.
Skills to build in the next twelve months
If you are a PM moving into AI work:
- Run one error analysis session on real traces and write binary evals from it.
- Learn to read a trace in whatever tool your team uses, and ask for OpenTelemetry export.
- Write one spec that states the agent's autonomy level and the approval gate for each action.
- Build one working prototype yourself, then list what it would need before production.
If you are an engineer eyeing FDE work:
- Practise discovery. Interview the people who will use the agent before you build it.
- Learn evals well enough to prove the deployment works, with numbers the client's managers accept.
If you are a designer, treat trace review as a research method and bring comprehension and trust findings to it. Our post on the full stack design engineer covers the build side of that move.
Where the two roles meet
The AI PM and the FDE are converging on one responsibility: owning whether an agent works for the specific people who use it. The PM carries it from inside the company. The FDE carries it from inside the customer. Both draw on the same core of evals, trace literacy and judgement about where humans stay in the loop. Titles will keep shifting over the next few years. The skills on this list will transfer to whatever the hiring posts call the role next.
If you want a structured way to build these skills, our AI Product Management cohort covers evals, autonomy design and the AI PRD on a feature you bring. For other formats, see our training programmes. If you are planning your own move into AI product work, our 1-on-1 coaching and consultation services cover it. For a defined, FDE-style engagement on one AI feature, get in touch.