AI engineer, ML engineer, product engineer, FDE: what AI job ads actually ask for in 2026

TL;DR: In September 2026 I read about 520 remote engineering job ads: 274 Polish B2B remote ads in full, plus 246 ads from foreign companies hiring remotely. The most frequent term across both sets was observability, far ahead of any AI term abroad. In the Polish ads RAG and AI agents lead the AI vocabulary, MCP is common, and coding agents are named by brand: GitHub Copilot, Claude Code, Cursor. LLM evals appear only four times, always in the best-paid roles. "ML engineer" almost always means Python model work, "AI engineer" means building products on existing models, "product engineer" signals ownership and is more common abroad, and Forward Deployed Engineer is rare but real. Only two hiring processes included a step that tests how you work with AI.

Why I read 520 job ads

I build and run my own products with coding agents every day, and over the past months I had accumulated habits that I never named: the way I load context per project, the guardrails around an autonomous agent, the evaluation harness behind a matching feature. The market clearly has names for these things, because job ads and interviews use them. So I read the ads to see whether the way I work matches what the market names, and where it does not.

The second reason was the titles. "AI engineer", "ML engineer", "AI solutions engineer", "product engineer", "forward deployed engineer": I wanted to know which of these describe different jobs and which are the same job with a different label.

How I counted, and what the numbers cannot tell you

The options were a keyword search across a large job board, a manual read of a few dozen ads, or pulling full ad text and counting terms. Keyword search overstates whatever you search for; a manual read does not scale past a sample. I went with full text and simple counting, and I read the ads behind the interesting numbers by hand.

• Polish set: 274 remote B2B ads in fullstack, frontend, backend and AI categories from a large Polish board, read in full (requirements, tasks, description).
• Foreign set: 246 engineering ads from companies outside Poland hiring remotely.
• Count: the number is how many ads mention a term at least once, not how many times it appears.
• Excluded: a small set of ads I found on another board by searching for tool names directly. They would have inflated every AI count, so they are out of the table.

Limits worth knowing before reading the table. Counting is pattern matching, so a mention is not a requirement; an ad can list a tool in its stack without anyone using it. It is one snapshot of one month. The raw data had traps of its own: one board returns a separate record per city, so a single ad could appear 17 times before deduplication, and pay came in four currencies and four billing periods. Most foreign ads do not disclose pay at all (41 of 246 had a range), so anything I say about pay comes from the Polish side only.

Which AI terms show up most in job ads?

What stood out:

• Observability is the most frequent term overall, 65 ads across both sets. Observability means the logs, metrics and traces that let you reconstruct what happened in a running system. It is also one of only three terms that appear clearly more often abroad, together with embeddings and product mindset.
• RAG and AI agents lead the Polish AI vocabulary. RAG, retrieval-augmented generation, means the model gets passages retrieved from your own data instead of answering from memory. An AI agent is a program that decides its own next steps and calls tools, as opposed to a single model call. AI agents are one of the few AI terms that are similarly common on both sides.
• MCP is a Polish phenomenon for now: 14 ads against 2. MCP is the open protocol through which an agent connects to external tools and data, and five Polish ads go further and ask about writing your own servers, not just using them.
• Coding agents are named by brand. In the Polish set GitHub Copilot leads (13 ads), with Claude Code (12) and Cursor (10) right behind. Abroad Cursor and Copilot are level. Windsurf did not appear once on either side, and neither did Kiro, Devin or Aider.
• Evals are rare and expensive. LLM evals, a fixed set of examples with expected answers plus metrics you rerun after every change, appear four times in Poland and never in the foreign set. All four were at the top of the pay range.
• Context engineering is still niche: three Polish ads. Context engineering means designing what goes into the model's context window and when: what loads always, what loads on demand, and what never gets in. Where it appears, it almost always sits next to prompt engineering as a pair.

ML engineer, AI engineer, product engineer, FDE: what is the difference?

Reading the ads behind the counts, four distinct jobs emerge, and the titles map onto them more cleanly than I expected.

ML engineer. "Machine learning" is the single most frequent AI term in the Polish set (31 ads), and it is a trap for anyone who builds products on top of models. It nearly always means a Python role around training and serving models. The adjacent vocabulary confirms it: deep learning, NLP, quantization and LoRA show up in 9 Polish ads, and the LangChain family appears in 10 Polish ads almost exclusively in Python roles.

AI engineer. Builds products on existing models: RAG, agents, tool use, MCP, guardrails, evals. Tool use, or function calling, is the pattern where the model returns a call to a function with a strict argument schema instead of prose; it underlies every agent. Among the 69 ads that named an AI coding agent as a requirement or a nice-to-have, the most common titles were Fullstack Developer (11) and Software Engineer (8), then AI Engineer (4), AI Solutions Engineer (4), Enterprise AI Architect (3) and Product Engineer (3). That 69 includes the targeted ads, so it overstates the base rate. The best-paid ads in that group described a role whose whole content was building reusable capabilities for a coding agent, and several of them looked like the same project posted by different intermediaries.

Product engineer. The title is less about a stack and more about owning a problem end to end. "Product mindset" or "ownership" appears in 7 foreign ads and 2 Polish ones, about three and a half times as often abroad. Phrases about "own product" or side projects matched 14 Polish ads, but almost all of them referred to the company's product. Exactly one ad said a shipped side project of your own was a plus.

Forward Deployed Engineer. An engineer who works directly with a customer and builds the solution in the customer's environment, instead of taking tickets from a backlog. The title came out of enterprise data software and is being picked up by companies building on LLMs. It appeared twice in the Polish set and not at all in the foreign one, so it is real but rare. Titles I specifically looked for and did not find a single time: "AI Product Engineer" and "AI Product Builder".

How are companies testing AI skills in hiring?

Seven of the ads I read closely described their hiring process. Most were the familiar shape: recruiter screen, one to three technical calls, a manager or client conversation. Two stood out because they include a separate step that tests how you work with AI:

• An AI-driven business case you present. You get a business problem, work it with AI tools, and later present the result and discuss it. It is a demonstration of how you drive an agent, not an algorithm puzzle.
• An AI testing skills round, placed right after the HR screen, before any manager interview.

One process went the other way and stated explicitly that no AI is used to screen candidates, run interviews or communicate with them. That is a useful signal in itself.

Where my own work matched, and where it did not

Going term by term against my own setup was the useful part. Several things I do daily had names I was not using: an agent harness (the permissions, hooks and blocks an agent runs inside; I wrote about the guardrails I run), context engineering (what I had been calling per-project context loading, described here), spec-driven development (a written spec in the repo that the agent executes and that gets reviewed), tool use, and multi-agent orchestration.

The strongest match was the rarest term. The matching feature in a product I run has an evaluation harness: a golden set built from admin decisions and user feedback, a pool for manual labelling, and information retrieval metrics such as precision at 10, NDCG, MRR and recall, versioned so pipeline changes can be compared. Some of the threshold work behind it is in tuning pgvector thresholds. Four ads asked for exactly this and I had never called it an eval harness.

The gaps were just as clear. I do not use LangChain or its relatives; orchestration in my work happens at the agent and MCP level. And on observability, the most frequent term of all, my products had product analytics and platform logs, but not the engineering side: structured logs with a request ID, error tracking, tracing. That is the one I would close first.

One more finding that cuts against the hype: in the Polish set, the best-paid frontend ads had no AI requirement at all. Roles defined around coding agents paid well, and pure React roles in finance paid more.

What this means for engineers and for hiring managers

For engineers:

• Name what you already do. Agent harness, context engineering, spec-driven development, tool use, evals. The vocabulary is how your experience gets matched to an ad.
• Treat observability as table stakes, especially for foreign companies. It outnumbers every AI term.
• Evals are the differentiator. Rare in ads, always at the top of the range, and hard to fake in an interview.
• Do not confuse ML roles with AI product roles. If you build on models rather than train them, "machine learning" in a title usually means a different job.
• Be ready to drive an agent live. Two processes already test it, and the format is a case you present.

For hiring managers:

• "AI-assisted development" says very little. It is the most common phrase and the vaguest. Name the actual work: building MCP servers, running evals, designing context.
• Test the skill in the process, the way the two processes above do, instead of asking about it.
• Say "ownership" if you mean it. Product engineers read for it, and most Polish ads leave it out.
• Publish a range. Five out of six foreign ads in my set did not, which makes them hard to compare with anything.