An IT services company in the European Union is hiring a remote AI Engineer for NLP and Agentic AI. The job ad wants a Master's or PhD, hands-on LLM expertise, production RAG pipelines, and agentic systems that interact with enterprise tools. The pay grade says entry level. The job description says build the autonomous future of the entire product line.
This is the funniest role on the internet right now. Let's talk about why.
What the human actually does here
Strip away the buzzword salad and the job is this: a software company has decades of unstructured data — site reports, soil surveys, PDFs of structural calculations, handwritten inspection notes, tender documents, and regulatory standards that change every time a municipality updates a footnote. None of it fits neatly into a relational database. A human AI engineer is being hired to build pipelines that read this mess, extract the useful bits, summarize them, classify them, and answer questions about them.
The role is a bridge between research prototypes and production. That means the human takes a notebook where something worked once, under lab conditions, and tortures it into a service that survives contact with real users uploading real garbage. They write retrieval pipelines. They tune prompts. They build agents that call tools — query the database, fetch the document, summarize, return. They document the prompts so the next person knows why the model hallucinates load-bearing capacities.
How AI eats this role, course by course
The delicious irony is that an AI Engineer builds the exact systems that make AI Engineers redundant. Here is the pipeline that replaces them, built from tools that already exist.
Stage 1: The prototype phase
The human is supposed to stay up to date with the latest LLM research and translate prototypes into robust, scalable solutions. An AI coding agent does this faster. Give an autonomous coding tool the API docs for a hyperscaler's managed LLM service, point it at the company's document store, and say: build a RAG pipeline that ingests PDFs, chunks them, embeds them, and answers queries with citations. It will write the ingestion script, the chunking logic, the embedding calls, the vector store schema, and the retrieval step. It will pick a reasonable chunk size. It will write tests. The human's job was to know how to do this. The AI's job is to actually do it.
Stage 2: The prompt engineering
The JD asks for prompting strategies and documentation of model behavior for reproducibility. This is the most automatable task in the history of software. You do not need a PhD to run a prompt optimization loop. An agent can generate prompt variants, test them against a held-out set of construction documents, score the outputs, and keep the winner. The documentation is a git diff of the prompt file. The knowledge sharing is a README that says we tried 47 prompts, this one won. A human with a Master's degree used to do this manually, in a Jupyter notebook, with a furrowed brow.
Stage 3: The agentic systems
The role calls for agentic AI systems capable of reasoning, acting, and interacting with enterprise systems. The standard agentic loop is: receive a task, decide on a tool, call the tool, observe the result, decide again. This is a while-loop with an LLM call inside it. The tools — query the construction database, fetch a regulation PDF, calculate a volume — are the same tools the human engineer would have written. The agent framework is open source. The glue is Python. An autonomous coding tool writes the agent, writes the tools, writes the error handling, and writes the retry logic for when the enterprise API times out because someone in procurement didn't renew the token.
Stage 4: The production hardening
This is where the JD says the human earns their keep — bridge research prototypes into production-ready solutions. But production hardening is mostly: add logging, add retries, add rate limiting, add a fallback model, add a cache, add monitoring, add a queue. These are known patterns. An AI coding tool that has read every infrastructure repository on the internet knows them cold. It will add the retry with exponential backoff before you finish your coffee.
The catch: where the human clings on
Here is where it actually breaks, and why the company is still hiring a human.
Construction is not a forgiving domain. When an LLM hallucinates a movie recommendation, nobody dies. When an LLM hallucinates a load-bearing calculation embedded in a helpful summary of a structural report, people can die. The gap between the model summarized the document and the summary is legally and physically safe to act on is enormous.
The human clings on at the boundary between extraction and assertion. An AI pipeline can read a site report and extract that the concrete pour on Level 3 was delayed due to weather. It cannot reliably tell you whether that delay triggers a contractual penalty, because that requires reading the contract, the weather clause, the force majeure provisions, and the specific municipality's construction code — and then reasoning about the interaction between them. The pipeline gets you 80% of the way. The last 20% is where a human engineer with domain awareness looks at the output and says, that's not right, the model conflated two different delay types.
The other thing that doesn't automate: the product manager conversation. The JD says collaborate with product managers, engineers, and domain experts to identify high-impact AI-driven opportunities. This is a human reading a room, understanding what the construction side actually needs versus what they say they need, and deciding which of twelve possible AI features will survive contact with a job site. No agent does this well, because the input is not a document — it's a conversation full of implicit context, political undercurrents, and a product manager who doesn't know what an embedding is but knows the field teams will hate anything that adds a click.
The verdict
The company wants a Master's-degree-holding, hyperscaler-experienced, agentic-AI-fluent engineer at entry level to lead the AI function of a product line. They are hiring a human to build the system that will make the next human unnecessary. The job posting itself is a confession: the work is real, the tools exist, and the only missing ingredient is someone willing to wire them together for entry-level pay.
The first thing this engineer will automate is their own onboarding documentation. The second thing they will automate is the prototype-to-production pipeline. The third thing they will automate is the prompt tuning. By the time they leave, the only thing that will require a human is sitting in a room with a construction engineer and deciding what to build next.
Apply. Build the pipeline. Watch it eat your job. That's the whole point of the role — the core role of people in software development is over, and the person who builds the replacement gets to watch it happen from the inside, with a salary, for about eighteen months.
