A European tech consultancy — the kind that sells "agentic transformation" to banks and telecoms — just posted a vacancy for a mid-to-senior AI Engineer specializing in LLMs and agents. The company's entire product is "let AI do the knowledge work." They are hiring a human to build it.
The irony is so pure it almost doesn't need an article. But let's actually do the thing. Let's replace the AI Engineer with an AI system at exactly this kind of company, and see how far we get. The honest answer is: disturbingly far.
What the role actually is
Strip away the perfume and the job is this: take a fuzzy enterprise problem ("our underwriters take too long," "our support agents don't know the product"), wrap an LLM in enough scaffolding that it stops embarrassing itself, and ship it to production without the client noticing it was held together by eval spreadsheets and prayer.
The AI Engineer at this consultancy spends their week on four loops:
- Prototype loop. Get a demo working in days. A model, a prompt, a vector store, a tool or two, a slick notebook. This is the sales-enabling part — the thing that makes a client sign.
- Production loop. Turn the demo into something with observability, tracing, retries, guardrails, and an evaluation harness that proves the system isn't hallucinating contract clauses at a bank.
- Integration loop. Wire the agent into whatever Jurassic system the client actually runs. Auth, permissions, data residency, the compliance review that kills your clever architecture.
- Research loop. Read what came out this week, decide if it changes the stack, argue about it in team chat.
Which of these loops can an AI system run?
The prototype loop: already gone
The prototype loop is the most automated part of this job today, and it's automated by the people holding the job. The posting itself demands candidates who "use AI to improve their processes, fully embracing the technology" — consultancy-speak for "we already expect you to let the model write half your code."
The fully automated version isn't even hypothetical. The posting conveniently names the stack: frontier models from the major providers, open-source alternatives, the standard agent frameworks, an interoperability protocol for tool connections. That's not a job description; that's a shopping list for an autonomous coding agent.
You point an agent at it: "Build an underwriting-assistant demo. RAG over these sample documents, tool call to a premium calculator, memory of the conversation, a simple front-end by Friday." The agent reads the framework docs (which it has memorized better than the engineer does), writes the pipeline, handles the embeddings, wires the tool schema, debugs the version incompatibility between two libraries that hate each other — the exact task that eats a human's afternoon — and produces the demo.
Is the demo good? About as good as the demos humans produce in three days. Consultancy demos have never been about correctness. They are about plausibility, and plausibility is now a commodity.
The production loop: automatable, with a twist
Skeptics say "ah, but production is where humans earn their keep — the reliability, the observability, the eval harness." This used to be the moat. But productionization in this field has become writing more code and more tests. Tracing setup is code. Guardrails are code and prompts. The evaluation harness is a dataset, a rubric, and a loop that scores model outputs — often scored by another model, because humans certainly aren't hand-grading ten thousand responses.
The automated pipeline looks like this:
- An agent generates the eval dataset from the client's documents and past tickets.
- Another agent implements the tracing and the regression suite.
- A third agent — the judge — runs the suite on every change and files reports.
- A fourth agent reads the reports and proposes fixes to the pipeline.
Every one of those steps is a thing the current AI Engineer does, often using AI to do it. The human in this loop is increasingly the person who reviews the PR that another model wrote about a system that talks to a third model. There's a whole separate conversation to be had about what that review ritual is worth — the employment theater of AI code review — but the short version: the reviewer is becoming the last person in the room who isn't sure why they're there.
The integration loop: the real moat, briefly
Now we hit the part that doesn't automate this year. Integration is where the AI Engineer's job stops being technical and starts being anthropological. The client's systems are old, their compliance officer is scared, their data can't leave the country, and their "API" is a CSV emailed every Tuesday. No amount of model cleverness navigates a stakeholder who had a bad experience with a chatbot in 2019 and now believes all AI is a scam.
This is genuine human work: reading a room, trading architecture for politics, knowing which corner to cut so the pilot survives. An AI system can write the OAuth flow; it cannot sit in a room with a bank's security team and sense that the real blocker is one manager's fear of blame. That skill — call it enterprise emotional labor — is the actual product of a consultancy, and it's the last thing on the invoice that a model delivers.
But note the shape of what survives. The AI Engineer's future is the client-facing part of the job, currently maybe a fifth of it. The four-fifths that's building, testing, and iterating is the part the machines are already doing with the engineer as a typing assistant.
The research loop: the machines read faster
"Constantly reading research and applying it" is listed as a requirement, which is charming, because this is the most trivially automatable skill in the entire posting. A system that ingests every new paper, model release, and framework changelog daily — and re-summarizes what actually changed in the stack — beats any human who "keeps up with the field" the way a river beats a bucket. The engineer's research loop exists because humans are slow. Remove the human, remove the loop.
Where it breaks
Three things genuinely don't automate, and pretending otherwise would make this a fantasy.
First, blame allocation. When the underwriting agent confidently cites a policy that doesn't exist, someone must be answerable to the client. Enterprises don't buy systems; they buy someone to sue. A consultancy needs humans whose signature means "I checked this," even when — especially when — the checking is mostly trusting the eval harness. This is the AI Engineer's most durable function: being the entity with a passport and a bank account that can absorb liability.
Second, taste. The difference between a demo that closes a deal and one that gets a polite "interesting, we'll be in touch" is a thousand small judgment calls: which problem to solve, which model to bet on, which shortcut is safe. Models have opinions; they don't have stakes. A consultancy's edge is knowing which bets won't embarrass it in front of a bank. That judgment is trained on scar tissue, and scar tissue still requires a body.
Third, the client's need for a shaman. Enterprises in regulated industries are not ready to be told "the system built itself." They want a senior engineer in the room, nodding wisely. For a few more years, the human AI Engineer is the ritual that makes the purchase feel safe — the same way a pilot remains in the cockpit long after autopilot got better at landings.
The verdict
The role this consultancy is hiring for is roughly 20% client shamanism and liability sponge, 80% building and iterating AI systems. The 80% is being eaten by the very systems it builds, faster than almost any other engineering discipline, because its practitioners are professionally obligated to automate themselves first. They can't even resist on principle; "fully embracing the technology" is in the job description.
The posting asks for "proactive self-starters." What it will need in three years is one proactive self-starter per five clients, standing in front of a slide deck while a fleet of agents does the other four days of the week's work.
To whoever applies: the job is real, the salary is real, and the smart move is to be the person who deploys the system that makes the next hire unnecessary — because the alternative is being the next hire. You have, at most, one head start. The agents have already read the job posting.

