A video published by IA et Stratégie | Le SamourAI starts from an appealing claim: “AI Engineer” postings have become so vague that the recruiter may not know much better than the candidate what the job actually contains.
Its description gives a striking number to support the point. Out of nearly 900 postings, only 1.4% supposedly do what the title says.
The problem is that the cited source does not say that.
It says almost the opposite.
The number comes from Alexey Grigorev's AI Engineering Field Guide, an open-source project that keeps the job postings, structured data, analysis code and much of its methodology public. In its first pass over 895 postings, 69.4% are classified as AI-first: roles working directly on AI systems, RAG, agents, models or LLM integrations.
Among postings whose title is exactly AI Engineer, 97% are classified as AI-first.
The famous 1.4% measures something else: the share of AI-first roles asking for GenAI skills and nothing else, without web, cloud, ops, ML or other engineering layers.
In other words, the number does not show that the profession is imaginary. It mostly shows that the job almost never consists of writing three LangChain prompts and going home.
The title predicts the work surprisingly well
The first study was produced in early February 2026 from 895 descriptions collected from Built In. The search looked for jobs containing the “AI Engineer” keyword around Los Angeles, New York, London, Amsterdam and Berlin.
That distinction matters: these were not 895 postings all carrying that exact title. The corpus also contains Applied AI Engineer, Software Engineer AI, AI Platform Engineer, ML Engineer and many other variants.
In that first snapshot, Grigorev classifies 621 roles as AI-first, 255 as AI-support and 16 as ML-first.
For the most explicit titles, the signal is far from random:
- AI Engineer: 97% AI-first;
- Applied AI Engineer: 88%;
- AI/ML Engineer: 95%;
- Software Engineer, AI: 91%;
- AI Product Engineer: 100% in this small sample.
Ambiguity increases mainly when the title includes platform, infrastructure, data or DevOps. There, the word AI can indeed describe platform support, data or infrastructure rather than building AI applications directly.
That is less entertaining than “nobody knows what they are hiring,” but more useful: some titles are ambiguous, while the canonical one is considerably less so.
The role is becoming more coherent over time
The Field Guide did not stop at 895 postings. The repository kept collecting monthly snapshots. Its published analysis across six scrapes, from February through June, covers 4,894 jobs.
The direction is fairly clear.
The AI-first share rises from 69.4% in the first scrape to 75.9% in June. Among 1,472 postings grouped under “ai engineer,” 92% are AI-first. The group is strongly associated with RAG and agents.
The more interesting shift is what those engineers actually do.
The “trainer” stack, PyTorch, TensorFlow, fine-tuning and CUDA, declines. The “integrator” stack, RAG, agents, model APIs and function calling, grows. Roles focused only on training become smaller while integrating AI systems into products takes more space.
The rising skills tell the same story: SQL, prompt engineering, CI/CD, LangGraph and even MCP become more common in postings. The occupation increasingly resembles software engineering specialized in AI, wrapped in data, evaluation, deployment and orchestration.
The ambiguity has not disappeared. It is turning into stack width.
The source itself has limits
One simplification should not replace another.
The Field Guide is valuable because it is auditable, but it is not a universal census of the profession.
The postings come from one platform, Built In. The early analysis covers five cities. Consecutive scrapes are independent cross-sections rather than a panel tracking the same postings over time.
Most importantly, the repository's current pipeline documents an enrichment step using GLM-4.7. Each job description is sent to the model to classify the role as AI-first, ML-first, AI-support or unknown and to extract skills, responsibilities and use cases.
The raw postings remain available, so the outputs can be inspected. That is enough to say the current pipeline is not simply a manual read-through. It does not, however, establish exactly how much of the very first February analysis was classified manually. The video's “read one by one by hand” wording is therefore stronger than the public methodology I could verify.
The classification also depends on a definition chosen by the author: a role becomes AI-first when it works directly on models, agents, RAG, prompts or AI deployment. That definition is coherent. It is not an official occupational standard handed down from some sacred committee between Python and human resources.
The harder problem is showing up at the bottom of the ladder
The other sources cited by the video become more interesting once we stop asking them whether “AI Engineer” is a real title.
Indeed finds that US software development postings have rebounded by roughly 15% since Claude Code launched in late February 2025. But 71% of the increase between May 2025 and May 2026 comes from senior roles, and 37% of the increase comes from postings mentioning AI in the title.
The market can therefore hire more AI-related technical people while making it harder to enter.
Revelio Labs finds a similar tension. Firms that have actually adopted AI have grown headcount more than non-adopters, but the growth is much stronger in senior roles than junior ones.
PwC reaches the same paradox through a different dataset. In its 2026 barometer, postings requiring AI skills grow much faster than the overall job market and carry a large wage premium. At the same time, highly AI-exposed entry-level roles are much more likely to demand skills traditionally associated with seniority, such as judgement and leadership.
The profession does not necessarily disappear. The beginner may simply be expected to arrive already experienced. The industry has apparently found a new implementation of the old “junior, five years of experience” bug.
Even the junior decline does not have one clean cause
Stanford's Canaries in the Coal Mine paper spread widely because it finds a relative employment decline among 22-25 year olds in occupations highly exposed to generative AI. Its live dashboard continues to show the strongest divergence among early-career workers.
But a June 2026 Centre for Economic Performance paper at the LSE complicates the story.
The authors analyze 243 million new hires and 407 million online job postings across four countries. They note that GenAI exposure is strongly correlated with exposure to working from home. When each factor is estimated separately, both appear to predict a decline in junior hiring. When both enter the model together, the work-from-home effect remains robust while the GenAI coefficient shrinks sharply and is often statistically indistinguishable from zero.
That does not prove AI has no labor-market effect. The authors explicitly avoid that conclusion. It shows that part of the broad “AI broke the entry ladder” story may be mixing together several transformations that arrived almost at once.
Labor markets make excellent natural experiments, except for their irritating refusal to change only one variable at a time.
In France, AI skill is already escaping the title
France's Apec adds one more useful piece.
In 2026, half of French managers and professionals surveyed say they use AI at least weekly. Two in ten companies already give weight to AI proficiency when selecting candidates, and more than half of large companies expect to give it more weight in future management hiring.
That points to another shift: AI skill is spreading into existing professions.
Not everyone therefore needs to become an “AI Engineer.” One part of the market is creating specialists in AI integration. Another part is adding AI capabilities to jobs that already existed.
The role is not empty, it is broad
After following the video's sources, the problem looks less like a crisis of job titles and more like a reorganization of technical work.
“AI Engineer” is beginning to describe something fairly recognizable: a software engineer who builds and ships features on top of existing models, with RAG, agents, data, evaluation, observability and infrastructure around them.
What remains unstable is how many separate jobs a company sometimes tries to fit inside that one person.
The title is therefore a reasonable starting point when reading a posting. The useful questions come immediately after it: what does this person actually have to ship, what part of the system do they own, and how will the company know their work functions correctly?
A recruiter may know perfectly well that they want an AI Engineer.
The problem starts when they also want a backend engineer, ML engineer, data engineer, DevOps engineer, product engineer and already-senior adult, then decide the bundle is an entry-level role.