Who the AI Engineer is in Spain
A sample of more than 1,800 profiles carrying the title in Spain. Where they come from, where they work, and why searching by title misses most of them.
In a sample of more than 1,800 AI engineers analysed in Spain, only 2.3% show up as open to work. And 85.8% of the people who declare generative-AI skills do not carry AI Engineer in their title.
Search by title and you miss 85%. And the ones who do carry it, you are not going to find by searching.
This is the third piece of the series, after what an AI Engineer actually does and what one earns in Spain. Here we switch datasets: we leave the job ads and go to the people.
The corpus is our own: more than 140,000 profiles in Spain that we have sourced for real searches. It is neither the whole market nor trying to be. Out of it comes the sample we analyse here: 1,878 profiles currently carrying AI Engineer, Ingeniero de IA or a variant in their job title. That is the base: unless we say otherwise, the percentages run over that sample, and when we give an absolute figure, read it as “of the sample analysed”.
For context, that same corpus holds 42.1K software profiles, 3.8K data engineers, 2.8K data scientists and 793 ML engineers. AI Engineer comes out 2.4 times more numerous than ML Engineer, and even so it is 4.5% of the software pool.
That proportion says more about us than about the market: these are the profiles we have sourced for real searches, and lately we search for far more AI engineers than ML engineers. In the real market there will be a good deal more ML engineers than show up here.
How many AI Engineers are there?
Classifying by title alone leaves out anyone who has not updated their profile or never applied that label to themselves. So we looked at it the other way round, by what people say they can do.
85.8%
of the people who declare generative-AI skills in Spain do not carry AI Engineer in their title.
849
carrying the AI Engineer title
5,131
carrying another titleSoftware Engineer (1,877), management and lead (818), consulting (443)…
This is how all the talent declaring AI skills splits by the title they carry:
The largest group of people building with LLMs in Spain is called Software Engineer
Different base: the 5,980 profiles declaring generative-AI skills, not the 1,878 with the title.
Software Engineer
AI Engineer
Management and lead
Other
Consulting and architecture
Data Scientist
Technical leadership
Data Engineer and analytics
ML Engineer
Research and academia
Product
The consequence for anyone searching is direct: the 1,877 Software Engineers in the first row show up in no search that filters on “AI Engineer”, and they are the largest group in the chart — twice the size of the ones who do carry the title.
The two figures coexist and measure different things: 1,878 carry the title, and 849 of them also declare some generative-AI skill. Put the other way round, more than half of the people calling themselves AI Engineer in Spain declare none. Skills are written by the candidate, so that says as much about what they fill in as about what they know.
With a stricter filter — only those literally declaring LangChain, LangGraph, LlamaIndex, RAG or LLM — you get 4,124 profiles. Of those, 734 have the title and 3,382 do not (another 8 have no title on the profile).
That changes the arithmetic if you are hiring. Even looking only at the sample analysed, the pool runs from 4,100 to 6,000 people, not a thousand-odd, and most of them will never show up in a search by title.
How much the role has grown
The role has multiplied by six since 2022
People holding an active AI Engineer job, as of 31 December each year. The last column runs to 24 September 2026, so it is an unfinished year.
2022 2022
2023 2023
2024 2024
2025 2025
Septiembre 2026 · en curso sep 26 · en curso
The chart counts anyone in the corpus who held an AI role that year, even if they no longer carry the title. If we restrict it to the 1,878 who do carry it today, 79% took it on in 2024 or later. Only 187 had it before 2023.
How much experience do they have?
A three-year-old role held by people with more than six years of experience
Two distributions over the same 1,876 profiles: total experience, and the experience they brought before touching AI.
Years of experience in total mean 8.0 · median 6.4
Years of career before the first job with an AI title median 6
87.7% already had two years or more of craft behind them when they moved into their first AI job. The 2-to-5-year band is the most populated, with 664 of the 1,876, but 6-to-10 is four points behind it: there is no typical moment of entry.
The median has 6.4 years of experience in total, and the median career before the first AI job is 6. They are two medians from two different distributions, so they do not subtract: what they say together is that the role is held by people who already had the craft, not that they have four months of AI behind them.
Where do they come from?
Let’s look at the job each of them held right before taking the title on:
Almost three in ten arrive from software; eight in a thousand, from frontend
The job held immediately before the first one with an AI title. N = 1,807 profiles in an AI role.
Software engineering
Data science and ML
Other (long tail)
Data engineering and analytics
Intern or work placement
Academia and research
Founder or CTO
Consulting and freelance
Infra, cloud and architecture
Product and project
Going down to specific job titles, which overlap and therefore do not sum: 238 came from data scientist, 172 from a generic software job, 142 from a machine learning job and 103 from data engineer.
The frontend figure — eight in a thousand — measures arrivals in 2024 and 2025. If that route is opening up now, we will see it next year and we will say so.
Of the 1,807 who held a previous job, only 18.8% (340) became an AI Engineer inside the same company they were already at. The remaining 81.2% (1,467) changed employer to do it.
Where do they work?
What happens if we look at the size of the company AI engineers are at?
They cluster in large companies more than software profiles do, but less than data profiles
Different base: the 1,287 profiles whose current employer we know along with its headcount. Split by company size, in Spain.
AI Engineer n = 1,287
Data (engineer, scientist, analyst) n = 6,585
DevOps and platform n = 3,447
Software (backend, frontend, full-stack) n = 28,925
Set against other roles: 64.4% of AI engineers work at a company of more than a thousand people, against 52.9% of software profiles, eleven points lower. But they are not the most concentrated: data profiles reach 66.7%.
The top employers for the role in Spain start with Accenture (44 across its two entities), NTT Data, Minsait and Indra (28) — Indra is Spain’s public-sector and defence contractor, Minsait its IT arm — then Factorial, AstraZeneca, Santander, IBM, Multiverse Computing, Deloitte, Avanade and NEORIS. Large IT services firms, banking, pharma and industry, with the odd product startup among them.
The mental image of the AI Engineer at a small startup does exist — 114 are at companies of under 50 people — but it is the exception. Most of the applied-AI capacity in Spain today sits inside corporations and IT services firms. Although it is true that more and more startups come to us to hire these profiles.
Where they are, in proportion: Madrid 43.6%, Barcelona 19.8%, Valencia 7.2%. Madrid holds more than double Barcelona.
What technologies do they know?
Python, SQL and cloud come ahead of everything AI-specific
Skills declared by the profiles carrying the title. It is a software role with a new layer on top, not the other way round.
Python
SQL and databases
Cloud (AWS / Azure / GCP)
JavaScript or TypeScript
Docker or Kubernetes
PyTorch / TensorFlow / Keras
RAG
Agents
LLM frameworks
NLP
Computer vision
FastAPI
MLOps
Prompting
Vector databases
MCP
Fine-tuning
The order matches what the job ads ask for almost point for point, with one exception: profiles declare SQL in 58.6% of cases and the ads only ask for it in 23.3%.
In how they describe themselves — LinkedIn’s “About” section, n=1,516 — 31.8% mention agents and 24.3% mention RAG. The same order the job ads carry, where agents show up in 73.3% and RAG in 56.7%. When we measured this in August the distance was far wider, with RAG almost absent from the self-descriptions; it has closed in a few weeks.
What shows up most in those self-descriptions is the craft side of the job. 46.6% talk about production, deployment or scalability, and 43.1% about business, customers or impact. 14.1% mention evaluation or monitoring.
What kind of work do they focus on?
The first piece explains that the title covers four different kinds of work. How do they split? We cannot replicate Grigorev’s clustering, because he groups job descriptions by skill signature and all we have is the self-declared skills on the profile. Classifying those skills by priority gives this:
Six in ten sit in the application layer; training models stops at 3%
Split of the 1,878 profiles with the title, classified by the skills they declare.
RAG / LLM app builder
LLM app with no RAG or agent signal
Agent builder
Platform / MLOps
Training or adapting models
Unclassified
Building applications on top of LLMs — adding agents, RAG and the apps with no signal of either — is 60% of the role, and platform another 10%. In the application layer the split resembles the international one, where the two equivalent clusters add up to 53.9%. In platform it does not: they have 18% and we have 10%.
The exception is training models: 2.9% here against 8.2% in the international analysis. That is almost triple, not a decimal. In August we were getting an order of magnitude, but that count only looked at English-language skill tags and LinkedIn stores them in the profile’s own language; counting the Spanish ones too narrows the distance. And our figure comes from classifying people by their skills while theirs classifies job descriptions, so it is not the same thing measured in two places either. It fits with the 2.9% who declare fine-tuning as a skill. In Spain, the people who train models under this title remain a very small minority.
The Forward-Deployed Engineer is the fastest-growing title in the international dataset. Here it already shows up in scattered ads — two of the six in the screenshot that opens the first piece carry it — but not yet as a category with weight of its own. We will give it an article of its own, because it deserves more detail.
What did they study?
Of 1,865 profiles with education data:
- 1,231 (66.0%) hold a master’s. Spanish master’s degrees run one to two years and are far more common than in the UK or the US, so read the share against that.
- 709 (38.0%) hold a master’s specifically in AI, machine learning or data science.
- 919 (49.3%) have some formal training in AI or data science, at bachelor’s or postgraduate level.
- 133 (7.1%) hold a doctorate.
- 62 (3.3%) went through a bootcamp; 74 come from vocational training.
Fields that show up in their education (one person can hold several qualifications, so these do not sum to the total): computer science 537, AI or data science as a degree in its own right 463, other engineering 152, business and economics 125, maths and statistics 112, telecommunications 102, physics 71, life sciences and humanities 46.
The universities that show up most: Universidad Politécnica de Madrid (216), Carlos III (121), Complutense (116), Universitat Politècnica de Catalunya (111), UOC (67), UNIR (61), Autónoma de Madrid (58), UPV (55), Autònoma de Barcelona (46), Sevilla (43), Granada (40). UOC and UNIR are Spain’s two large online universities, so roughly seven in a hundred studied at a distance.
Recommendations
If you are hiring, only 2.3% are openly looking. The profiles you want are not going to apply if you post the job.
Applications per ad
67
median
Well above the market median of 45, though below frontend (105).
With open to work switched on
43
of 1,878 · 2.3%
And 53.2% have been under a year in their current job, with a median of 11 months.
These are people who have just landed in a new role and are not looking at ads. Posting the vacancy does not solve it: almost none of those applications come from the people you wanted to reach. You have to go and find them.
Filtering on “AI Engineer” leaves out 85% of the people declaring GenAI skills, starting with the ones called Software Engineer. And if you are looking for someone who trains models, you are searching inside the 3% and it will take months; if what you need is someone to build a production system against your data, the pool is far larger and the profile looks like a senior backend engineer with cloud.
If your company has fewer than 100 people, you are competing for this profile against Accenture, Minsait and Indra, which is where 64.4% of those whose employer we know are sitting. Your advantage is going to be the problem and the autonomy, and you have to know how to tell that story.
If you want to get into the role, the front door barely exists: only 16 of the 476 ads carry junior, trainee or intern in the title. Almost everyone arrives by jumping from somewhere else, and the data says from where. The two routes holding almost six in ten are data (29.1%, adding data science, ML, data engineering and analytics) and software engineering with craft behind it (27.0%), and in both you arrive by adding the AI-systems layer on top. The route the roadmaps describe — start with model fundamentals and work down toward the product — is the one fewest people have taken.
If you already have the software craft, what you are missing you can demonstrate without anyone’s permission. Take a system with an LLM in it, read at least 100 real traces, label every failure by hand, and publish the taxonomy with a number before and after. It is what Hamel Husain calls error analysis, and it is what job ads are asking for when they say “evaluation”. The 100 is the starting point he recommends; you stop when you stop finding new failures, not when you hit the number.
Data and method
| What it is | Enriched sourcing corpus |
| Window | 2 May 2025 – 24 Sep 2026 |
| Base | 288,128 profiles, 142,936 in Spain |
| Sample analysed | 1,878 profiles with AI Engineer or a variant in their current title |
How we measured it and what we cannot claim
It is not a census. These are the profiles we source for real searches, so the absolute values are a floor. The real market is bigger and we do not know by how much.
LinkedIn skills are written by the candidate. They say how someone positions themselves, not what they can do.
We search rag as a word, not as a substring. As a substring that 24.3% barely moves, because leveraging hardly appears in Spanish: it is in 6.8% of the self-descriptions.
Qualifications almost never carry an end date, so we do not know how many did their AI master’s before or after 2023.
The skills pool (5,980) and the split across the four jobs are proxies. The second leaves 27.5% unclassified for want of declared skills, and the method is an approximation of Grigorev’s, not the same method. It only serves to rank magnitudes.
Applications per ad come from the other dataset, the job-ads one, and are explained in the salary piece.
Other people’s data. The international percentages are from Alexey Grigorev’s dataset: 6,964 descriptions from builtin.com across six markets, between February and August 2026 (our reading). His clusters come from a k-means over skill signatures; ours from classifying the profile’s self-declared skills by priority. The four-jobs taxonomy comes from Yarchi’s synthesis of the February-to-June cut, when the dataset held 4,894 descriptions.
This article gets updated. When there is new data it gets rewritten at this same URL, rather than published again.