Job postings for "AI Engineer" are everywhere, and they ask for everything: machine learning expertise, cloud infrastructure, product thinking, prompt engineering, security, deployment, monitoring. They pay like senior roles but expect junior-tier candidates to do all of it. Most positions sit open for three months and then go unfilled.
It's not a talent shortage. It's a mismatch between what companies think they need and what the market actually has to offer.
What companies think they need
Most hiring managers list what they think covers "AI work:" machine learning theory, deep learning frameworks, model training, fine-tuning, prompt engineering, LLM API integration, vector databases, RAG systems, agentic workflows, model deployment. The job posting becomes a list of every AI-adjacent technology that's been mentioned in a blog post in the last year.
Then they search for someone who knows all of it and is willing to work for junior or mid-level pay.
That person doesn't exist, or if they do, they're already working somewhere and not looking to move.
What actually needs doing
The work breaks down into much more specific specialties:
Prompt engineering and system design — someone who knows how to structure tasks for language models, manage context, test prompts empirically, and iterate. This is engineering, not ML. Most prompt engineers came from software engineering backgrounds, not research.
Integration and infrastructure — someone who wires models into product, handles API integrations, manages costs, sets up monitoring. This is software engineering with AI experience, not AI research.
Evaluation and measurement — someone who understands what it means to actually measure whether an AI system is working, designs tests that matter, monitors performance in production. This is more quality engineering and statistics than traditional ML.
Security and safety — someone who understands how to keep AI systems from breaking or doing unexpected things, who knows about prompt injection, and adversarial inputs. This is security engineering, not ML.
Performance and efficiency — someone who optimizes models for speed and cost, does quantization and model compression. This is closer to traditional ML but focused on the operational side.
Data and feedback loops — someone who manages the data that models learn from, collects feedback from production, handles retraining. This is data engineering plus domain knowledge.
Each of these is a full-time job for a skilled person. Most job postings seem to expect one person to do all of it.
What the market actually has
- Good prompt engineers who used to be customer success or DevRel and learned to work with APIs
- Good integration engineers who built APIs and now wire models into products
- Good ML researchers who don't want to do ops work and won't take a job that expects them to
- Good software engineers who have spent the last year building AI features and now understand the tooling but don't have formal ML training
- Good data engineers who understand the infrastructure side but might not know prompt optimization
What the market doesn't have a lot of:
- Unicorns who are equally good at ML research, infrastructure, security, product thinking, and communication
- Junior people who somehow understand all of AI already
- People willing to take MC+equity offers for work that used to pay $200k in big tech
How to actually staff for it
- Separate the jobs. Stop looking for one person who does everything.
- Hire for software engineering fundamentals, not AI knowledge. A good software engineer can learn to work with AI APIs in a month. A person who knows LLM APIs but can't write clean, maintainable code will create problems.
- Hire for specific roles that have real leverage. If your problem is "our prompts drift and we don't notice," hire for prompt evaluation and measurement, not "AI engineer." If your problem is "we're spending $50k a month on API calls," hire for performance and cost optimization, not "ML engineer."
- Build seniority gradually. Your first AI hire probably shouldn't be a staff engineer trying to solve everything. Hire a senior engineer to own one piece, let them build deep expertise, then hire the next person for a different specialty.
- Consider contractors and fractional roles. Not every company needs a full-time security engineer for AI. Not every company needs a full-time performance engineer. Fractional roles and contractors can cover gaps while you're building the team.
What's actually missing from the market
The real shortage is in judgment and experience running AI systems in production at scale. That's not teachable from a book or a course. It comes from having shipped features, had them break in unexpected ways, and learned to think about failure modes.
That person is expensive and rare because they have live production experience and they know what can go wrong. They're also the person most worth hiring, because they're the difference between "we shipped something with AI" and "we shipped something with AI that we can actually maintain and scale."
The hiring opportunity
If you're building an AI team right now and you hire for *specialties* instead of "AI engineer," you'll beat companies that are waiting for unicorns who don't exist. You'll move faster, spend less, and build something more sustainable.
We cover all of this in our career guides — prompt engineer, AI infrastructure engineer, agentic engineer, and others. If you're hiring for any of these roles, we can help you find the right people. See our approach at how we staff AI teams.
