The AI and ML skills to hire for in 2026
The skills that make AI teams effective are changing quickly. Here is where the real value is, and which requirements cost you good candidates.
The skills AI teams need most
Most AI teams now build on foundation models rather than training their own from scratch, so the skills in short supply sit around the model: building reliable LLM applications (retrieval, tool use, agents), fine-tuning and judging when it is worth it, making inference fast and affordable, and running models in production.
The established specialties still matter. Computer vision, speech, recommendation and forecasting need people who know the domain well, and data engineering underpins all of it. Review your pay bands for these skills regularly; the market moves faster than annual cycles.
Skills in demand
- 🧪 Evaluation design and model testing
- 🧩 LLM applications: retrieval, tools, agents
- 🎛️ Fine-tuning and adaptation
- ⚡ Inference optimisation and GPU efficiency
- 🏗️ MLOps: training, serving, monitoring
Full-stack ML is the default for applied roles
Applied ML roles increasingly expect one person to take a problem from data to a model in production: preparing the data, training or adapting the model, evaluating it, and serving it. Teams that hand work between separate data, modelling and deployment specialists move more slowly.
Deep specialists remain essential in research and infrastructure. A researcher who never touches serving, or an inference engineer who never trains a model, is fine. Decide which kind of role you are hiring for, write the job description to match, and test only what the role needs.
Evaluation is the scarce skill
Calling a model API is easy; knowing whether the result is good enough is hard. The candidates worth competing for can design evaluations that catch regressions, explain why a model fails on a slice of the data, and judge when a simpler approach, or no model at all, is the better answer.
Test for it directly. Ask how a candidate measured whether something worked and what they found when it did not. If your pay bands for these roles have not been reviewed recently, check them against current listings.
What to stop requiring
A PhD is the first requirement to question. Research roles may need one; most applied, LLM and MLOps roles need evidence of the skill instead, which a portfolio, shipped models or open-source work can show.
Also drop years-of-experience requirements on tools that are only a few years old, long lists of every framework you have ever used, and anything a capable engineer can learn in a week. They filter out good candidates without filtering out weak ones.
Requirements to remove from your job descriptions
- ✗ A PhD for roles that are not research
- ✗ “X+ years” of experience with new tools such as LLMs
- ✗ Every framework in your stack
- ✗ Tools that can be learned in a week
- ✗ Vague requirements like “strong communication skills”
The habits that predict success in ML
The most useful non-technical skills in ML are concrete. Good experimental habits mean tracking what was tried, keeping results reproducible, and writing up findings so others can build on them. Candidates who do this make the whole team faster.
Judgement and communication complete the list: knowing when a problem does not need ML, explaining uncertainty to people outside the team without overstating results, and making progress when the data or the goal is unclear. These are worth designing your interviews to assess.