How to Hire a Head of AI

Research vs Applied: The Defining Distinction

AI research leaders — those who have come from academic institutions, DeepMind, OpenAI, or Google Brain — are optimised for advancing the state of the art. Their success metric is novel contributions to the field. Shipping production systems within resource and time constraints is a secondary consideration at best. Hiring a research leader into a role that requires shipping AI product features in a commercially constrained environment is a mismatch that produces frustration on both sides.

Applied AI leaders are optimised for production — taking the best available models and infrastructure and deploying them in ways that create customer value at the speed the business requires. They are comfortable with pragmatic model selection, fine-tuning foundation models rather than training from scratch, and making the architectural tradeoffs that ship-fast demands. This is the profile that growth-stage companies with product shipping goals almost universally need.

How to Define the Head of AI Role

Before sourcing any candidates, answer three questions: What specific AI-powered product capabilities do you need shipped in the next 12 months? What is the existing data and infrastructure foundation the AI leader is inheriting? And what is the organisational structure — will they be managing a team, building one from scratch, or functioning as a technical leader with engineering resources borrowed from other teams?

The answers to these questions determine the seniority, the technical profile, and the managerial experience required. A Head of AI building a team from scratch needs different skills than one inheriting a 15-person ML engineering organisation. A candidate optimised for fine-tuning foundation models for specific use cases needs different evaluation criteria than one expected to build proprietary model infrastructure.

What Strong AI Candidates Evaluate in an Opportunity

The strongest applied AI candidates are evaluating three things beyond compensation: the data assets the company has (because the quality of training data determines the quality of AI products), the compute resources available (because constraints on compute directly limit what is achievable), and the organisational authority they will have to make technical decisions without interference. Companies that cannot credibly describe their data assets or that have CEOs who want to direct model architecture decisions will consistently lose the best AI leadership candidates to better-positioned competitors.