AI-native companies hire for demonstrated capability, not titles: they run compact work-sample evals instead of trivia interviews, decide in 72-hour loops instead of 40-day pipelines, and slot people into small pods with real ownership. The pattern is consistent because the market forces it, strong AI engineers are gone within weeks, and credentials predict little about who ships.
The five patterns
- 1Define the capability, not the title: 'can take a RAG feature from prototype to reliable production' filters better than 'Senior ML Engineer, 5+ years'.
- 2Replace interview trivia with a paid or time-boxed work sample: a small, realistic task with your kind of data and constraints, scored against a rubric.
- 3Compress the loop: screen, work sample and decision inside 72 hours to one week. Every added week loses a measurable share of top candidates to other offers.
- 4Hire into pods: a new engineer joins a 2-4 person pod with a mission and a weekly shipping cadence, ownership from day one is both the pitch and the filter.
- 5Blend the models: permanent hires for the core, embedded engineers for speed and scarce specialties, fractional leadership until scale justifies full-time.
Evals as interviews
The strongest signal in AI hiring is watching someone do a small version of the actual job: given a rough prompt pipeline and a failing eval set, can they diagnose, fix and explain? Companies that interview this way report far fewer false positives than title-and-trivia loops, because the exercise measures the job, not interview practice. Keep it short (2-4 hours), pay for longer versions, use a written rubric, and have the pod that would work with the candidate do the scoring.
Why speed is a filter, not a shortcut
A 72-hour loop is not lowered rigor, it is concentrated rigor. One deep work sample plus one values-and-collaboration conversation carries more signal than five staggered interviews, and it signals to candidates that your company decides the way it ships. In a market where senior candidates hold multiple offers within two or three weeks, the slow process does not get to choose, it gets chosen against.
