Both, and the order is the whole answer: embed senior AI talent first, then train your existing team through daily contact with them. Companies that start with training, courses, certifications, an 'AI enablement week', produce enthusiasm and no shipped systems, because nobody in the room has ever taken an LLM feature to production. One embedded senior changes what everyone around them believes is normal.
The three strategies, honestly compared
| Strategy | What you get | Where it breaks |
|---|---|---|
| Training first | Vocabulary, enthusiasm, certificates | No production experience in the room; pilots stall |
| Hiring only | A capable AI silo | Dependency; product team never levels up |
| Embed senior, then train | Shipped system plus rising team capability | Requires deliberate pairing, not just seating |
Why osmosis beats curriculum
- AI engineering is craft knowledge: how to build an eval set, when to trust output, how to debug a prompt, none of it sticks from slides.
- A senior working in your codebase transfers judgment in code review, not in a workshop.
- Live stakes make lessons permanent: your team remembers the failure mode that almost shipped, not the one on a quiz.
- The senior also calibrates your hiring bar, after six months with them, your interviews get dramatically better.
How to run the sequence
- 1Embed one senior AI engineer into a real product team with a real deadline, not a lab.
- 2Pair them explicitly: every AI task gets a shadow from your existing team.
- 3Make knowledge transfer a stated deliverable, docs, evals, review patterns, not a hope.
- 4After the first shipped feature, hand the second one to your own engineers with the senior reviewing.
- 5Only then invest in formal training, now your team knows which questions matter.
