Should you be hiring AI engineers, or building the team you already have?

Two headlines that describe the same week

In 2026 so far, employers have cut roughly 120,000 tech roles and named AI as the most common reason, according to TechCrunch’s running tally updated on July 6. Oracle put it in a securities filing, in plain language, that AI deployment “resulted in reductions to our workforce.” That is one headline.

Here is the other, from the same market, the same month. LinkedIn has AI engineer sitting as the fastest-growing job title in the country. PwC’s 2026 Global AI Jobs Barometer, published June 15, measured the wage premium for AI skills at a level that makes a finance lead sit up. And a separate TechCrunch piece from June 24 reported that engineering roles are turning out to be among the most resilient to AI, not the most disposable.

So the same three letters are cutting jobs and creating the most expensive hiring category on the board. The founders I talk to feel both of those at once. They read that AI is coming for headcount, and they cannot fill a single AI engineering seat without a bidding war. The real question underneath the noise is not whether AI matters to the team. It is this: hire a scarce, expensive AI engineer now, or build that capability from the people already on the payroll.

How to make the hire-versus-build call before posting the job

The mistake is treating this as one yes-or-no decision. It is really five smaller ones, and running them in order keeps a founder from paying frontier-model prices for a chatbot.

1. Name the exact workflow, not the title

   Write down the specific thing that needs to exist in ninety days. "An agent that drafts and files support responses against our knowledge base," not "AI engineer." A title is a guess. A workflow is a spec, and half the specs I see do not actually require a specialist.

2. Sort the work into frontier versus applied

   Frontier work means training or fine-tuning models, novel research, hard latency or safety constraints. Applied work means wiring existing models into a product with good prompts, evaluation, and guardrails. Applied work is most of what a Series A ships, and it is buildable by strong generalists.

3. Check who on the team is one course away

   Deloitte's 2026 enterprise survey found the most common talent move is not hiring, it is teaching. A backend engineer who already owns the codebase and the customer context often closes the gap faster than an outside hire closes the context gap.

4. Price the true cost of the hire, then the true cost of the wait

   A senior AI engineer is a total-comp number well past two hundred thousand dollars, plus three to six months of recruiting in a 3-to-1 demand-to-supply market. Compare that honestly against upskilling someone who starts Monday. Sometimes the hire still wins. Now the math is visible.

5. Make the one specialist hire count

   If frontier work is real, hire one strong specialist and point them at the hard problem only they can solve, with generalists building everything around them. One senior lifting four builders beats four seniors doing applied plumbing.

The specialist hired for a problem that is not there

Here is the pattern that costs the most. A founder reads that Meta was hiring AI engineers with packages reported near a hundred million dollars, decides the war for talent is existential, and goes looking for the same caliber of person to build a feature two generalists could have shipped. Then Meta froze that same hiring and started relocating thousands of people into AI roles they did not choose. The company with the deepest pockets on earth could not turn raw talent-spend into output fast enough, and it hit the brakes.

That is the tell. At frontier labs, the constraint is genuinely a tiny pool of researchers who can push model capability. At almost every company below that tier, the constraint is not model skill. It is workflow ownership: who understands the process well enough to redesign it, instrument it, and stand behind the output. That cannot be imported with a signing bonus. It lives in the people who already know where the work breaks.

This is why the enterprise data leans so hard toward building. In Deloitte’s 2026 State of AI in the Enterprise survey of more than three thousand leaders, educating the broader workforce and upskilling existing staff both outrank hiring specialized talent as the way companies are closing the gap. The organizations with the most AI exposure are not out-hiring the shortage. They are out-teaching it.

What the AI-skills premium is actually telling us

The premium is real, and it is worth reading correctly. PwC analyzed more than a billion job ads across twenty-seven countries for its 2026 Barometer, so this is not a vibe.

Robert Half found US postings for AI, machine learning, and data science roles jumped 163% from 2024 to 2025, and ManpowerGroup’s 2026 survey of more than thirty-nine thousand employers now ranks AI model development and AI literacy as the hardest skills in the world to hire for, ahead of engineering itself. A premium that steep is not a signal to bid harder. It is a signal that supply is the bottleneck, and the cheapest supply a company controls is the talent already inside it.

There is a calmer story hiding in the layoff headlines too. CNBC reported on July 1 that a wave of employers who cut roles citing AI are quietly reversing course and rehiring humans, having discovered the automated version could not carry the work. IBM said it would triple its US entry-level hiring. That is the market correcting an over-rotation, and it is a useful warning for the opposite mistake: do not fire the people who hold the context, then pay a premium to import context that just walked out the door.

A 62% wage premium is not a reason to bid harder. It is proof that the scarce resource is capability, and the cheapest capability a company owns is the team already holding the context.

The first AI hire worth making, and the four you can skip

If I had one AI seat to fill at a Series A right now, I would hire the person who can turn a vague AI ambition into a shipped, measured workflow: strong generalist engineer, product instinct, comfortable with evaluation and guardrails, allergic to demos that cannot be measured. That hire pays for itself by making every existing engineer more useful on AI work. The four “AI specialist” reqs that usually follow are the ones to pause and re-run through the five steps first.

None of this means talent does not matter. It means the founders who win the next year are not the ones who won the bidding war. They are the ones who looked at a 62% premium, understood it as scarcity rather than status, and decided to grow their own. That is slower on paper and faster in practice, because the capability compounds inside the building instead of walking out with the next counteroffer.