How to Hire AI Engineers When Even Google Can’t Keep Them

Last updated: August 5, 2026

By Mike Carter, Director of Partnership Success, KORE1

Hiring AI engineers in 2026 means competing for a national pool of roughly 40,300 computer and information research scientists, so build the plan around an eighteen-month productive window instead of a four-year tenure. That reframe changes what you screen for, what you pay for, and what you put in writing before the person starts. It also changes who you should be hiring in the first place, which is the part most companies get wrong.

On August 5, Google lost four of them in a morning.

Not to a competitor. To a startup Google itself is funding. Jeff Dean, who spent 27 years there and ended as chief scientist, left to co-found Discovery Loop, a public benefit corporation built to automate the experimental loop of science itself. Sanjay Ghemawat went with him. So did Oriol Vinyals, a Gemini technical lead, and Quoc Le, who co-founded Google Brain. Alphabet is a founding investor and the cloud partner. Khosla Ventures and Radical Ventures are in too. Friendly exit. Funded by the employer.

Same day, Demis Hassabis stepped back from running Google DeepMind day to day to become its chair and Alphabet’s chief scientist, with Koray Kavukcuoglu taking over operations as SVP. CNBC covered the whole reshuffle as one story, which it is.

I run growth and partnerships at KORE1, we staff AI and machine learning engineering roles for a fee, and we get paid when somebody gets hired, so read the rest of this knowing there is an invoice at the end of the funnel. I am going to spend a section arguing that a large share of you should not hire an AI engineer at all, and another one arguing that the resume credential you are screening hardest for is close to worthless for your use case. Both cost us money. I would rather write them down than pretend.

Google Is One of the Best in the Industry at Keeping Engineers

Retention risk in AI hiring is the probability that a technical hire leaves before the system they built reaches production stability. It is now the dominant cost in AI team building, ahead of salary, recruiting fees, and ramp time combined.

SignalFire’s engineering talent report names Apple, Google, Microsoft, and Adobe as the companies where engineers routinely stay past the three to five year mark. Google is a retention leader. Measured, benchmarked, named in the report.

And it still could not hold Jeff Dean.

The same SignalFire analysis found four-year engineering retention across the industry slid from nearly 59 percent in 2015 to just over 52 percent in 2024, while the engineer-to-manager ratio climbed from 5.87 to 7.65 over nine years. Flatter orgs, fewer people watching for the wobble, more engineers who quietly decide in month nine and hand in notice in month fourteen. SignalFire also found Anthropic, OpenAI, and Meta growing engineering teams two to three times faster than they lose them, which tells you where the pull is coming from.

So here is the uncomfortable version. If a company with Google’s compute, Google’s problems, Google’s money, and Google’s measured retention advantage cannot keep the person who wrote the infrastructure that modern AI runs on, then your retention plan is not a plan. It is a hope.

Plan for the departure. Then be pleasantly surprised.

The Pool Is 40,300 People

The Bureau of Labor Statistics tracks the closest thing the government has to this role under computer and information research scientists. The numbers are worth sitting with.

Source: U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, 2024 base year.

Forty thousand three hundred. Nationally. That is the whole pool. Every frontier lab, every hyperscaler, every hedge fund quant desk, every defense contractor, and every Series B startup with a demo is drawing from that same pool, and the four people who left Google on August 5 sat at the very top of it.

The BLS occupation is narrower than what most companies mean when they write “AI engineer,” which is the point. When your job description asks for someone who has published, trained models from scratch, and owns production inference, you have written a req against 40,300 people. When it asks for someone who can build reliable systems on top of models other people trained, the pool is an order of magnitude larger. Same business outcome in most cases. Wildly different search. Wildly different price.

Our AI and ML desk sees this show up as a scoping problem rather than a sourcing problem. The req is usually the thing that is broken.

How a Week Like This Reaches Your Req

Four people leaving Google does nothing for you directly. You were never hiring them.

What matters is the cascade underneath, and it is fairly predictable once you have watched it a few times. Discovery Loop is going to staff up. It will pull from DeepMind, from Google Research, and from the labs, because that is who those four know. Those teams then backfill. The people they hire come from the tier below, and the tier below hires from the tier below that, which is roughly where your applied AI req is sitting.

The cascade takes about two quarters to arrive.

It shows up in ways that are easy to misread. Candidates who ignored your outreach in the spring answer in the fall, not because your company got better but because the person two rungs above them left and the ladder shifted. Counter-offers get more aggressive at large employers, so your close rate on passive candidates dips before it improves. Comp expectations at the top of the market spike and then, oddly, the applied tier gets slightly easier to close, because the people who were holding out for a lab offer finally stop holding out.

One practical consequence. If you have been sitting on an AI req you keep meaning to open, open it now rather than in November. The pipeline is about to get better before it gets worse, and reqs that are already live when the movement starts are the ones that catch it.

None of that is a reason to lower your bar. It is a reason to have the bar written down before the good candidates show up, which most teams do not.

Hire for the Eighteen-Month Window

Most hiring plans are written as if the person stays four years. Comp is modeled on four-year vesting. Ramp is budgeted against a four-year payback. Knowledge transfer gets scheduled for “later,” which is a word that means never.

Try eighteen months instead.

Not because everyone leaves at eighteen months. Some stay six years and become the best hire you ever made. The reason to plan for eighteen is that every decision improves when you do, and none of them get worse. Watch what changes:

• Documentation stops being a nice-to-have and becomes an acceptance criterion on the offer letter. You write it into the first ninety days, not the exit interview.
• Two people touch the eval harness. Always two. This one rule has saved more AI projects than any tooling decision I have watched a client make.
• Vendor lock-in gets priced honestly. If your only Bedrock expert walks in month fifteen, what does that cost? Answer it before you sign, not after.
• You stop over-indexing on the ten-year veteran who wants a landing spot and start looking at the sharp four-year engineer who wants a hard problem, because over eighteen months the second person usually ships more.
• Contract-to-hire stops feeling like a downgrade. It is a rational instrument in a market this liquid.

One client of ours, a healthcare payments company in Costa Mesa, ran the opposite play last year. Single senior ML hire, $215K base, no second reader on any of the pipeline code, a retrieval system that only he understood. He got a call in month eleven. Not from a lab, from a Series C in Seattle. The rebuild took two engineers five months and they are still finding assumptions nobody documented. The offer was never the expensive part.

If you want the step-by-step mechanics of scoping, sourcing, and closing one of these searches, our complete AI engineer staffing guide covers the process end to end and our hiring manager’s guide to AI engineers covers the interview loop in detail. This piece is about the market conditions those processes now run inside.

What Actually Holds These People

Ask a recruiter why AI engineers leave and you will hear “comp.” Ask the engineers and you get a different list, in a different order.

Dean did not leave Google for money. He left to run thousands of parallel experiments on chip design, biology, materials science, and drug discovery, which is a thing Google could have let him do and apparently did not. Vinyals and Le went for the same reason. Ghemawat, who has spent most of his career building the systems underneath other people’s ambitions, went along with the people he trusts.

Four things hold AI engineers in my experience, and base salary is the weakest of them.

Compute access. The single most common reason a good applied AI engineer quits a mid-market company is that they spent nine months waiting on a GPU budget approval that never landed. If your infrastructure story is “we will figure it out,” you are going to lose people to companies whose story is “here is your cluster.” That is the whole conversation. It is not close.

A problem that is actually hard. Fine-tuning a support-ticket classifier is a real job. It is not a career. Engineers who came out of a research environment need at least one thread that is genuinely unsolved, even if it is 20 percent of their week, or they start reading recruiter emails in month seven.

Permission to publish or speak. Cheap. Enormously effective. Most companies never offer it because nobody asks.

Autonomy over the eval loop. If a product manager can override the quality bar, the engineer is not accountable for the system, and they know it.

Comp still matters. Obviously it matters. But we have watched candidates take $25K less to go somewhere with a real infrastructure budget more times than the reverse, and if you cannot compete on the four things above, you will end up competing on the one thing where the frontier labs cannot be beaten.

The Roles Most Companies Should Be Hiring Instead

Here is the section that costs us search fees.

A meaningful share of companies writing “AI engineer” reqs right now do not need one. Not one. Not yet. They need somebody who can take models that already exist and build something dependable around them, which is a different job with a different title and a much deeper candidate pool. We wrote about the widening split between the two markets in our analysis of the senior software engineer glut and the AI infrastructure drought, and the gap has only gotten wider since.

Three profiles worth considering before you post the req:

A platform or infrastructure engineer with AI exposure solves the deployment, latency, cost, and observability problems that kill most projects. These are the problems that actually kill them. Not model quality.

A strong backend engineer who has shipped against LLM APIs will handle retrieval, orchestration, guardrails, and evals for most enterprise use cases at a substantial discount to a research-adjacent hire. The pool here is enormous.

An applied data scientist is the right call when the actual problem is forecasting, ranking, or anomaly detection and somebody attached the word “AI” to it in a board deck.

If you are not sure which side of the line your project sits on, our breakdown of AI engineer versus ML engineer is the fastest way to sort it, and building an AI team from scratch covers sequencing if this is your first one. For teams making a genuine first hire, hiring your first AI engineer gets into what that person needs to be able to do alone.

The One Interview Exercise Worth Adding

I am not going to relitigate the whole interview loop here. The loop is covered in detail elsewhere and most of it is fine.

One exercise, though, predicts performance in a churn market better than anything else we have watched clients run. Give the candidate a working evaluation harness that is subtly wrong. Not broken. Wrong.

Build it so the test set leaks into the examples, or so the metric rewards a behavior nobody wants, or so it silently drops the 4 percent of inputs that fail to parse. Then ask them to review it the way they would review a teammate’s pull request.

Watch what happens next. You learn four things in forty minutes:

• Do they read the data, or do they read the code? The good ones open the examples first. Almost every weak candidate goes straight to the code and never looks at what is actually flowing through it.
• Whether they will say the uncomfortable thing to someone senior. This is the entire job in month fourteen when a launch is at stake and the numbers are soft.
• How they handle being partially right. Many candidates find one flaw, get satisfied, and stop. The ones you want find one, then keep going, because finding a bug is evidence there are more.
• Whether they can explain the failure to a non-technical stakeholder without either condescending or hiding behind jargon. Ask them to. Out loud.

A candidate who spots the leak in six minutes and then spends the rest of the session arguing with your metric definition is the hire. Even when they are wrong about the metric. Especially then, honestly, because that argument is the one you need them having internally when nobody is grading it.

This exercise also does something a whiteboard round cannot. It shows the candidate that your team takes evaluation seriously, which is one of the four things that actually holds these people. We have had candidates tell recruiters the eval exercise was the reason they took the offer. Nobody has ever said that about a system design round.

What You Will Pay, and Why Every Source Disagrees

Salary data for this role is a mess, and the mess is informative. Genuinely informative. Not an excuse.

BLS puts 2024 median pay for computer and information research scientists at $140,910. The offers we actually see cross our desk run from roughly $145,000 to well past $300,000 for the same nominal title, which we break down by level in our AI engineer salary guide. Neither number is wrong. They are measuring different populations, and the spread between them is the most useful thing on this page, because it maps the shape of the market rather than the middle of it.

Averages are useless here. Genuinely useless. The distribution has two humps and reporting the valley between them tells you nothing about what your specific req will cost.

Most companies reading this belong in row one and are writing reqs for row three. That single mismatch explains more failed AI searches than every other cause I can think of.

Beyond base pay, the total cost of hiring an AI engineer covers what the search itself runs, and the salary benchmark assistant will price a specific role in a specific metro. KORE1 has placed technical talent across 30-plus U.S. metros since 2005, and the regional spread on this role is wider than any other engineering title we work.

Writing the Offer for a Market That Moves

Standard four-year vesting assumes a four-year relationship. We covered why that assumption is shaky. The offer should reflect it.

Front-load the equity. A one-year cliff with heavier vesting in years one and two beats an even four-year slope for a candidate who has watched three friends change jobs in eighteen months, and it costs you very little if they stay, because you can always refresh. Refreshes are cheap. Losing the only person who understands your retrieval pipeline is not. Ask anyone who has.

Put the infrastructure commitment in writing. Not in the offer letter necessarily, but somewhere they can point at later. “Dedicated GPU budget of $X per month, reviewed quarterly” is a real sentence that has closed real candidates for us. Vague enthusiasm about “investing in AI” closes nobody, because every company says it and the candidate has heard it four times this month.

Name the hard problem. In the offer conversation, out loud, with specifics. Not “you will work on interesting challenges.” Something closer to “our claims classifier is at 84 percent, the last six points are the job, and nobody here knows how to get them.”

Then there is the thing most companies skip, which is writing down what happens if it does not work out. A ninety-day scope review, agreed in advance, where either side can say the problem was different than expected. It sounds like hedging. In practice it is the opposite, because a candidate who knows there is a structured off-ramp is far more willing to take a bet on a company that does not have a famous AI team, and the reviews themselves surface scope drift early enough to fix.

Whether you structure the engagement as a permanent hire or start it through contract staffing changes the mechanics but not the principle. Write down what the person owns. Write down what they get. Revisit it on a date you picked before day one.

When You Should Not Hire One at All

Four situations where the honest answer is no.

You have no data infrastructure. An AI engineer arriving before your data is queryable will spend eleven months doing data engineering while resenting it, then leave. Hire the data engineer first. Every time. No exceptions I have seen.

Your use case is genuinely solved by an off-the-shelf product. Some are. Buying is not a failure.

You need the capability for one quarter. That is a contract technical staffing engagement or an AI staff augmentation arrangement, and treating it as a permanent req wastes four months you do not have.

Nobody internally can evaluate the work. If no one on staff can tell a good eval harness from a bad one, you cannot interview for this role and you cannot manage it afterward. Fix that gap first, even if fixing it means bringing in a contractor to run the loop with you.

Things Hiring Managers Are Asking This Week

Is a week like this one a good time to be hiring, or a terrible one?

Good, and it is not close. Senior AI talent moving at the top of the market loosens the tier below it, which is the tier you are actually hiring from. Every founding team that leaves a large company backfills from somewhere, and the ripple reaches the applied layer within a quarter or two. Reqs opened during the churn tend to close faster than reqs opened into a quiet market.

Realistically, how long does an AI engineer search take?

Eight to fourteen weeks for a well-scoped applied role, longer if the req is written for a research profile. KORE1’s average time-to-hire across IT roles is 17 days, and AI engineering is the clearest outlier we work. The delay is rarely sourcing. It is usually a hiring team that has not agreed internally on what the person will own, which surfaces during debriefs and restarts the loop.

Does a frontier-lab name on the resume buy me anything?

Less than you think, and sometimes it works against you. Someone who spent three years with unlimited compute and a research mandate may struggle badly with a $4,000 monthly inference budget and a product manager asking about latency. Evaluate for the environment you actually have. Some of the best applied hires we have placed came out of unglamorous companies where constraints were the whole job.

What do I do when my best AI person gets the recruiter call?

Assume it already happened and that they did not tell you. Counter-offers work about a third of the time and buy roughly seven months on average in our experience. The durable move is upstream, which means fixing compute access, giving them one genuinely hard thread, and making sure a second engineer knows the system well enough that you are negotiating from choice rather than from panic.

Contract or direct hire for our first AI role?

Contract-to-hire, in most first-hire situations. You learn whether the problem is what you think it is before committing to a permanent headcount, and the engineer learns whether your data is real. Once the scope is proven, direct hire is the better instrument for anything you expect to run past a year.

Our band tops out around $180K. Are we out of the game?

No, but you are out of one specific game. You will not win a bidding war for a research profile and you should stop trying. At $180K you can hire an excellent applied AI engineer, particularly outside the Bay Area and Seattle, if the role comes with real infrastructure and a problem worth talking about at a meetup. We place in that band constantly.

How many AI engineers do we actually need?

Two, for most companies, and almost never one. A single AI hire is a single point of failure with a recruiter’s phone number. Two engineers who both understand the eval loop cost more up front and cost dramatically less the moment one of them leaves, which the retention data says will happen sooner than your plan assumes.

What is the fastest way to lose one of these people?

Ship their model without their eval gate. Nothing else comes close. It tells a technical person that quality is negotiable and that their judgment is decorative, and in a market where the next offer is a LinkedIn message away, that is usually the last thing you get to tell them.

What to Do This Quarter

Rewrite the req against the eighteen-month window. Decide honestly whether row one or row three of that table describes your problem. Put a second engineer on the eval harness, even part time, before you need to. Ask your one AI person what infrastructure they have been waiting on, then actually go get it. This week. Not next quarter.

The Google news is a useful shock precisely because it is not about you. Nobody reading this is competing with Discovery Loop for Sanjay Ghemawat. What the week proves is narrower and more useful, which is that tenure assumptions at every level of this market are softer than the org chart implies, and the companies that plan for it build teams that survive the churn instead of getting reorganized by it.

If you want a second set of eyes on a req before it goes live, or a read on whether your comp band is realistic for your metro, talk to our recruiting team. We have been placing technical talent since 2005, our recruiters average 15-plus years in their verticals, and 92 percent of the people we place are still there at twelve months. I will tell you if you do not need us.