For twenty-five years, hiring engineers was a search for throughput. Could this person build the thing, and how fast? The interview loop was designed around that question, the leveling ladder was calibrated to it, and compensation was priced against it. It worked because execution was genuinely scarce. Most of what a company wanted to exist did not exist because nobody had the hours or the skill to make it exist.
That constraint is gone, or at least badly loosened. Generation is cheap now. Google’s DORA program found that 90% of technology professionals now use AI at work, and more than 80% believe it has made them more productive1. Every organization I talk to has noticed this.
What almost none of them have noticed is that their defect rates did not improve.
That gap is the entire story. It is also the reason your talent strategy no longer works.
Velocity went up. Correctness didn’t.
The DORA data is unusually blunt about this. In 2024, the program measured a 1.5% decrease in delivery throughput and a 7.2% decrease in delivery stability for every 25% increase in an organization’s AI adoption2. By 2025, the throughput relationship had reversed and turned positive — teams learned to ship faster. The stability relationship did not. Higher AI adoption remains associated with higher delivery instability, and DORA’s own summary is that AI improves nearly every measured outcome except the one that determines whether the system holds together.
The time saved in creation doesn’t disappear. It gets moved downstream, into verification, and verification is the part you can’t automate away.
DORA names the mechanism directly: the time saved generating code is reallocated to auditing and verifying it. The industry has taken to calling this the verification tax. Effort saved on typing doesn’t vanish; it relocates downstream into review and validation, and most delivery pipelines were never built to absorb that volume3.
Developers feel this even when their organizations don’t measure it. Stack Overflow’s 2025 survey of more than 49,000 developers found adoption at 84%, up from 76% — while trust collapsed, with 46% saying they distrust the accuracy of AI output, up from 31% a year earlier4. The single most-cited frustration, named by 66%, was output that is almost right, but not quite5, with 45% adding that debugging generated code eats more time than it saves.
Then there is the METR randomized trial, which deserves to be cited carefully. In mid-2025, METR ran 16 experienced open-source developers through 246 real tasks and found that allowing AI increased completion time by 19%, while the same developers estimated afterward that it had made them 20% faster6. That perception gap is the interesting part. It is also fair to note that METR has since labeled the result historical and cautioned that it doesn’t necessarily describe current tools or workflows7. The tools improved. The methodological caveats were real. Take the headline number with appropriate salt.
But don’t discard the finding underneath it, because DORA’s much larger sample points in the same direction: individual effectiveness rises, system stability does not. Two independent research programs, different methods, same shape.
Velocity is easy to measure, and judgment is not, so teams optimize the thing on the dashboard.
The organizational failure mode follows from that asymmetry. Output rises, review quality falls, the defect curve stays flat, and everyone congratulates themselves on the throughput chart. The rework is real, but it lands in a different quarter and gets attributed to something else.
Most developers are not equipped for the new bottleneck. I’d put the number somewhere around nine in ten, and I say that without contempt — it isn’t a character deficiency. The industry spent a generation selecting and training for execution discipline inside a defined scope. That is a different muscle from deciding what the scope should be. The people who can hold the whole system in their head, see three moves ahead, and say no, not that, they were always rare. They were just less visible when execution was the binding constraint, because the throughput chart didn’t distinguish them.
Now they’re the only ones who matter to your product’s trajectory. And every one of them has a better outside option than they had two years ago.
The outside option repriced
Here is what changed underneath the talent market, and it has almost nothing to do with AI hiring hype.
The people who can drive the shift were always able to build things alone. What stopped them wasn’t vision or discipline. It was that a serious product requires competencies outside any one person’s depth — the frontend you don’t want to write, the infrastructure you’ve never operated, the compliance work nobody enjoys — and hiring for those competencies requires capital, which requires investors, which requires giving up control.
The constraint on the solo builder was never ambition. It was the inability to transfer a systems-building process to collaborators fast enough to matter.
That constraint has been substantially dissolved, and the data is no longer ambiguous. Carta’s cap-table data shows solo-founded companies rising from 23.7% of new startups in 2019 to 36.3% by mid-20258. Outside venture entirely, the U.S. Census Bureau counted 117,060 businesses with zero employees crossing $1 million in annual revenue in 2023 — roughly double the 2021 figure9. Recent growth in American business formation is coming almost entirely from non-employer firms rather than new employer businesses.
One in three new startups is now founded by a single person. Six years ago it was fewer than one in four. That is your retention problem, expressed as a market statistic.
Meanwhile, the inside option got worse — structurally, not cyclically.
Equity’s value depends on liquidity, and liquidity keeps receding. The median age of a startup at IPO was 13.5 years in 2024, against a 5-to-9-year median from 1980 through 200710. A four-year vest against a thirteen-year horizon is not an ownership stake. It is a lottery ticket with a stranger holding the drum.
Under private-equity ownership, the compression is more direct. Revelio Labs’ workforce analysis found that after a PE acquisition, attrition among high-paid roles rises and replacement roles are posted at lower compensation than pre-acquisition levels11, with employee sentiment toward leadership and culture declining further from an already-depressed baseline.
Honesty requires the counterweight: Josh Lerner’s work at Harvard, tracking outcomes worker by worker rather than firm by firm, found the picture more nuanced than the slash-and-burn caricature; employees were only about 2% less likely to be employed three years out, and those who found new jobs within that window saw wages fall by just 0.5%12. PE ownership is not automatically a catastrophe for the average worker.
But you are not trying to retain the average worker. You are trying to retain the person who can already see, in detail, what they’d build if they weren’t building yours.
Inside: a good salary and a lottery ticket on a thirteen-year horizon. Outside: full ownership of something they can now actually build, for the price of a tooling subscription.
A credible one-person software stack now runs roughly $3,000 to $12,000 a year13. You do not need to believe most of these attempts will succeed. You only need to notice that the expected-value calculation flipped, and that the people capable of running it correctly are precisely the ones you cannot afford to lose.
What has to change
The instinct is to respond with retention spending… comp bands, bonuses, a title. That treats an ownership problem as a cash problem, and it will not work, because cash is the dimension on which the outside option is weakest. You are competing on upside and agency, so you have to compete on upside and agency.
Internal venture terms that are actually terms. Not innovation time, not a hackathon, not a suggestion box. A written structure: if someone builds a thing inside the company that becomes a line of business, they hold a defined economic stake in that line. Percentage points, revenue share, phantom equity in the unit — the instrument matters less than the fact that it is documented and enforceable rather than promised.
IP policy that isn’t a trap. Most employment agreements assign everything the employee thinks about, in any domain, forever. The people you want to keep know this, and they will simply not build anything near the edges of your business while under that agreement. A clear carve-out process — with real answers, in writing, on a reasonable timeline — costs the company almost nothing and is worth a great deal to the exact person you’re trying to retain. The ones who ask are the ones you should worry about losing.
Ownership of the defect budget, not the ticket queue. If you want to select for judgment, put judgment on the scoreboard. Make someone accountable for the correctness curve of a system over time rather than the completion of items within it. Given that AI adoption’s measured cost shows up in stability rather than speed, this is the metric that now carries the information. It is also the single cheapest change available, and almost nobody makes it, because it requires admitting the velocity chart was measuring the wrong thing.
Real authority without formal headcount. Much of the highest-leverage work in an AI-augmented organization is done by people with no direct reports, whose influence is entirely a function of whether the organization routes decisions through them. If your structure only grants authority through management chains, your best technical judgment is architecturally powerless — and it will notice.
Differentiate, and accept the cost of differentiating. Not everyone should get moonshot terms. Pretending otherwise produces a diluted program that motivates no one and irritates everyone. The uncomfortable version is that a differentiated retention structure is legible to the people excluded from it. That is a real cost, and it should be paid deliberately rather than avoided by making the program meaningless.
The trap this is actually protecting against
The reason this matters more than a normal retention problem is what companies tend to do when they feel their core position eroding. They extend. They push into adjacent domains where they have no accumulated advantage, on the theory that growth has to come from somewhere.
Extension outside your domain is the single most judgment-intensive thing an organization can attempt. It is where accumulated pattern-matching fails, where the defaults are wrong, where you need people who can distinguish a real signal from a plausible one in unfamiliar territory. DORA’s framing of AI as an amplifier applies at the strategic level too: it magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones. Extension is exactly the maneuver you cannot execute without the people who left.
The compensation structures get hollowed out during the period of financial pressure. The best people leave during that period. Then the company attempts the extension, eighteen months later, with the population that stayed.
And the sequencing is cruel. The failure gets attributed to market conditions or poor strategy. It was neither. It was a staffing decision made two years earlier by people who thought they were managing a cost line.
The honest limit
Not every company needs this. If you operate a mature product in a stable domain where the job is genuine efficiency and reliability, then a workforce of strong executors is the correct workforce; moonshot terms would be a waste, and the people described here would be miserable working for you anyway. There is no shame in that, and it is probably most companies.
But you should know which one you are. The expensive mistake is being a company that needs to keep evolving, staffed and incentivized like a company that doesn’t, run by executives who believe the throughput chart.
The talent doesn’t leave with a bang. It leaves quietly, one person at a time, having already run the numbers.