Mind the gap. Every Tube passenger knows the warning. It may also be the best advice anyone can give about AI right now.

Most of the debate asks what AI will replace. I think a more useful question is what will it pull apart? Often, AI speeds up one side of a process and leaves the other side behind. Much of the danger, and much of the opportunity, now sit in the space between them.

The impact of these gaps was highlighted in a new a16z essay, The Exploit Was the Bottleneck produced in response to the latest cyberattacks. The key point it makes is that for twenty years, turning a software flaw into a working attack took scarce skill and time. AI has made that step cheap and fast. Fixing the flaw has not sped up at all.

The numbers are stark. According to figures in the essay, the average time from disclosure of a flaw to its exploitation fell from 63 days in 2018-19 to seven days before disclosure in 2025. Meanwhile, the median time to fully fix a known-exploited flaw has risen from 32 days to 43. In August, one operator using a few hundred AI agents compromised 395 organisations in 48 countries, about half of them schools.

The authors do not predict catastrophe. They expect a long, expensive era of outbreaks, falling hardest on those who cannot update quickly. The issue is not that attackers have become brilliant. It is that one side of the race changed speed. The other hasn’t.

Software developers saw this gap growing first. The bottleneck moves to specification and verification, as Google's whitepaper on the new software lifecycle argues. The data bears that out.

Faros telemetry from 22,000 developers links heavy AI adoption with 34% more tasks completed per developer. But median review time rose fivefold, and bugs per developer rose 54%, while the number of changes merged with no review at all grew by 31%.

Writing code was never the whole job. Now it is being written faster than anyone can read it, review it, or deploy it.

The same pattern appears in professional work. In a Mercor study, Claude Opus 5 scored 100% on four simplified accounting tasks, finishing each attempt in under ten minutes. On Handshake's much harder ATLAS-Finance banking benchmark, it passes only about one attempt in ten.

One failed attempt says it all. The agent produced a complete, polished workbook and passed all thirteen of its own checks. It was still wrong, because the checks were built on the same mistaken assumption as the work.

Producing a professional document now takes minutes. Knowing that it is right still takes an expert, and with AI-generated materials appearing so quickly, an expert's time is getting ever more scarce.

Yet another gap runs through our careers. Stanford's Digital Economy Lab reports that workers aged 22 to 25 in highly AI-exposed jobs are about 19% below their expected employment level, mainly because of reduced hiring. Experienced workers in the same occupations show no comparable decline. The researchers link this to the kind of knowledge involved: AI can increasingly reproduce the codified knowledge that newcomers bring from their studies, but not the tacit knowledge that comes with experience.

AI appears to reward people who already know what good looks like. In the failed banking assignment, Handshake notes that experienced bankers would have avoided the AI's tax error; a graduate might not have. But that judgement takes years to build, usually by doing the very work that AI now handles: drafting the first version, checking the numbers, and being corrected by someone senior. Yet firms are already rethinking the traditional pyramid, expecting AI to shrink the junior base while the need for experienced judgement at the top grows. Hiring fewer juniors saves money today but risks a shortage of experienced reviewers tomorrow. No model release shortens that apprenticeship. Closing this gap means designing new ways for people to learn, not just new ways to produce.

These four stories share a cause. AI accelerates whatever a single system can do alone at a keyboard: finding a flaw, writing code, drafting a report. It does far less for the work that depends on people, permissions and trust. That side moves at human and institutional speed: a hospital testing an update on a medical device before it reaches a patient, a law firm partner reviewing an AI-drafted contract line by line, a graduate learning their trade one correction at a time. The fast side improves with every model release. The slow side is held back by safety rules, accountability and the time it takes people to learn, often for good reason.

So, speeding up one side does not remove the bottleneck. It moves it, usually to somewhere less visible and less well funded. That may help explain a puzzle that Mercor itself points to: benchmark scores have soared, yet AI seems to have added little to US productivity growth so far.

None of this is permanent. AI can help on the slow side too, in finding fixes and testing work, and whoever narrows a gap will create real value. But the gaps will not close on their own. The fast side improves by default; the slow side improves only when someone decides to invest in it.

How do we address these growing gaps? When your organisation next celebrates an AI speed-up, ask three simple questions:

1.      What has to keep up? Which step downstream now receives more, faster?

2.      What is your slowest step? That is probably where your real risk now sits.

3.      Who owns it? If nobody does, the gap will keep widening.

Remember: Faster is not the same as better. Mind the gap.