Estimated reading time: 7 minutes
Key takeaways:
- AI is turning your best engineer into your biggest single point of failure. The engineer who ships the whole stack alone is most productive, and least likely to spread what they know.
- The friction AI removed wasn’t only waste – it was how teams learned. Asking a colleague transferred knowledge.
- Putting humans together now has to be deliberate. Pair two humans with the AI. If only one person can explain a system, that belongs on the risk register.
Imagine your strongest engineer, the one who ships more than anyone and who, with AI, now works across the whole stack: API, frontend, migrations, infrastructure. Work that used to need three people arrives as a single pull request with her name on it. If the 10x engineer was ever a real thing, it’s her. Nobody quite notices that nobody else understands what she built.
That last part should worry you. She is the most productive person on the team and its largest single point of failure. The tool that made her so capable is why her knowledge never spreads.
AI is shifting engineering from collaborative towards independent work, compressing the coordination layer that once held teams together. This article examines the other half of that shift: what happens to a team, and to the connection between the people on it, when they stop needing each other?
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
The friction we engineered away
For a decade we treated friction as waste, and we spent years designing it out. Asking the person who owned a service, pairing on a bug you couldn’t crack alone, waiting for a review… all of it was a tax.
AI is the best friction remover we’ve ever had. You no longer ask the colleague who knew; you ask the model, which answers in seconds without the social cost of interrupting anyone.
The catch is that friction was never only a cost. Asking a colleague moved knowledge between two people; pairing taught a junior how a senior thinks. A team’s culture and shared understanding were a byproduct of needing each other, and that came for free only because the work demanded it.
Remove the necessity and those effects go too, and you only notice once they’re gone. The small tasks and pairing that used to grow people are the first things AI absorbs.
This isn’t just remote work again
We’ve seen something like this before. Remote and async work pulled engineers apart, and teams adapted, so why is this different? It’s because remote work relocated the colleague, while AI works in their place.
When your teammate moved to Slack and Zoom, there was still a human on the other end, able to tell you what isn’t written down. When the model answers instead, there is no human on the other end at all, and therefore no knowledge is transferred.
Even the thinner version left marks. A study of over 61,000 Microsoft employees in Nature Human Behaviour found that fully remote work made the collaboration network more static and siloed, with fewer bridges across the company and less new information shared. That was with every colleague still one video call away.
Decades earlier, Thomas Allen found the same: communication between engineers drops sharply with distance, and messaging tools never reversed it. AI doesn’t add distance to your colleague, it removes the reason to cover it.
More like this
Where the loss shows up
The first place is the one that engineers illustrate, where knowledge concentration is already worse than most leaders assume.
One study of 133 open-source projects found 65% would be in serious trouble if one or two people walked away, and that was before a single engineer could own the frontend, backend, and infrastructure at once. The ‘AI hero’ doesn’t only write more code, they cover ground once shared between specialists, widening the silo and shrinking the ‘bus factor’ at once.
The second loss is more subtle. When the cheap way to answer a question is to ask the model, the person who held the answer never gets asked. The model knows the public internet, not your system’s odd history, like why the retry logic looks so paranoid.
As Michael Polanyi observed, we know more than we can tell, and the knowledge in someone’s head moves by working alongside them, not through a prompt. Route around them often enough and their understanding never spreads, while the team grows confident on answers never right for its context.
The most dangerous loss is the easiest to miss. A capable engineer paired with a confident model is a closed loop with nobody to catch its blind spots, because the model is fluent even when it’s wrong. That confidence rubs off, so you stop looking for the bug it introduced three files away. Human review surfaced those unknown unknowns rather than merely gatekeeping, and it’s the first thing to go when one person can ship the whole thing alone.
None of this shows up as a productivity problem at first, which is what makes it hard to see. The 2024 DORA report, surveying roughly 39,000 professionals, found that adopting AI lifts individual productivity and satisfaction, while at the team level a 25% rise in adoption was associated with an estimated 1.5% drop in throughput and 7.2% drop in stability.
The individual feels faster while the system underneath gets shakier. When Stack Overflow asked developers what they wanted from AI tools, improving collaboration came last, chosen by under 8%: we adopt these tools for what they do for us alone, and barely think about what they do for the team. Some of what’s lost never shows up in throughput or stability at all, because it lands on the people rather than the work.
The evidence here is inferential, and worth being plain about. No study has yet measured AI directly making engineers lonelier or teams more siloed. The conclusion comes from joining three findings that each stand on their own: the tools make individuals more autonomous, an earlier shift towards autonomy measurably fragmented collaboration networks, and the knowledge that binds a team needs human contact to move. That’s reason to take it seriously and watch for it on your own team, rather than wait for a single study to go looking.
Berlin • November 9 & 10, 2026
Engineering leadership has never moved this fast.
See how other leaders are keeping pace at LeadDev Berlin.
Putting the humans back
The answer is not less AI. Refusing the biggest change to how software gets built in a generation isn’t serious, and this is no plea to go back. The move is to put humans back around the AI on purpose, because the work no longer forces them together on its own.
Pairing should now mean two humans and the AI together, not a human instead of AI, so the model accelerates the work while a second person catches its blind spots and carries the context home.
Make code review a synchronous conversation more often than a comment left for later, which puts a second human in the loop and, in my experience, is often quicker than comment-and-wait anyway.
Measure the spread of understanding, because if only one person can explain a system, that belongs on the risk register, not the list of your stars.
None of this survives as a mandate, which is the hard part. At Shopify, pairing was an expectation rather than a policy, so it happened without anyone policing it. Pairing and synchronous review stick when a team believes the reasons – better code, a faster route to impact, and work that’s more enjoyable than doing it alone.
If your budgets and process are set far above you, the good news is that none of this needs permission or money. You don’t need a reorganization to pair on the next gnarly change, turn one weekly review into a conversation, or ask in a retro who else could explain the service you depend on. One default practice that puts two people back on the same problem is available to any team tomorrow.
The stakes for the people involved are real, even outside the delivery metrics. The US Surgeon General’s 2023 advisory put the health cost of weak social connection on par with smoking up to 15 cigarettes a day, and named the workplace as a setting that shapes it.
Engineering used to supply connection as a side effect of the work. It won’t anymore. Your engineers are about to become more capable than ever. Whether they also become more alone is the part you still get to decide.
