AI made writing software cheap. That is the obvious story. More code, more prototypes, more pull requests, more “I built this in one afternoon” moments. Very impressive. Very dangerous. Very LinkedIn.
Because the interesting story starts one step later: once quantity becomes cheap, quality becomes the bottleneck.
And by quality I do not mean “does it compile?” or “did the model also generate tests that heroically confirm its own assumptions?” Those things still matter, obviously. But they are not the scarce thing anymore. The scarce thing is confidence: confidence that the change is right, that the team understands it, that the product still makes sense, and that the system will not turn into archaeological material for future developers.
It was never easier to build software. It was never harder to build good software.
AI increases output faster than teams can increase understanding. That is why code review is becoming the new bottleneck. O’Reilly’s article on agentic code review makes this visible with some wonderfully numbers: code churn up 861%, incidents-to-PR up 242.7%, defect rates rising from 9% to 54%, median review duration up 441.5%, and more PRs merged with zero review.
That is not a productivity graph. That is a smoke alarm with axis labels.
Traditional review was designed for a world where humans produced code at human speed. A developer wrote a change, another developer reviewed it, and the system was slow but at least somewhat balanced. AI breaks that balance. Now code can arrive faster than humans can deeply review it, and the bottleneck moves from “can we produce a solution?” to “can someone competent become confident this solution is right?”
Reviewing everything line by line is not the answer. That does not scale. But reviewing less while pretending nothing changed is worse. That is how you get a codebase that technically works, spiritually screams, and requires a priest before every refactor.
A lot of current AI best practices work extremely well for experienced, skilled one-person shows. And that makes sense. If you know the domain, know the architecture, understand the tradeoffs, and can smell nonsense through three layers of confident autocomplete, AI is incredible. It is like having a very fast junior developer who never sleeps, never complains, and occasionally invents a database migration because it felt emotionally right.
For a strong individual, this can be a genuine superpower. The human keeps the system model in their head. The AI produces options, drafts, tests, refactorings, explanations, and glue code. The expert filters. The loop is tight. The context is local. The quality gate is one brain. Teams are harder.
In a team, the bottleneck is not only implementation. It is shared understanding. Who knows why we chose this architecture? Who remembers the product constraint from three weeks ago? Who understands which stakeholder request was intentionally not implemented? Who can tell whether this new feature fits the domain model or just sounds plausible in a PR description?
AI makes the individual loop faster, but teams do not scale by making everyone individually faster. They scale by making understanding transferable.
That is where many current practices hit a wall. Prompting harder does not solve it. More agents do not solve it. A bigger context window helps, but it does not magically create agreement. Team-level AI needs shared context infrastructure: explicit decisions, visible assumptions, domain language, product principles, review criteria, and a way to surface misunderstandings before they become architecture.
This is why quality practices are not bureaucracy anymore. They are the scaling layer.
ADRs and Open Knowledge Format are not “nice documentation.” They are how AI usage becomes team-compatible. They turn private reasoning into shared artifacts. They let humans and agents argue over the same object. They make it possible to review not only the code, but the intent behind the code.
The one-person AI rocket is impressive. But if you want a team to fly, you need more than rockets. You need air traffic control.
We still need to look for bugs, security issues, test gaps, and architecture drift. But we also need to review product quality: is this intuitive, is it consistent, does it solve the actual user problem, does it add clarity or complexity, does it reflect taste?
This matters because AI makes stakeholder wishes dangerously affordable. In the past, some bad ideas died peacefully because they were too expensive to implement. A beautiful natural selection mechanism. Now AI can build them all. The backlog no longer has brakes; it has a rocket engine and a questionable sense of direction.
So restraint becomes a technical skill. Taste becomes part of engineering. Saying “no” becomes architecture.
Only because we can build everything users and stakeholders say they want does not mean we should. That sentence sounds obvious until every request comes with a working prototype by lunch.
For teams, this is even more important. A single expert can silently apply taste while coding. A team cannot rely on silent taste. Taste has to become discussable. Product judgment has to become reviewable. Architecture has to become visible. Otherwise every developer and every agent optimizes locally, and the product slowly becomes a collection of individually reasonable decisions that collectively require an apology. Clean Code (or even better Vertical Slices), Domain-Driven Design, Behavior-Driven Development, better tests, clearer naming, useful documentation, Architecture Decision Records (and for the js/python devs even type-systems are not a given) — all the things teams used to call important right before saying “but we do not have time” suddenly become much easier. AI can extract domain language from tickets, draft BDD scenarios, compare code against product intent, find inconsistencies, propose refactorings, summarize tradeoffs, and generate documentation that is at least good enough to start a fight.
Starting the right fight is underrated.
Used badly, AI creates more code than understanding. Used well, AI creates more understanding than code. That is the difference between acceleration and acceleration directly into a wall.
And this is where the team-scaling benefit appears. AI can help turn tacit knowledge into shared knowledge. It can turn vague intent into examples. It can turn a messy discussion into a decision record. It can turn a PR into a reviewable explanation of tradeoffs. It can turn “I thought we agreed on this” into “here is the artifact where we can see that we absolutely did not.”
That is not just documentation. That is organizational debugging.
One of the best examples is Architecture Decision Records. I used to think of ADRs mainly as documentation: useful, responsible, slightly boring, and therefore doomed to be written “later,” which in software means sometime after the heat death of the universe.
AI changes the economics. You can generate ADRs after the fact. That sounds suspicious, and yes, ideally decisions are documented when they are made. But even after-the-fact ADRs can be incredibly useful because they expose what the system appears to believe.
The generated ADR says: “We chose this architecture because…” and someone on the team immediately says: “Wait, no we didn’t.”
Perfect.
That is not a documentation failure. That is a misunderstanding becoming visible. A slightly wrong ADR can be better than no ADR because it creates a concrete object for discussion. It gives the team something to correct, challenge, refine, and align around.
Software engineering is a specification gathering exercise.
The same applies to Open Knowledge Format (OKF): decision logs, domain concepts, assumptions, risks, examples, open questions, product principles. The first version does not have to be perfect. In fact, if it is slightly imperfect, it may be better. Perfect documentation can create silence. Imperfect documentation creates useful friction, and useful friction is where shared understanding is built.
This is especially powerful in teams because the artifact becomes a meeting point between people, code, and agents. The developer can challenge it. The reviewer can use it. The product person can correct the intent. The next AI run can use it as context. The new team member can understand not only what exists, but why it exists.
That is how quality scales. Not by making every individual remember more, but by making the important things harder to lose.
The danger is not simply that AI writes bad code. Humans have been writing bad code for decades with far less impressive hardware. The real danger is that AI writes plausible code faster than teams can understand it.
That creates cognitive debt: more features than product coherence, more tests than confidence, more documentation than agreement, more code than ownership. At that point AI has not made the team faster. It has made confusion scalable.
We fell into this trap too. AI made progress feel effortless, so we produced more. Then review became heavier, assumptions became harder to track, and misunderstandings appeared in places where the code looked perfectly fine. The important part was noticing early.
The countermeasure was not to use less AI. It was to use AI differently. Not only for implementation, but for review surfaces: ADRs, explanations, domain notes, product questions, test scenarios, risk summaries, alternative interpretations. AI became less of a code generator and more of a misunderstanding detector.
A surfaced misunderstanding is a gift. A hidden misunderstanding is an incident wearing fake glasses and a mustache.
This is the team version of the AI sweet spot. Do not only ask AI to produce the next change. Ask it to make the reasoning behind the change inspectable. Ask it to expose assumptions. Ask it to compare the change against ADRs, OKF, product principles, and domain language. Ask it to help the team disagree earlier and more precisely.
The goal is not fewer conversations. The goal is better conversations with less fog.
Here is the uncomfortable part: using AI for quality often slows you down at the beginning (at least when compared to turning the slop to 11). It adds review work. It creates more discussion. It surfaces disagreements. It asks the team to think before merging. Very rude.
This can feel frustrating when AI has made implementation feel instant. But the slower feeling is not necessarily waste. Often it is the system learning. Better decisions reduce rework. Better ADRs reduce confusion. Better domain language reduces translation errors. Better product review reduces feature bloat. Better tests reduce fear.
In the long run, quality becomes productivity. Maybe even in the not-so-long run.
The trick is surviving the awkward middle, where the team is doing more thinking but has not yet received the compound interest. This is where culture matters. If the culture worships output, quality work will look like delay. If the culture values understanding, quality work looks like investment.
This is also where scaling starts to work. At first, ADRs, OKF, and richer reviews feel like extra weight. Later, they become rails. New people onboard faster. Agents produce more relevant drafts. Reviews become less repetitive. Decisions stop being rediscovered every sprint. Product discussions become less emotional because the assumptions are visible on the table instead of hiding inside everyone’s head wearing camouflage.
The early cost is real. But the alternative is paying interest on cognitive debt forever, which is less fun than it sounds, and it already sounds terrible.
AI is a multiplier, but multipliers do not care what they multiply. If your process rewards output over understanding, AI multiplies output over understanding. If your team says yes to every stakeholder request, AI multiplies product complexity. If your review process is already overloaded, AI does not magically fix it; it delivers more boxes to the burning warehouse.
But if your culture rewards clarity, taste, review, architecture, shared understanding, and useful disagreement, AI multiplies those too. That is the opportunity.
The first wave of AI software development was about producing more. The next wave is about understanding better. The winning teams will not be the ones with the most generated code. They will be the ones with the best filters: the best review systems, the clearest product judgment, the strongest shared understanding, and the discipline to say no when building would be easy but wrong.
For individuals, AI can be a rocket. For teams, quality is the guidance system. Without it, everyone moves faster in slightly different directions, which is a very efficient way to create a distributed mess.