What we learned when we stopped optimizing for speed
When AI tools became useful enough to integrate into real engineering workflows, the conversation in most companies took a predictable shape. Productivity multipliers. Hours saved per developer. Time-to-merge reductions. The framing was almost always about going faster.
We followed that framing for a while. Then we noticed something more interesting going on, and it had very little to do with speed.
What AI actually changed for our R&D team wasn't the pace of the work. It was the scope of what we could realistically take on.
The work that lives at the bottom of every list
Every engineering team carries what we have started calling a shadow backlog. It is the list of things everyone agrees should happen, that never quite reach the top of the priority queue. End-to-end test coverage. Internal documentation. Knowledge capture from the people who built a system three years ago and have moved on. Tooling that would help the team but cannot be justified against the next customer-facing release.
This shadow backlog is not a sign of bad management. It is the mathematical result of finite engineering capacity meeting infinite legitimate demands. Every team has one, and every team manages it as best they can with the capacity that's left over.
When AI agents enter the picture, the conversation in most companies is about clearing the visible backlog faster. We started there too. What changed our approach was noticing that AI was unusually well-suited to a different category of work entirely, the one that gets postponed not by choice, but by the math of what fits in a quarter.
One project we had been postponing for years
A specific example. Across our R&D teams, years of context lived in places that were almost impossible to retrieve outside the people who had built them. In people's heads. In old Slack threads. In decisions made during meetings nobody documented.
We had tried to fix this before. Wikis, documentation sprints, structured interviews with senior engineers. Each approach made sense in theory. None of them survived the reality of an engineering team with shipping deadlines.
This time we tried it differently. We let AI do the structural work. Transcription of interviews. Extraction of key decisions and trade-offs. Cross-referencing across sources. Structuring the output into something searchable. The estimated month of focused work became a few hours of human review. What changed wasn't our speed. It was that the slow, structural part of the work could finally happen alongside everything else.
The output was richer than what manual work would have produced. Decisions linked to the context that explains them. Trade-offs visible alongside the alternatives that were considered. The team now has a shared layer of context that didn't exist before. It belongs to everyone.
The lesson
The temptation is to read this as a productivity story. AI made knowledge capture faster, therefore knowledge capture happened. That reading misses what actually changed.
The bottleneck was never engineering ability. Anyone on our team could have run the interviews and written the documents. The bottleneck was the disproportion between the time it would have taken and how visible the result would be. Senior engineering time spent on something that doesn't ship to a customer is always going to lose to whatever is shipping that week. That is not a problem you fix with motivation. It is a structural problem.
AI didn't solve it by being faster. It solved it by changing which side of the disproportion we had to fund.
This is why we think the productivity framing of AI is a distraction. The interesting question for any organisation isn't how to make existing work faster. It is what work could now exist that couldn't before. Those are usually the projects that compound for years afterwards, the ones that don't show up on a dashboard the week they ship.
For our R&D team, knowledge capture is one of those. The system we built doesn't need a customer-facing justification. It needs to exist, and now it does. Every person who joins the team inherits something that wasn't there before.
The most valuable use of AI is not what it lets us do quickly. It is what it lets us do at all.
