Every AI gain has a cost attached. Most dashboards only show one column.

AI takes friction out of engineering work. Some of that friction was quietly holding your organization together. Here is what it costs, and how to see it before it shows up as an outage, a turf war, or a resignation.

Luca Rossi, who writes the Refactoring newsletter, recently shared his favorite read of the week on LinkedIn: The Manager’s Path in the Age of AI by Camille Fournier, based on the talk she gave at LDX3 New York. Luca singled out one idea in particular. AI reduces how much engineers need to talk to each other, and we usually count that as pure upside. Fewer interruptions, more flow. Ask the assistant instead of bothering a colleague.

At Waydev we spend our days measuring AI adoption, impact and ROI across engineering organizations. The gains are real and we can measure them. The point of this piece is the other side of the ledger, the part most dashboards never show.

The ledger nobody is keeping

Every AI win in engineering has a matching entry that rarely gets recorded. Here are four we see most often.

AI does not respect Conway’s Law

Conway’s Law says systems end up mirroring the communication structure of the organization that builds them. Fournier makes a sharper point: for humans, Conway’s Law is also useful. When you work on a team, you know which systems are yours, which teams sit close to you, and that reaching into a distant team’s code takes a conversation first.

An agent has none of that cultural context. It takes a dependency on anything it can see, and the team that kicked off the work may never realize it did something wrong. Add the fact that code is now cheap, and you get a new kind of politics. Someone gets mildly annoyed at an existing system, builds their own in an afternoon, and now ownership is up for debate.

The cost does not show up in velocity metrics. It shows up months later as duplicated services, surprise breakages across team boundaries, and arguments about who owns what.

The trust channel goes quiet

The second effect is human. Those daily asks to a colleague can be annoying, but they are how people learn each other’s strengths, build a network inside the company, and develop the trust that makes hard conversations possible later. When the first stop for every question is an assistant, those exchanges simply stop happening. Nobody decides to stop collaborating. It just erodes.

The what without the why

In the comments under Luca’s post, Michael Burns, an IT Director in product engineering, pointed at the group that worries him most: new hires. Those small questions to a colleague are how people learn the reasoning behind a system.

Luca pushed the thought one step further. Even if the why were perfectly documented and the assistant could answer it, would we want to work on a team where talking to each other is mostly unnecessary because AI sits in the middle? That is not a tooling question. It is a question about what kind of engineering culture you are building.

Visibility is not understanding

This one matters to us directly, because we build an engineering intelligence platform. Fournier warns that AI-generated summaries of who did what are a useful starting point, but the further they span beyond one person, the less they explain. Leaders who rely on them can lose the situational awareness that comes from asking people directly.

We agree. Data should start the conversation, not replace it. A metric that tells you a team’s output doubled is a reason to go talk to that team, not a reason to skip the meeting.

How to measure what is hard to see

You cannot manage a cost you never measure. If you are tracking AI adoption and output, pair those numbers with signals that show what the speed is costing you. These are the ones we encourage engineering leaders to watch.

Spend some of the gains on people

Fournier’s advice to managers sounds counterintuitive: spend more time talking to people, not less. When one engineer can produce what used to take several, teams drift apart faster than any summary can catch. Some of the time AI saves should be reinvested in design reviews, real conversations and following up on the surprises your data surfaces. In her words, this only works if you “talk to humans more than you talk to Claude.”

AI is the biggest productivity shift engineering has seen in a generation. Take the gains. Just keep the full ledger, because the costs you do not measure are the ones that end up running your organization.

Sources and credit: Camille Fournier, The Manager’s Path in the Age of AI (September 2026), based on her LDX3 New York talk. Luca Rossi’s LinkedIn post on the article, and the discussion in its comments with Michael Burns.