We opened with an ice breaker: if AI were a new hire on your team, how would your organization evaluate whether it was worth keeping? Nobody picked "we'd look at how many hours it logged," and nobody picked "we'd just let it quietly renew." The room split between two more honest answers: 36% said they'd ask managers if the team feels more productive, and 64% said they'd check if it shipped anything that actually mattered. Progress, but that second answer still isn't something you can put in a board deck.

Then we dove into the recent news that Uber burned a year's worth of AI spend in four months. Gartner projects that by 2028, AI coding costs will overtake the average developer's salary. And the one that got a real reaction in the chat: MIT found that 95% of enterprise gen AI pilots still show no measurable impact on the bottom line.

It's Not a Visibility Problem, It's an Attribution Problem

Most orgs actually have plenty of AI data. Tokens, prompts, model calls, adoption numbers, all of it exists. What's missing is the bridge from that data to an actual business outcome. We polled the room on exactly this: how does your organization currently track AI spend, and the results told the story. Nobody said they don't track it at all. 29% have aggregate totals with no team or feature-level breakdown. 64% track by team, but haven't connected that to delivery outcomes. Only 7% have real attribution down to work items, correlated to business results.

It's the early days of cloud all over again. Everyone rushed in; the bills came in bigger than expected, and the panic wasn't "we don't have data"; it was "we have no idea what any of it bought us." Same story now, just with tokens instead of compute.

Capping Spend Is the Easy Answer and the Wrong One

One company hit a $500 million AI spend problem and responded by pulling everyone's licenses. Full stop, no more AI, until someone could explain the bill. Understandable, and also a bad move. A blanket cap doesn't ask whether any of that spending was working; it just assumes none of it was.

The better move is tagging every AI call with the team, feature, and work item it's tied to. Most orgs already have the pieces for this. Your pipeline already knows which PR something came from, what story it closes, and when it shipped. You're connecting systems you already own, not building new ones. Attributable spend is defensible spend, and defensible spend is spend you can actually invest behind.

So Who's Actually Supposed to Own This?

Last question: Who owns AI governance in your engineering org today? Half the room said engineering leadership, with FinOps as a partner. A third said it's embedded in the platform team as part of delivery. Only 8% said nobody owns it, and another 8% pointed to finance or procurement without an engineering context. That's a healthier split than we expected; most rooms we talk to still default to "security will figure it out."

The Part That's Actually Hard

Tagging is the easy part. The harder part is patience. Before you can flag anomalous AI usage, you need a real baseline, 30 to 60 days minimum of watching what "normal" looks like before concluding. It's the same discipline FinOps teams built for cloud spend over the last decade, and it still doesn't feel natural to teams moving fast on AI. Without the baseline, you're not managing spend; you're reacting to whichever number scared you most that week.

The goal was never less AI. It's knowing which AI is earning its keep, so you can double down on that and stop funding the rest on faith.