Visibility comes before governance

The advice for enterprise AI has narrowed to a single line, and you’ve almost certainly heard some version of it. Your pilots aren’t scaling because your organization isn’t ready. The models are fine. The bottleneck is governance: the rules for who’s allowed to do what, and who owns the risk when something goes wrong. Build those first, get leadership to own them, then scale.

Real evidence sits behind that advice. When the consultancy Deloitte surveyed 3,235 technology and business leaders across 24 countries this spring, only about one in five had mature oversight for AI agents, while three in four planned to deploy them by 2027. McKinsey’s survey of more than 10,000 executives found the same shape one step later: most have put AI into production somewhere, far fewer can point to real profit from it. The reading writes itself. Companies are rich in pilots and poor in transformation, and the missing piece is the governance to coordinate hundreds of teams at once.

Then a company actually made the leap and told us how it happened. Priceline, the travel booking company, employs more than 12,000 people and works in the heavily regulated world of travel and payments. In a July 15 account for the DX Newsletter, CTO Sejal Amin and senior director Pedro Gutierrez described moving their engineering organization off a project-delivery model, where teams ship a defined chunk of work and disband, onto long-lived product teams that own something over time. The interesting part is where the change began: with measurement, before any governance mandate.

Priceline had invested in developer experience signals, meaning measurements of what shipping software actually feels like inside the org, such as how long a change waits before it moves, how many teams it has to pass through, and where the work stalls. Those signals made the trouble legible. The handoffs, the tangled dependencies, and the quiet friction between teams that everyone half-felt but nobody could point to showed up as numbers on a screen. That visibility, Amin and Gutierrez say, is what built the internal case for the reorganization that followed.

You can’t redesign the rules for problems you can’t see. In the one large enterprise with this transition on record, seeing came first: the measurement is what told the org which governance it actually needed.

Line the two stories up and they run in opposite directions. The survey playbook says governance is the binding constraint, so build governance and transformation follows. Measurement appears in that playbook, but late and small. Microsoft’s widely-cited framework for enterprise agents lists “measure usage, quality, and cost” as one capability out of six, tucked inside the governance layer as a way to tune what you’ve already built. Priceline ran it the other way. Measure first. Let the visibility expose what’s broken. Build the case for change from what the numbers show. Then redesign the rules to fix it. Measurement there sits upstream of governance and feeds it.

That reordering exposes a step the standard advice quietly skips. “Run pilots, build governance, then scale” never says what the governance gets built from. You cannot write oversight for dysfunction you can’t see. The answer to “built from what” is the part Priceline names and the surveys leave blank: built from what the measurement surfaced. Governance-first assumes the org already knows what’s broken. Priceline’s account says that assumption is the problem.

There’s a second reason this should land, which is that we already believe it somewhere else. The going wisdom about AI agents is that the teams who win are the ones who instrument quality: who measure whether the agent actually gave a useful answer, and build a durable advantage out of that measurement. It’s the same logic Priceline ran, one level up. The pattern is the same either way: you measure the thing, and what you see is what lets you move it. We’ve been glad to call measurement a moat when it’s aimed at an agent’s answers. Priceline aimed it at the organization and got the same effect.

The governance-first case isn’t wrong, and it deserves stating fairly. Governance really was what Priceline redesigned. The reorganization mattered, and nobody here is claiming the rules didn’t need changing. The correction is narrower, and it’s about order. Governance may well be the binding constraint. But a lever sits upstream of it that the standard advice doesn’t name, and in the only case on record, that lever moved first.

Now the caution, because this is one company. The public account is thin. There’s no timeline, no list of the specific structures that changed, and none of the outcome numbers you’d want before believing it worked. The claim that measurement came first is my read of a short summary, not a controlled finding. This is not a proven reversal of the governance-first advice. It’s the first look at the path in rather than the end state, and it leans clearly enough one way to be worth testing properly.

So the question worth carrying isn’t whether governance matters. It’s whether Priceline’s order is the general one. Does visibility come before governance across enterprises making this leap, or is it a quirk of one large, regulated, unusually mature org? And if the order holds, which measurements did the work: deployment frequency, or the sheer count of handoffs a change has to survive? There’s a real hazard buried in the answer. An organization that builds its governance before it can see its own dysfunction is writing rules for the wrong problem, and doing it with confidence. Instrument first, and the org will tell you what needs governing. If a second case confirms the sequence, that order stops being a footnote and becomes the first thing you do.