Most enterprises rolling out AI are quietly optimizing for the wrong thing: speed, volume, lines of code shipped. Manu Narayan, CIO of GitLab, thinks that instinct is about to run companies into a productivity ceiling they can’t engineer their way out of. A faster version of a pre-AI workflow, he points out, is still a pre-AI workflow.
Narayan is GitLab’s first-ever CIO, brought in to lead enterprise technology, internal AI strategy, and data infrastructure, essentially putting GitLab to work inside GitLab. In a conversation with Talking AI host Matt Paige, he lays out what it actually takes to move an organization from incremental AI adoption to a genuine, first-principles rebuild.
There are AI haves and have-nots, and it’s not the have-nots’ fault.
Narayan describes a familiar split inside enterprises: people who’ve embraced AI and folded it into their workflows, and people who haven’t. The gap, he stresses, isn’t about effort: it’s about enablement. His job is to capture the wins happening in isolated pockets and expand them in ways that scale, which is why GitLab runs a hub-and-spoke model: a central team owns governance and technical builds, while embedded “AI transformation owners” find the repeatable, time-consuming work inside each division that’s ripe for change.
Token maxing is the wrong scoreboard.
Gamifying usage can drive early adoption, Narayan allows, but GitLab doesn’t ascribe to token maxing. There’s a direct cost at the end, and it rewards the wrong behavior. Instead the team ties measurement back to business KPIs: innovation velocity, release cadence, time to first response, time to resolution. “We’re not thinking just about things like lines of code,” he says. The goal isn’t incremental efficiency; it’s finding nonlinear, step-function change.
The human in the loop is moving to a higher level of abstraction.
A year ago, the human in the loop was a hard gate, someone reviewing every line of AI-generated code. That’s no longer possible when teams generate more code than any person could read. Narayan sees the role evolving into orchestration: agents working with other agents under an oversight agent, with a person applying taste and judgment to the output. The critical discipline is deciding, deliberately, which checkpoints still require a human. Code reviews might be fine to skim, but security scans may always warrant a person.
Context and traceability are the new differentiators.
As frontier models have improved dramatically, the ability to access context hasn’t kept pace, which makes context the real lever. Narayan’s example: technical customer support isn’t a solved problem, because so much of what an experienced engineer knows about deployment types, versions, and past failures is never documented. Where context is rich, lean on AI; where it isn’t, lean on the person. GitLab’s knowledge graph, Orbit, is his illustration of stitching context together to consume fewer tokens and get faster, more accurate output.
It’s easy to get to 90%. The last 20% is where the real work hides.
Anyone can vibe-code a proof of concept that’s 80 to 90% of the way there. But the remaining slice is where the deep problems live: role-based access controls, versioning, immutable records, the governance a public company with regulated customers actually needs. Narayan has built plenty himself; only a handful of things make it to market. That last mile, he argues, is exactly why systems of record and governance platforms don’t disappear in an AI-native world, and why the SaaSpocalypse is likely overstated.
The first move is alignment, not tooling.
For a CIO who suspects they’re optimizing the old workflow instead of rebuilding it, Narayan’s advice sounds trite but isn’t: get executive-team alignment on the scope of the change. He aligned his peers early around the fact that GitLab wasn’t chasing incremental gains but a wholesale transformation, a partnership that runs especially tight with the chief people officer as roles, AI literacy, and enablement all shift. In an era he compares to living in dog years, speed matters more than the traditional metrics, and it’s okay to pivot.
