Most enterprises rolling out AI are quietly optimizing for the wrong thing: speed, volume, lines of code shipped. Manu Narayan, CIO of GitLab, argues that efficiency gains alone are about to drive companies straight into a productivity ceiling they can’t engineer their way out of. The reason is simple and uncomfortable—a faster version of a pre-AI workflow is still a pre-AI workflow. The real unlock isn’t speeding up what you already do; it’s rebuilding it from first principles.
In this episode of Talking AI, Matt Paige sits down with Manu Narayan, GitLab’s first-ever CIO, who owns the company’s internal AI strategy, enterprise technology, and data infrastructure—in effect, putting GitLab to work inside GitLab. Manu makes the case for moving beyond incremental AI adoption toward a genuine operating model for enterprise AI.
The conversation covers GitLab’s hub-and-spoke operating model and its embedded “AI transformation owners,” why the team measures adoption against business KPIs instead of token counts, how “human in the loop” is evolving into an orchestration role, and why context and traceability—not raw speed—are the new differentiators in software development.
In this episode, you’ll hear about:
- Why efficiency gains alone lead straight into a productivity ceiling
- The gap between AI “haves and have-nots” and how to close it
- GitLab’s hub-and-spoke (really hub-spoke-hub) operating model
- What an “AI transformation owner” does inside each division
- “Full stack” people: stretching roles end-to-end across a life cycle
- The difference between a skill and an agent—and why it matters
- Building an internal skill library with governance built in
- Why token maxing is the wrong scoreboard, and what to measure instead
- How human-in-the-loop shifts to a higher level of abstraction
- What “loops” mean and the move to being a manager of agents
- Why context and traceability beat commoditized speed
- Local vs. repo-side development and where guardrails belong
- Handling shadow AI with a genuine “happy path to production”
- The first move for a CIO stuck optimizing the old workflow
Key Moments
- 00:03:11 — The AI “haves and have-nots” inside every enterprise
- 00:04:30 — The hub-and-spoke operating model and “AI transformation owners”
- 00:07:00 — “Full stack” people: stretching roles across the whole life cycle
- 00:09:06 — Skills vs. agents — human-invoked versus autonomous
- 00:12:00 — The daily to-do skill that briefs Manu every morning
- 00:12:58 — Building an internal skill library with a review-and-promote pipeline
- 00:16:13 — Why GitLab doesn’t ascribe to “token maxing”
- 00:18:02 — Measuring adoption by role — beyond lines of code and MRs
- 00:24:30 — Local vs. repo side: where governance and guardrails actually live
- 00:27:39 — How “human in the loop” is evolving as agents outpace review
- 00:30:49 — What “loops” really are, and the manager-of-agents shift
- 00:33:52 — Why context and traceability are the new differentiators
- 00:37:29 — The maintainability fear and the bottleneck that moved to review
- 00:39:55 — SaaSpocalypse, agent sprawl, and the limits of MCP
- 00:42:51 — Shadow AI and the “happy path to production”
- 00:45:29 — The first move Monday morning: executive alignment on scope
- 00:47:33 — Advice to his pre-AI self: stay nimble, it’s okay to pivot
Key Links