About me
University
KashStartupIntel
QashioContact me
Kulu
Eric de La GoublayeFintech AI CTO
Back to the globe

AI moved the bottleneck

My point of view, from running it in production at Kulu

AI-native engineering

~5xengineering output at Kulu

AI did not replace engineers. It moved the bottleneck: writing code stopped being the slow part. The job of a CTO today is to move AI to where the bottleneck went.

AI did not replace my engineers. It moved the bottleneck, so I moved the AI to where the bottleneck went.
  • Productwhat is worth building
  • Buildwriting code
  • Review
  • Testing
  • Usersdid it work?
  • AI agentsticket or Sentry error to pull request
  • Second agentadversarial code review
  • AI-made testsend-to-end, synthetic data
  • AI on real usagesession replay and analytics
  • MeasuredFinOps and runway alerts
  • Engineerssame tool, very different output

The thesis · 1/5

The thesis

AI did not replace engineers. It moved the bottleneck. For years, writing code was the constraint. With AI agents, development stops being the slow part, so the bottleneck moves to product, distribution, review, testing and understanding users. The job of a CTO today is to move AI to where the bottleneck went.

The thesis · 2/5

Where the bottleneck went

If development is no longer the limit, product has to decide what is worth building, because we do not want to build things nobody uses. And engineering has to stay close enough to users to know we built the right thing. The hard conversation moved from "how do we build it" to "what do we build, and how do we know it worked".

The thesis · 3/5

What that means in practice

Adversarial code review: a second agent tries to break a change before any human looks. End-to-end tests generated and maintained with AI. Synthetic data for automated testing. AI over session replay and analytics to understand real usage. The full loop, from a support ticket or a Sentry error to a reviewed pull request.

The thesis · 4/5

Measured, not believed

~5xengineering output at Kulu

AI earns its place with numbers: complaint-to-fix time, output per project, and AI spend tracked with FinOps and runway alerts. At Kulu that looks like under an hour from a customer complaint to a reviewed pull request, and about 5x engineering output.

The thesis · 5/5

The honest nuance

The multiplier is uneven. Same tool, different engineer, very different output. So I size teams project by project, with the target set in parallel, and the numbers tell me the right size. I do not cut blind, and I do not promise a flat headcount cut.

See it running at Kulu
  • <1 hfrom a customer complaint to a reviewed pull request
  • ~5xengineering output
  • 1+year with an agentic platform in production

What does "AI moved the bottleneck" mean?

It means AI did not replace engineers, it shifted the constraint. Development is no longer the slow part, so the bottleneck moves to product, review, testing and understanding users, and that is where AI should go.

How do you measure the impact of AI on an engineering team?

With numbers: complaint-to-fix time, output per project, and AI spend tracked with FinOps and runway alerts. At Kulu, under an hour from complaint to reviewed pull request and about 5x output.

How do you build an AI-native engineering team in production?

Build systems around AI: adversarial second-agent code review, AI-generated end-to-end tests, synthetic test data, and a loop from ticket or Sentry error to reviewed pull request, behind guardrails no agent can skip.

Drag to turn the globe, scroll to zoom, click a place.Drag to turn the globe, pinch to zoom, tap a place.