01
Leadership talk

Engineers transformed.
Engineering didn't.

Individual engineers write code far faster. No organisational number moved that anyone could attribute to it. This is about the gap between those two sentences.

Sangeethkumar · engineering leader, AI platform

02
A fair question

"Is engineering actually transforming with AI — or just using it?"

It deserves a real answer, and the honest one is uncomfortable: a great deal changed for individuals, and very little changed for the organisation.

Both halves of that are true at once, and holding them together is the entire leadership problem.

03
Three tests

I applied the same tests I'd apply to any initiative

01Is there a system?Something a new engineer inherits — not a habit that leaves when the person does.
02Did a process get retired?Real change removes work. If nothing was retired, we added a tool on top of the old shape.
03Did a number move?Not activity. Outcome. Cycle time, escaped defects, throughput.

We failed all three — while the individual gains were completely real.

04
The gains were real

Where it worked

Automated review adoption went from a few per cent to most changes in two months — on one codebase.

First review feedback in minutes, where waiting for a human took most of a day.

Code written roughly three times faster.

Where it didn't

The same tool, on another codebase, stayed near zero.

The difference was not the tool. On one codebase it ran on its own; on the other it was triggered by hand, with each person's own rules and context.

05
The pattern

The gain leaks out before it reaches delivery.

An engineer writes code three times faster. Then it is validated line by line and waits for review. Everything downstream of writing got harder.

Speeding up a step that wasn't the constraint produces a happier engineer and an unchanged organisation.

06
The reframe

Experiment vs. promoted default

An experiment

One person, one workflow, no owner, no measurement, no obligation on anyone else. Valuable — and it stays where it started.

A promoted default

Owned. Measured. Documented. On by default. Something a new engineer inherits without being told.

Every engineer had individual practices, and none of them were systematised. That is not an adoption problem. It's a distribution problem — and distribution was the thing we did worst.

07
Four capabilities

What decides whether individual wins become organisational ones

01ContextAssembling what a system must see — module, history, architecture — instead of carrying it in one person's head.
02HarnessThe rules, tools and guardrails agents run inside. When something fails once, it is engineered so it never fails that way again.
03EvaluationProving output is right without redoing it by hand. Today this is the gap that turns every speed gain back into review load.
04DistributionHow one person's win becomes the default. The one we do worst, and it costs the most.
08
The operating rule

Agents propose. Evidence gates. Humans promote.

The same sentence governs the technical systems and the organisational ones, which is why I trust it. An experiment becomes a default when someone owns it, evidence supports it, and a person decides to promote it — not when it goes well for one engineer.

09
Leading through it

The part that isn't about AI

This ran through a period of turnover on the team.

Transformation programmes are run by the same people absorbing everything else. Any plan that assumes otherwise is fiction.

10
Measuring honestly

I scored my own forecast and it was wrong

I built a capacity model projecting when the backlog would break. Two months later I checked it against what happened. The bug projection was over-pessimistic by more than twice over — the broader forecast, by about 30%.

The interesting part was why: the bug backlog improved because fewer bugs came in, not because throughput rose.

Had I only looked at the outcome, I'd have congratulated the team for something they didn't do.

11
The scoreboard question

Same thirty people. Different work?

That's the only question worth answering, and it needs a number rather than a story. If the same people are doing the same work slightly faster, we bought tooling. If they're doing different work, we changed something.

I'd genuinely like to know how other organisations are measuring this — because adoption dashboards answer a different question than the one leadership is actually asking.

shipscale.org · sangeethcloud@gmail.com

← All decks to move 1 / 11