Velocity Is Rising. Quality Isn't.

In the past six months, I’ve heard the same sentence in executive reviews and retrospectives:

“AI has doubled our team’s productivity.”

Velocity is up. More stories are closing. Pull requests are flowing.

The metrics look like something is working.

But sit with it a little longer and something feels off.

Engineers are spending more time in review. Security teams are flagging more issues. Refactors quietly bleed into future sprints. Bugs are subtler now. Harder to catch. Harder to trace.

This isn’t AI being bad at generating code.

It’s AI being very good at generating plausible code.

And plausible isn’t the same as durable.

AI and the Illusion of Productivity

In many teams, output has accelerated faster than validation capacity has grown.

Developers feel faster. Leadership expects faster.

But review discipline — the ability to absorb, evaluate, and confidently ship what’s generated — hasn’t scaled with generation speed.

The result isn’t collapse.

It’s review debt.

The system looks productive on the surface. Downstream, friction compounds. Change failure rates drift upward. Senior engineers become quiet bottlenecks.

Nobody announces this shift.

It becomes the texture of the work.

Velocity measures throughput.

It does not measure coherence, maintainability, or strategic fit.

A team can close fifty stories and still degrade the system.

Speed is visible.

Structural fragility is not — until it is.

If you want to understand what’s really happening in an AI-assisted environment, stop watching velocity alone.

Watch a different pairing:

Deployment frequency versus change failure rate.

If both rise together, you’re improving.

If deployment frequency climbs while change failure rate drifts upward — even slightly — you’re accumulating silent instability.

This is exactly why DORA pairs those metrics. It distinguishes busyness from health.

As noted earlier in the backlog article, faster implementation without faster validation simply shifts the bottleneck.

AI hasn’t broken delivery.

It has exposed weak feedback loops.

Over the next decade, the winning teams won’t be the fastest coders.

They’ll be the fastest learners.

What are you actually measuring right now?