Faster code, same release dates: the AI delivery paradox

Ask a team that has adopted AI coding tools whether they are faster, and most will say yes without hesitation. Ask the business whether software is reaching customers faster, and the answer is often a shrug. Both can be true at once. Understanding why is the first step to getting a real return from AI in software delivery.

The evidence does not match the feeling

The research so far points in an uncomfortable direction. Individual tasks may feel faster, but delivery as a whole does not automatically follow.

The second finding is the more striking one. In METR’s study, experienced open-source developers expected AI to make them about a quarter faster. After finishing, they still believed it had. The stopwatch said the opposite.

Change in task completion time for experienced developers using AI tools

SlowerFaster

• Expected before starting24% faster24% faster
• Believed after finishing20% faster20% faster
• Actually measured19% slower19% slower

Source: METR randomized controlled trial, 2025. 16 experienced open-source developers, 246 real tasks on their own projects.

None of this means AI tools are useless. It means that the sense of speed at the keyboard is a poor guide to what is happening across the whole delivery system. Leaders who manage by that feeling will make the wrong calls on staffing, estimates and investment.

Where the time actually goes

Delivery is a chain: shaping the work, writing the code, reviewing it, testing it, integrating it and releasing it. AI tools mostly accelerate one link, writing code. The other links stay the same, or get harder, because there is now more code arriving, in larger batches, written by a tool that nobody on the team fully reasons like.

Share of lead time by stage, for a typical change. Illustrative, not measured data.

Before AIBaseline

ShapeCodeReviewTestIntegrate

With AI coding toolsCoding is faster, review and testing absorb it

ShapeCodeReviewTestIntegrateRelease

• Shape
• Code
• Review
• Test
• Integrate
• Release

The result is a familiar pattern from any system with a bottleneck. Speeding up a step that was not the constraint simply moves work into the queue in front of the step that was. In most teams we see, that queue forms in three places:

• Review. Senior engineers become the bottleneck, reading code they did not write and that is often “almost right”.
• Testing and environments. More change means more to verify, on test infrastructure that was sized for the old pace.
• Integration and release. Larger batches collide more often, and failures after release take longer to diagnose.

Why process, not tooling, decides the outcome

Most delivery processes were designed when writing code was the expensive part. Handoffs between analysts, architects, developers, testers and operations existed to make the best use of that scarce capacity. When code becomes cheap, those handoffs become the cost: each one loses context and delays feedback.

That is why adding more AI tools rarely fixes the problem. The constraint is no longer the typing. It is how quickly a team can understand, verify and integrate what was produced, and how much context is lost on the way.

What to do about it

The organizations getting real gains treat AI as a reason to redesign the flow, not just to equip the people in it.

1. 1Measure the flowTrack lead time from idea to production, not developer activity or tool usage.
2. 2Find the queueLook for where work waits. It is rarely in coding.
3. 3Shrink the batchSmaller changes, reviewed and tested earlier, so queues stop forming.
4. 4Automate the gatesMove routine quality checks into the pipeline so people review what needs judgment.
5. 5One pod, one outcomeA small senior team owns a change end to end, in short loops.

The last step matters most. A small, senior, forward-deployed pod that shapes, builds, verifies and ships its own work removes most of the handoffs where time and context are lost. That is the idea behind our Forward Loop.

The bottom line

AI has made writing code cheap. It has not made delivering software cheap. The gap between the two is where most of the return on AI is won or lost, and closing it is a matter of flow, roles and process far more than of tools.

If release dates have not moved since your teams adopted AI, that is not a failure of the tools. It is a signal that the system around them needs to change.