More code is not more value: measure what AI actually ships

AI coding tools make teams produce more code. That is not the same as delivering more to customers. A 2026 study of 718 firms found code output rose sharply after AI adoption while finished work barely moved. Before reporting a return on AI, measure what reaches production, not what is typed.

The headline number measures the wrong thing

A 2026 Harvard working paper by Chen and Stratton, Artificial Intelligence in the Firm: Bottlenecks in Software Production, studied roughly 300 million work events across 718 firms from January 2021 to March 2026. After firms adopted AI coding agents, lines of code rose about 30%, commits about 20% and pull requests about 23%.

Then the authors looked at finished work: issues and larger features completed. The effect was small and not statistically significant. Their estimates rule out an increase of more than 12% of the baseline. More code did not turn into more software.

It is a working paper, not a peer-reviewed result, and its firms all use an engineering analytics platform. But the pattern matches what many leaders see: dashboards report productivity gains that the customer never notices.

Speed up one stage, and work piles up at the next

Writing code was never most of the job. Software passes through several stages: deciding what to build, writing it, reviewing it, testing it and releasing it. Making one stage faster does not make the whole line faster. The work simply queues at the next slow stage.

With AI, that next stage is usually review and verification. More code arrives, often larger and less familiar, and the same number of experienced people must check it. Review becomes the bottleneck, and quality risk rises if checking is rushed.

So the business question is not "how much faster do people code?" It is "where is work waiting now, and what is that delay costing us?"

Measure the system, and decide where the time goes

Leaders can act on three things.

Choose the number the board hears. Report delivery measures, such as how long work takes from start to customer and how much finished work ships, rather than code volume or tool usage. These are the numbers that connect technology spend to business outcomes.

Make review a designed stage, not a queue. Treat checking and testing as capacity to plan, with clear standards and the right people, rather than as a side task that absorbs the extra load.

Decide where saved time goes. If AI frees capacity, choose deliberately whether it goes to quality, to more ambitious work or to cost. Left undecided, it disappears into the queue.

One caution matters most. The same study found no significant change in employment attributable to AI. Cutting teams on the assumption that more code means fewer people needed is a structural decision based on a number that never reached the customer. Look at the delivery trend first.

What to do

• Ask for one chart: finished work delivered per month, before and after AI adoption.
• Find the stage where work waits longest today, and give it an owner.
• Hold staffing decisions until delivery measures, not code measures, show a sustained change.

If you need to show what your AI investment returns, let's discuss how to measure it.