AI adoption metrics can climb while lead time barely moves, and the developers using the tools may report a speed-up that a stopwatch would not confirm. Bharat Sharma makes the first half of that case in a recent article. This piece follows the second half: the signals leaders use to judge AI tools tend to point the same way whether delivery has improved or not.
Sharma found that a near-perfect adoption rate told him nothing about delivery
Sharma writes that his organization tracked AI adoption closely this year and reached numbers close to full adoption. When a peer asked how business outcomes had changed, he had no answer. He then compared the lead time of an initiative built mostly with AI assistance against similar work done without it. Coding had taken about 10% of lead time without AI, so a large gain in coding speed cut lead time by 4%. He is careful to say that figure comes from one comparison in one organization.
His argument is that adoption rate, pull requests merged and suggestion acceptance all measure creation, the one step coding assistants help with. Judged by those numbers, an organization keeps funding a step that is not its constraint. His fix is to report lead time, broken down by where the work waited and why, next to the adoption figures. In his words, “lead time belongs to the organization, not the individual.”
The arithmetic caps the gain before any tool is chosen
Sharma’s own numbers set a hard ceiling. If coding is 10% of lead time and becomes twice as fast, the coding step drops from 10 units to 5 out of 100, and lead time falls by at most 5%. If AI wrote the code in no time at all, lead time would still fall by at most 10%. His measured 4% sits just under the first bound. No improvement in the coding tool can lift that ceiling. Only shrinking the other 90% can.
That makes the share of lead time spent coding the first number a leader needs before approving more spend on coding tools. It is also a number most adoption dashboards do not show.
Developers’ sense of their own speed is a weak instrument
Sharma treats adoption metrics as incomplete. There is evidence that the most common supplement, asking engineers whether the tools help, is unreliable as well. In July 2025, the research group METR published a randomized controlled trial with 16 experienced open-source developers working on 246 real issues in repositories they had contributed to for years. Each issue was randomly assigned to allow or disallow AI tools. With AI allowed, developers took 19% longer. Before the study they expected AI to speed them up by 24%, and after it they still believed AI had made them 20% faster.
METR frames the result as a snapshot of early-2025 tools in one setting: large, mature codebases and developers who knew them well. It does not show that AI slows every team. It does show that a developer’s felt speed-up and the measured one can point in opposite directions. A leader who pairs an adoption rate with a satisfaction survey has two signals that rise together and neither of them is lead time.
At the organization level, adoption has been measured moving against delivery
The same divergence shows up at a larger scale. Announcing its 2024 report on 22 October 2024, Google’s DORA team wrote that as AI adoption increased, it was accompanied by an estimated 1.5% decrease in delivery throughput and an estimated 7.2% reduction in delivery stability. The team said improving the development process does not automatically improve delivery without the basics, such as small batch sizes and robust testing.
These are modeled estimates from survey data, and they describe association rather than cause. They are still the opposite sign from what a rising adoption dashboard implies. Read next to Sharma’s second explanation, where faster coding pushes more work into review and release, the stability figure is the one to watch. Work that arrives faster at a saturated review step can come out with fewer checks.
Lead time has to come from system timestamps, not from reports
Put the three sources together and a narrow conclusion follows. Adoption counts measure use. Self-reports measure perception, and METR found perception can run 39 points from measurement. Neither can tell a leader whether delivery got faster. The only instrument that can is lead time recorded by the systems the work passes through: ticket created, first commit, review requested, review approved, merged, deployed. Those timestamps also give Sharma’s wait-time breakdown almost for free, because each gap between two of them is a wait with an owner.
From an AI governance view, this is a familiar pattern. A decision to expand an AI system should name the outcome it is meant to change and how that outcome will be measured before the money is spent. For coding assistants the outcome is lead time with stability held steady, and the measure is the timestamp trail. Agreeing that in advance stops a high adoption rate from becoming the reason for the next purchase.
Why This Matters for Engineering Leadership
Sharma’s peer asked one question that his dashboard could not answer. The evidence from METR and DORA suggests the usual follow-up, asking the team, may not answer it either. Engineering leaders who want to know what AI tools have done for delivery need the share of lead time spent coding, the timestamp trail for real work, and a stability measure beside both. With those in hand, adoption numbers become context. Without them, rising adoption and a confident team can coexist with a delivery system that has not changed.
FAQ
What are AI adoption metrics?
They count use of AI tools: the share of engineers using them, suggestions accepted, pull requests merged. They describe activity at the coding step and say little about delivery.
Why can lead time stay flat when coding gets faster?
If coding is a small share of lead time, even a large speed-up there saves little. At a 10% share, doubling coding speed saves at most 5% of lead time.
Can developers tell how much AI speeds them up?
Not reliably. In METR’s 2025 randomized trial, developers took 19% longer with AI and still believed afterwards that it had made them 20% faster.
What did the 2024 DORA report find about AI adoption?
Google’s DORA team estimated that as AI adoption increased, delivery throughput fell by 1.5% and delivery stability fell by 7.2%.
Does this mean AI coding tools slow teams down?
No. The studies are narrow and correlational in places. They show that adoption counts and self-reports cannot settle the question, so lead time must be measured directly.
What should leaders measure instead?
Lead time taken from system timestamps, split into wait time, active work and rework, with a reason recorded for each wait, reported next to adoption figures.
