Aircury

Aircury

AI made our engineers faster. Then we found a different bottleneck.

9th November 2026

What we learned from measuring AI-assisted delivery on a mature platform, before and after.

On one long-running client platform, we billed 18% fewer engineering days per sprint in 2026 than in our 2023 to 2025 baseline, and delivered 32% more tickets for each billed day. What we found most interesting came afterwards, when we saw what started slowing us down next.

The platform is around ten years old. It integrates with a dozen other systems, holds sensitive personal data and is used every day. Like most mature software, its budget goes mainly on changing code that already works: new features, maintenance, integrations and a steady stream of security fixes, all without breaking what depends on it.

AI promised to make that work cheaper. On a system like this it also carried obvious risks: plausible code that is wrong, new dependencies, and changes nobody fully understands. We wanted to adopt it, and we wanted to be able to show the client what it had changed.

Adopting it slowly, with controls first

Our infrastructure and data teams began trying AI assistants in 2024. In October that year we asked the client for permission to use AI tools on its code, and got it. A four-week joint experiment followed in June 2025, at our cost. The approach we set out in writing was simple to state: write the specification first, build the tests before automating anything, and have an engineer read and answer for every change.

The first automated agent arrived in September 2026, more than two years after our teams first tried AI assistants. It went onto one product only, where the team had first written specifications and tests for every observable behaviour. Test files on that product grew from 390 to 1,028 before the agent changed a line. We publish the method as open source in the Aircury AI Framework.

What changed

We have timesheets, code history, pull requests and tickets for this account going back to March 2023, so we could compare before and after instead of relying on impressions. Every rate is divided by days billed, because that is what the client pays for. Comparing the 2023 to 2025 average with 2026, three results stand out.

  • Billed days per sprint fell by 18%.
  • Tickets delivered per 100 billed days rose by 32%, from 20.2 to 26.6.
  • Bugs reported by the client’s staff fell by approximately 85 to 88%, from between 393 and 499 per half-year in 2023 to 59 in the first half of 2026.

The supporting measures point the same way. Releases per billed day nearly doubled. The median time from starting a ticket to finishing it roughly halved, to 8 to 10 days. Code written in 2026 remained in place after 90 days at least as often as code from earlier years, which suggests that wider AI adoption has not reduced code retention.

We are careful about cause. This is a comparison between periods, and other things changed over the same years, including team size and how much testing the team does. On the main platform, output per day started improving in 2024, before AI left any trace in the code. What we can say is that delivery got faster and quality improved while AI use grew. We left lines of code out entirely, because AI inflates any volume metric.

Maintenance that finally fits in a budget

Every mature platform carries dependency upgrades that never make it into a plan, because doing them by hand costs more than anyone will approve. On the prepared product, our Overboards agent took that work on one ticket per package. In three weeks it merged 37 major version upgrades to an integration branch, including React 16 to 19, with about five days of human supervision. Along the way it logged 22 items of debt nobody had written down, two security bugs among them.

None of this is in production yet; it waits for a person to review the integration pull request. Even so, the economics have changed. Work that had no budget line, and would not have fitted in one, became affordable, and the client paid nothing extra for it.

The unexpected part

Faster engineering moved the queue somewhere else. In the summer of 2026 it moved to our own QA team, and a temporary reinforcement cleared it within two sprints. Since then the waiting has happened either side of engineering: in getting work specified clearly enough to start, and in getting finished work reviewed and approved. At the end of some recent sprints, more tickets were waiting for review than were in development.

Many product teams have processes and approval cycles designed around engineering being the slowest step. When that changes, the organisation around the code has to change with it, and on this account we are working that out together with the client.

What we take from it

The day rate says little about value. On this account the client is billed for fewer days, and each one delivers more tickets with fewer defects. That is engineering efficiency. What it means for total spend and for business outcomes is a separate question, and one we are discussing openly with the client.

The method matters as much as the result, and it should be agreed before any experiment starts. Our data covers engineering time. A complete comparison would also count the hours the client’s own people spend specifying and reviewing, and that is the next thing we want to measure with them.

The bottleneck will move. If engineering speeds up and nobody is ready to say what to build next, the gain waits.

One rule stays fixed. Prototypes built with AI are welcome. Any change that reaches sensitive data goes through an engineer who can read it and answer for it, whoever or whatever wrote it.

If you look after a mature platform and want to bring in AI without adding risk, or need to show what it has changed, the Aircury AI Framework is open source and a good place to start: github.com/aircury/ai-framework.

We are also happy to talk through how we measured all this, because the method transfers more easily than the numbers.