Astragal

Evidence

Businesses all over the world are moving the same way right now: handing repetitive work to automation and agents, and building platform teams so growth does not mean breakage. It is one of the largest shifts in how companies operate in a decade, and unusually for a trend, quite a lot of it has been measured in public.

Four of those results are below, each with a link to the source. Each one also carries the part that did not go to plan, because a case study without its correction is advertising. If you read one thing on this page, read the caveats. None of these companies are clients of ours.

Automating busywork

Klarna's customer service assistant

Klarna put an AI assistant in front of its customer service queue. In the first month it handled 2.3 million conversations, roughly two thirds of all its customer service chats. Klarna reported that people resolved their issue in under 2 minutes rather than 11, that repeat enquiries fell by 25%, and that satisfaction scores were about level with human agents.

The part most write-ups leave out: in 2025 Klarna walked some of this back and brought human agents back for complicated cases. The widely repeated "work of 700 agents" line was also a modelled equivalence, not 700 people who left.

We think the correction is the more useful half. Automation absorbed the volume and did not remove the need for people at the hard end. That is exactly the shape we design for: the machine takes the repetitive nine tenths, and a person is deliberately kept in place where being wrong is expensive.

Source: Klarna press release

Building what you grow on

Spotify's internal platform

Spotify grew to the point where its own engineers could not reliably find their own systems. Rather than hiring around the problem, they built one place to see everything, called Backstage. New engineers reached their tenth piece of shipped work 55% faster. Spotify open sourced it, and it is now used by over 3,400 companies.

Be clear about scale: Spotify is enormous and you almost certainly are not. Adopting their tool because they use it is how companies end up maintaining a developer portal nobody asked for.

The transferable finding is not the software. It is that the hours engineers lose to not knowing where things live is a real and measurable cost, and that it can be roughly halved. On a team of ten that is worth doing with a wiki page and some discipline, long before it is worth doing with a platform.

Source: Backstage by Spotify

Measured properly

The one that was a real trial

Most claims about AI at work are company blog posts. This one is a controlled experiment. Researchers gave 95 professional developers the same job, building an HTTP server, and gave half of them an AI assistant. That half finished 55.8% faster. The result held up statistically, with a p-value of 0.0017.

Read the confidence interval before you get excited: the true effect sits somewhere between 21% and 89%. That is an enormous range. It was also one self-contained task written from scratch, which is the friendliest possible case, and nothing like changing a large system you have to keep running.

We include it because it is the most honestly measured number on this page, and because it still comes with a caveat that wide. Anyone quoting a single confident percentage at you has not read their own source.

Source: Peng et al., arXiv 2302.06590

The whole field, not one company

Thousands of organisations, measured yearly

The four numbers on our home page come from the DORA research programme, which has surveyed engineering organisations every year for roughly a decade. Its central finding is the one most people refuse to believe: teams that ship more often also break things less often and recover faster. Speed and stability travel together.

It is survey data, so it is self-reported, and it shows correlation rather than proof of cause. Being fast does not by itself make you stable. The practices underneath, small changes and good tests, are what produce both.

This is the evidence base we actually work from. One company's story can be luck. A decade of the same pattern across thousands of them is something you can plan around.

Source: DORA research programme

None of these organisations are clients of ours and none has any connection to Astragal. They are here because they published figures that can be checked, and because we would rather show you someone else's real numbers than invent our own. The benchmarks in the fourth card are the ones we measure against on the home page.