Most modernisation programmes start with a discovery phase, and most discovery phases end in a slide deck. By the time it reaches the executive team, parts of it are already wrong. A team has shipped something, a contract has renewed, a dependency has changed. The programme starts on the back foot.
We learned this the hard way with a regulated financial services client.
Eight services, they said
We were told the client ran about eight services. On day one, I was given access to hundreds of code repositories. Which ones were actually in use? All of them, came the answer.
There was almost no documentation. When something broke, everyone scrambled to work out how it functioned, and often broke something else doing so. Reviewing the estate by hand would have given me about four minutes per codebase.
Our first instinct was to point an AI model at the code and ask it questions. That got us so far. But when every service depends on a dozen others, you can’t draw a box around one codebase and call it understood. And ask an AI the same question twice and you may get two different answers.
A different way of thinking
So we changed our approach. Instead of asking AI for the answers, we asked where the evidence lies - which code calls other services, which network logs show traffic between them. A second agent then wrote ordinary, deterministic code to extract that evidence the same way every time, so we could rerun it whenever the code changed, for almost zero cost.
Then we layered in the business - recordings of our value stream mapping and event storming sessions, even photos of the sticky notes. What started as a way to read code became a single bank of context about the business and its technology.
That became Atlas, CreateFuture’s approach to mapping a complex technology estate.
The scan is the easy part
Knowing what’s deployed, what it costs and who looks after it are the basics, and anyone charging a premium for it is selling you the easy part. The real challenge is knowing what to look for.
When it comes to a regulated lender, compliance and security are critical lenses. A business running on Azure has no use for an AWS cost model. Atlas is less a product than a method, and it’s only as good as the people asking the questions.
What we found
Some findings needed fixing immediately, including security issues that had gone unnoticed for years. Others were simply expensive. The client’s compute clusters were using just a fraction of the processing capacity they’d paid for. Nobody had spotted it, because nobody had a view across the whole estate.
The cloud savings alone more than paid for the discovery phase. Within eight weeks, the client had a fully costed menu of opportunities, and we agreed together which ones to pursue. The first quarter of delivery has been played back to the board, delivery metrics have moved significantly, and the second phase is under way.
Early on, staff told us an automated identity check failed around 70% of the time. We started planning a replacement. However, when we measured it, the failure rate was actually 50 times less.
Memory is shaped by frustrating cases. Decisions must rest on evidence, gathered the same way every time.
Discovery is not just for phase one
Because Atlas gives consistent answers, it’s more than a snapshot and becomes the instrument you plan against. You can model a change before you make it, then check afterwards whether it was delivered.
It changed how we worked, too. The client told us plainly that if we sent slide decks, they wouldn’t read them. So findings went onto a secure web app as we found them, and weekly playbacks became executives asking smart questions about work they’d already read.
The bottleneck in AI-assisted delivery
The industry conversation about agentic development is fixated on models. On legacy estates, the bottleneck is context. When the client’s engineers started using AI coding tools, an agent asked to add a feature that would cheerfully invent a new service, unaware that one already existed. Feeding Atlas’s context into their development process stopped that.
Our strong belief is that organisations that haven’t mapped their estate based on runtime evidence will find AI accelerates their mistakes as fast as their delivery.
Every technology leader will end up using AI to understand their estate. The choice is whether they settle for a one-off answer or build something they can steer by.
If you’d like to see what that looks like on your own estate, we’d be happy to show you. Find out more about how we rapidly turn complex technology estates into a living map here.
FAQs
Meet the author
Chris Daly is a Principal Engineer at CreateFuture. With over 12 years building cloud-native platforms across .NET, TypeScript and AWS, including roles at Kin + Carta, Valtech and AI-native start-up Wholepal, he helps organisations modernise legacy systems and take secure, AI-powered products from zero to production.
Industry Insights
Explore the latest thinking from our industry and tech experts.
What is AI native?
Why do AI experiments fail and how do you scale AI-native delivery?
CreateFuture expands its Anthropic AI capabilities as Version 1 Group achieves Preferred Services Partner status