AI coding requires better system understanding, Norwegian consultants argue
Norwegian consultants Silvija Seres (technologist and strategic adviser at Technorocks) and Ole Rabanal (architect and senior developer at Sonat) argue in a debate piece that AI-assisted coding can deliver three times the pace or slow development down, depending on how well a team understands the organisation it works in. Their point is that how fast AI writes code says nothing about how fast a team can make a correct change in a real system.
They note that Anthropic has just launched Claude Fable 5.1 and OpenAI GPT-6 Astra, both able to write and change code quickly, including reading large programs, finding bugs and making changes in complex systems. The hard part, they say, is often deciding what a change should involve. Most companies extend systems whose data and rules have grown over many years, so teams must establish which data applies, which exceptions to keep and how a change affects other systems. Without that, a model can produce working code that solves the wrong task.
They cite a study in which experienced developers expected AI to make them 24 percent faster in codebases they knew, but actually took 19 percent longer. The study used tools from the first half of 2025, and the researchers expect newer tools to give bigger gains.
The authors describe projects where large firms rebuilt older business-critical systems under tight time and cost limits. Teams interviewed users, focusing on exceptions: asking "what happened the last time an order failed?" can reveal an unwritten rule that "how does the sales process work?" will not. They then produced an architecture with diagrams and specifications for each task and component, which the AI agent could consult. Developers reviewed and tested the code. This approach gave up to three times higher delivery pace with quality maintained and less manual maintenance.
The authors call this making the enterprise "legible to AI": models need definitions of business terms, the authoritative data source, known exceptions, how to test a correct solution, and preserved decisions. They expect teams to spend more time on interviews, specification and control, and less on coding, which they call re-architecting for AI. The more code AI can write, they conclude, the more it matters that people understand what the business actually does.