Most companies have long since got started with AI. Employees use ChatGPT and Claude in their day-to-day work, developers are testing new tools, and the pilot projects look promising. But the leap from experiment to production is a big one, and this is where many stall. That is the message from Adam Peter Nielsen, CTO at Novicell, in an interview in Børsen on Wednesday 16 September 2026. Read the full article here.
We have taken the interview as our starting point and look more closely here at why so many AI initiatives never reach the business, and what it takes for them to do so.
Architecture determines whether AI creates business value
Most companies' IT landscapes were built long before generative AI and agents became part of everyday life. That is rarely an argument for replacing everything. But the systems need to be able to connect with new technologies, and data needs to be usable where it creates value, without compromising security or control.
This applies both internally and externally. Some data needs to optimise your own processes. Other data needs to be made available to customers, partners or digital agents, and here access alone is not enough. An agent also needs to understand what the data means and which rules apply. Otherwise, it will never become more than a promising trial.
"It's easy to make something that looks convincing on the surface. The hard part is making sure the solution can work with the right data, meet the company's requirements and work together with the rest of the business. That's typically where the experiment meets the reality of implementation." says Adam Peter Nielsen.
That's why we modernise the architecture gradually, with no big bang and without putting operations at risk. In practice, we open up controlled access to selected data and functions, so new AI solutions can be built on top of what already exists, while the older systems are modernised along the way.
"Being ready for AI isn't about buying one particular platform. It's about setting up your architecture so data can flow securely and be used where it creates value. The company needs to be able to open up its systems without giving up ownership, security and governance." says Adam Peter Nielsen.
The price of missing architecture isn't just failed pilot projects. As customers and buyers increasingly use AI agents to find information, compare suppliers or complete parts of a purchase, companies without the right architecture lose visibility – not because the product is poor, but because it can't be found. The same applies to your product data: if it's written for people and not for machines, the agent simply skips you.
A roadmap connects AI with the business
AI only becomes strategic when the technology is linked to revenue, costs and customer experience. We typically start a collaboration by drawing up a roadmap that prioritises where the potential is greatest, which prerequisites need to be in place in data and architecture, and what needs to be delivered first.
Most companies don't need more isolated AI trials. They need a rational plan that shows where the technology can increase revenue, reduce costs or improve the customer experience, and which data and systems that requires.
Adam Peter Nielsen
CTO at Novicell
How it's built in from the start
We have built our own agent infrastructure, which all new, larger projects are based on. Trained, specialised agents work end to end: from research and design to code, testing and documentation. This means AI is used to solve the AI problem, and it's built into the solution from the start instead of being added on top afterwards.
"Technology is no longer necessarily the biggest constraint. The challenge becomes creating ownership, prioritising the opportunities and making sure the organisation can keep up. When output becomes faster and cheaper, it places greater demands on the decisions." says Adam Peter Nielsen.
From promising experiment to competitive advantage
AI shouldn't stand alone as an isolated project. Only when strategy, architecture and execution are connected can AI move from promising experiment to a competitive advantage that can actually be scaled.
Novicell and JABS Group unpack the issue at Copenhagen Composable Conference 26
Adam Peter Nielsen puts the same agenda up for debate at Copenhagen Composable Conference on 30 October 2026, where, together with JABS Group, one of Europe's leading manufacturers of windows and doors, he takes the topic a level deeper.
The point is sharp: an AI agent without access to your architecture and conventions is as useless as a new employee without onboarding. Without the right context, the work stalls every time something needs to be interpreted or decided.
Without that framework, most end up with AI agents becoming scattered one-off trials: each agent has to be briefed from scratch, quality varies from task to task, and the work comes to depend on the few people who know the system inside out. That makes it hard to scale, slows down time-to-market and makes it difficult for management to stand behind the result.
That's the challenge we have solved internally with Novicell's Agent Infrastructure: a structured workflow where specialised AI agents build, test and deliver digital platforms within clear boundaries. Each agent has access to exactly the context it needs: architecture, conventions and previous decisions, so the work doesn't stall every time something needs to be interpreted. The result is a significantly shorter time-to-market: tasks that used to take months can now be completed in days. A person approves and is accountable for the final result, but the heavy, repetitive work is carried by the agents.
At Copenhagen Composable Conference, we and JABS Group will show how the infrastructure is built, which problems it solves, and what it takes to make your development organisation agent-ready.