AI Adoption is an Integration Problem
Thu 16 Apr 2026
AI initiatives often struggle before model performance becomes the issue.
Across organisations, tools are deployed, pilots are launched and leadership signals intent. Yet systems underperform, degrade over time or are quietly abandoned. The difficulty tends to sit upstream — in integration, structure and organisational alignment — rather than in the models themselves.
Ahead of Tech Show Frankfurt on 6-7 May at Messe Frankfurt, Marie Kilg, AI Advisor and host of Der KI-Podcast (ARD), examines why AI struggles to move beyond isolated use cases, how organisational design and culture shape outcomes, and why meaningful adoption depends as much on system alignment as on technology.
– – – – – –
You’ve worked across product, journalism, and AI strategy. Where do you see organisations still misunderstanding why AI adoption fails?
Most organisations frame AI as a problem of adoption. It’s actually a problem of integration.
Traditional software you could implement in fairly isolated steps: you buy it, you connect it to your systems, you teach people how to use it. For AI to work really well, it needs context. That often means you have to put effort into your existing systems too: How documents are stored, how information flows between teams, how things are labelled. You have to look at the whole process holistically, not just the AI step.
Can you give an example?
Take employee onboarding. A company might build a chatbot that new hires can ask questions, backed by documents from HR, internal guidelines, and policies. But the system only becomes genuinely powerful when the information behind it is complete and always current. Which means you may need to change where different teams store their documents and how they flag the latest version – before the AI layer even matters.
What I see repeatedly is organisations building only the first step, finding it works maybe 70% well, and then watching it degrade because nobody manually maintains it, and the upstream processes are not well designed. Eventually, it gets abandoned, and the conclusion is “AI didn’t work for us.” The AI was fine. The integration wasn’t.
There’s a push from leadership to “use AI,” but that doesn’t always translate into day-to-day adoption. When organisations focus on scaling AI, what role do culture and behaviour actually play?
The change that’s needed is really fundamental. You need to change the culture, not just learn a new software. That takes time, and you can’t just mandate it.
A lot of organisations are going from an assembly-line structure — with lots of separate teams and clear handoff points — to a system that’s supposed to be interconnected, well-documented, and structured enough that relatively basic machines can understand it. That is a huge change. You need buy-in from as many teams as possible, because every little cog can endanger the mission if they resist.
But here’s what leadership often misses: employees might have learned that they need to resist external demands to a certain extent — to protect their deliverables, to not be ground down by requests from teams that don’t understand their work. If you’re supposed to use a new AI system that’s only half-integrated and doesn’t match the reality of your workflows, it’s neither useful nor fun. So people resist — and they’re right to. Leadership shouldn’t see that as a failure or unwillingness to adapt. They should try to find out why the system isn’t useful. Employee feedback is a great resource.
You’ve spoken about trust as a deciding factor. How does trust in AI form inside an organisation, and what tends to erode it?
Trust plays out on multiple levels.
When you’re asking teams to open up their processes, share their documentation standards, and accept interdependencies they never had before, you’re asking them to become vulnerable. A team that’s spent years protecting its deliverables by keeping a healthy distance from external demands now has to collaborate deeply with teams that may not understand their work. That only happens if people believe their openness won’t be used against them — that they won’t be penalised in the next round of layoffs for being easily automated. Or labelled as “resistant to innovation” when they’re raising legitimate concerns.
I’ve seen this play out in journalism. Newsrooms used to function well as relatively independent teams: reporters research and write, then pass the article to layout, then it goes to print. As long as humans were managing the handoff points, each team could change how they worked internally without affecting the chain. With AI, you suddenly need data compatibility across the entire workflow. That limits the flexibility teams are used to — and it forces new ways of collaborating that don’t come naturally. A technology team might insist that journalists can’t install their own tools. Journalists might feel that engineers don’t understand editorial needs. Without trust, those tensions become blockers.
Leadership has to model this. It means taking concerns seriously, giving teams genuine autonomy in how they adapt, and, critically, making sure that honesty is rewarded, not punished.
Many organisations talk about experimentation. In practice, what tends to limit people’s ability to test and adopt AI tools?
There’s a chicken-and-egg problem I see everywhere. A team wants to know whether a tool works for their process, so they need to test it. To test it, they need approval from Legal, IT Security, or Procurement. But those teams want justification and specifics that can only come from having tested it. Everyone is waiting for someone else to go first.
AI tools make this worse than usual. The market moves so fast that you can’t rely on competitor studies or established best practices. You genuinely have to try things yourself. At the same time, the terms and conditions can be alarming, and you certainly can’t take every vendor’s claims at face value. So companies find themselves stuck in grey zones, needing to experiment but lacking the structures to do so safely.
Some organisations are agile enough to experiment on the fly. For everyone else, you need sandbox environments: dedicated “labs” with high-level permission to operate outside normal protocols, combined with clear risk management and decision frameworks. The third option is guerrilla innovation: individuals and small teams experimenting in the shadows. If you’re lucky, you’ll hear about what they’ve learned. If you’re not, they may be exposing the company to risks nobody is aware of.
Looking ahead, what will distinguish organisations that meaningfully embed AI from those that continue to struggle?
Strategic courage — and the structures to act on it. A lot of organisations have the ambition but not the decision-making infrastructure. They want to move on to AI, but decisions get stuck between departments, between leadership layers, between competing priorities.
The organisations that will pull ahead are the ones willing to make clear, sometimes uncomfortable decisions about how they work — not just which tools they buy. That means giving teams real permission to experiment, creating the cross-functional trust that integration demands, and accepting that meaningful AI adoption will change the organisation itself, not just its toolset. The companies that treat AI as something you layer on top of existing structures will keep struggling. The ones that treat it as a reason to rethink those structures will build something lasting.
