Why AI Adoption Fails Without Trust, Integration, and Grounding

At Tech Show Frankfurt, Marie Kilg, AI Advisor and host of “Der KI-Podcast” (ARD), focused on a recurring problem in enterprise AI: many organisations want adoption, but few understand why it does not happen.

Her keynote positioned AI adoption less as a question of access to tools and more as a question of trust, workflow design, infrastructure, and organisational behaviour.

Kilg opened with a personal example: losing access to a preferred AI model after it was deprecated. The story illustrated a broader issue for organisations: AI systems change quickly, but business processes and decision-making structures often do not.

This speed creates uncertainty. Leaders may feel they are already behind, while employees may fear replacement or hesitate to experiment. In that environment, adoption can stall before it begins—not because people lack tools, but because the organisation lacks structures that allow safe and useful experimentation.

As Kilg put it in the Q&A: “No, I think it’s a step before a lot of companies struggle to even start the experiments because the structures don’t allow for this.”

The Clever Hans Lesson

A central thread in the talk was the story of Clever Hans, the horse believed to perform calculations but later found to be responding to subtle human cues. Kilg used the example to challenge how people interpret AI capabilities.

“The language models don’t reason, they memorise right.”

The point was not to dismiss AI, but to caution against over-reading its abilities. AI outputs can appear convincing, especially when users refine prompts or select the best result. This can create a misleading sense of intelligence unless the output can be checked.

“He was reading the room. He was not Reasoning.”

For organisations, this means adoption should start with tasks where there is a clear way to verify results.

“It’s the grounding in reality.”

Grounding may come through calculation, expert review, audience testing, or comparison against known facts. Without it, the cost of checking AI outputs can outweigh the benefits of using them.

From Adoption to Integration

Kilg distinguished between giving employees access to AI and embedding AI into the way work happens.

“You can’t just tell people to use AI without giving them any type of integration into their workflows, into your systems.”

This is a practical distinction. A chatbot subscription may help individuals, but sustainable adoption requires integration into data sources, production chains, knowledge systems, and team processes.

Using a media organisation as an example, she described AI opportunities across research, production, distribution, audience interaction, administration, and infrastructure. Each part of the chain may require different tools, policies, and expertise.

Designing for Flexibility

Because AI models and commercial terms change quickly, Kilg argued that organisations should avoid becoming dependent on a single model or provider where possible.

“Sometimes you have to be modelling agnostic.”

This is not only a technical recommendation. It affects procurement, architecture, risk management, and long-term planning. If an organisation builds too tightly around one model, changes to pricing, availability, or performance can disrupt workflows built on top of it.

Experimentation With Guardrails

Kilg recommended using an application matrix to assess use cases by both technical readiness and ethical, legal, or business acceptability. Some applications may work technically but still be inappropriate for a brand or audience. Her example from media was photorealistic image generation: the technology may work, but it could confuse audiences or damage trust.

The deeper implication is that adoption decisions cannot be made by technical teams alone. They require input from legal, editorial, product, leadership, and audience-facing teams.

“You have to fail forward in AI, or you won’t get anywhere.”

Failure, in this framing, is acceptable when experiments are small, intentional, and linked to learning. It becomes risky when pilots are disconnected from strategy or deployed without clear evaluation.

Trust as the Adoption Layer

Kilg closed by returning to trust.

“But that’s the thing, trust is relationship.”

This applies internally as much as externally. Employees need to see leaders experimenting visibly, admitting uncertainty, and giving managers clear roles. Otherwise, AI adoption can be interpreted as a hidden productivity test or job threat.

The final message was that adoption depends on behaviour as much as infrastructure. Organisations need the right tools, but also the trust, communication, and psychological safety that allow people to use them well.