When AI comes up in client work, I usually meet two basic stances: enthusiasm and reservation. The cautious voices get pushed into the laggard corner pretty quickly. In my experience they are the ones who contribute most to a successful rollout, if you actually listen to them.
Why scepticism isn’t an information problem
When someone in a workshop asks how an AI decision can be reversed, it sounds like resistance at first glance. On second look, it’s a precise quality question. The same goes for the lawyer asking about liability, the data protection officer probing the training data, or the team lead checking how consistent the results are. These are the people who will carry the project later.
Explanations alone don’t do much against this kind of reservation. A single visible mistake can damage trust in an algorithm for a long time. Research calls this algorithm aversion: machine errors get judged more harshly than human ones, even when the algorithm is on average better. What changes the stance is the option to intervene. The mere chance to override or correct an output shifts attitudes more than another slide in the explainer deck.
Which kind of transparency helps, and where regulation sets the frame
Transparency means something different to everyone. The controller wants to see what a recommendation rests on. The data protection officer cares about training data and model documentation. The colleague in day-to-day operations needs a confidence indicator and the option to discard a suggestion. There is no single explanation that covers all of that at once.
This is where regulation actually helps. The EU AI Act sets a deadline for companies running high-risk AI: 2 August 2026. By then the systems have to be classified, assessed and registered. Article 13 obliges providers to enable users to interpret outputs in a meaningful way. These are things a well-run organisation should sort out anyway.
ISO/IEC 42001 adds a management system for AI governance with policy, controls and regular audits. The structure is familiar from information security and quality management, now applied to AI. For executives that’s useful because concerns can be answered with documented evidence.
The doubts inside the organisation come from the responsibility that stays with the company, even when the decision was made by the model. A working management system helps, because the answers to the good questions are documented and auditable.
In the end, the goal is to give people the means to answer good questions with confidence. The colleagues asking those questions tend to be the strongest allies of an AI rollout.
