AI is often sold as a smart assistant. Something that helps, supports, saves time. In reality it does something more fundamental: it redistributes power, accountability and trust.

Whoever designs the technology also designs the relationships within the organisation.

That is not an exaggeration. It is a direct consequence of how decision processes work once a model plays a role in them. I keep seeing three shifts recur, and none of the three is named out loud often enough before implementation is already well underway.

The first shift concerns who prepares decisions. Where a professional used to open a file themselves, recognise a pattern and form an assessment, a model now often delivers that assessment first. The question that arises with this is not technical but relational: do you trust the model, or do you trust the professional? Must someone who deviates from the model’s advice justify that, and to whom? Is dissent seen as difficult behaviour, or precisely the expertise you hired that professional for in the first place?

That question is rarely answered explicitly before the system goes live. The answer then emerges by itself, in practice, and usually in favour of the model, simply because the model appears faster, more consistent and, at first glance, more objective than a human judgement. That is not a technical truth. It is a social dynamic that unfolds whenever nobody has consciously organised a different answer.

Models reproduce the history of your choices. What was implicit becomes explicit and scalable.

The second shift concerns accountability. Suppose someone is wrongly disadvantaged by an AI recommendation. Who explains that? Who offers an apology? Who has the authority to correct it? Without clarity on this, a no-man’s-land emerges, a sentence I encounter in almost every organisation that has not thought this through: that’s just what the system does. That sentence sounds harmless at first glance. In reality it is a form of shifting accountability onto something that cannot be held to account, because it has no face and no voice.

That is precisely why this question must be answered in advance, not afterwards once damage has already occurred. Who is ultimately accountable must be known before the first incident arises, not only afterwards, when the pressure to quickly find a scapegoat gets in the way of careful thinking.

The third shift concerns who determines the norm. A model is trained on historical data, and that data carries the history of your choices and your biases, including biases that were never named as such. What used to be implicit, perhaps even unconscious, is made explicit by a model and applied at scale. That can be an opportunity: it makes visible patterns you no longer actually want. It can also be a risk: it confirms and reinforces a norm you actually wanted to revise, without anyone noticing, because the norm is now hidden inside an algorithm rather than in an explicit policy rule.

Beneath these three shifts lies a psychological layer that matters just as much as the structural one. Managers may experience a sense of lost status or lost control when patterns that used to remain hidden suddenly become visible in data. Professionals sometimes experience AI as a form of control, or as a devaluation of the craft they have invested years in. And a board can be tempted to steer primarily on dashboards, because that feels more manageable than a complex conversation, with the result that dialogue within the organisation impoverishes precisely at the moment it is most needed.

Normalising dissent is not a friendly piece of advice. It is a design task.

What does this require of management, concretely? Three things, I think, that reinforce each other and that none alone is sufficient.

First: make the division of roles explicit. Who remains ultimately accountable for which decisions? Where is human override not only permitted but required? Record that in mandates and processes, but also in the story you tell, internally and externally. A mandate that exists only in a policy document, never spoken aloud in a team meeting, does not exist in practice.

Second: normalise dissent, actively and repeatedly. It helps to literally state that it is professional behaviour to question an AI outcome and, where necessary, to override it. It helps even more to share examples of moments when that actually happened, with names and outcomes attached, so it does not remain an abstract statement but becomes a recognisable pattern.

Third: discuss the undercurrent out loud. Name the fear and tension these shifts cause, rather than pretending they are not there. Combine that honesty with a genuine development path, so that people feel not only their discomfort heard, but also receive a direction for how their role develops alongside a system that takes over part of their former tasks.

There is one more possibility worth mentioning, one that goes beyond risk management. Use AI as a mirror. Deploy it to make visible patterns you no longer actually want: systematic disadvantage to a specific group, steering on speed at the expense of care, decisions that consistently fall the same way without a clear reason. In that way AI becomes an instrument for the organisation’s maturation, not only an instrument for efficiency.

Concretely, you can start today with an override log: a simple record of moments when professionals deviate from the model’s advice, and what the outcome was. Over a slightly longer horizon, you can design a fixed dissent ritual, for instance a quarterly review in which a small team deliberately looks critically at the system’s outcomes. And what you should especially avoid is treating every deviation from the AI’s advice automatically as a mistake. That discourages exactly the behaviour you need to keep the system sharp and honest.

Does your use of AI mainly strengthen control and hierarchy? Or does it strengthen trust and professional dialogue?

The answer to that question is rarely entirely clear-cut. But asking it, repeatedly and out loud, is precisely the work this third instalment of the series is concerned with.

Notes for those who wish to read further:

1. Wanda Orlikowski, Sociomaterial Practices (2007, Organization Studies). On how technology and organisational relationships continuously shape each other, rather than technology being a neutral instrument existing apart from the organisation.
2. Amy C. Edmondson, The Fearless Organization (2018, Wiley). On psychological safety as a precondition for genuinely normalising dissent against automated outcomes.
3. Mary Uhl-Bien & Russ Marion, Complexity Leadership Theory (2009, Organizational Dynamics). On how leadership must adapt as decision-making becomes distributed between people and systems.
4. Sidney Dekker, Drift into Failure (2011, Ashgate). On how organisations gradually drift from their own standards when accountability becomes diffusely distributed between people and systems.
5. Cathy O’Neil, Weapons of Math Destruction (2016, Crown Publishing). On how models reproduce and reinforce historical bias at scale, often without that being the intention of the original builders.