How mature is your ethics when it comes to AI

Ethics sounds like the topic for a working group that meets twice a year. In practice, it is the dimension that determines whether an AI application is allowed to keep running once someone asks a difficult question. Who decides to use an output, who bears the consequences if that output turns out to be wrong, and on the basis of what consideration that happens: if there is no answer to that, a pilot stands on loose ground, even if the technology works.

What this dimension measures

The maturity measurement does not look at whether an ethical code exists on the intranet page. It looks at whether ethical considerations land somewhere in the work itself: in the way an AI application is approved, in who is allowed to reverse a decision, in how an outcome that no one had foreseen is handled. These are not issues that exist separately from the rest of the organization. They are connected to the dimensions that need to be in place first: an organization that does not yet have a structure to assign AI decisions can hardly carry an ethical framework that goes beyond a statement of intent. The same applies on the infrastructure side and data handling: whoever does not know which data a model has used cannot give a meaningful answer to the question of whether that use was responsible.

The five levels, from baseline to intelligence

At baseline level, ethics exists as a theme mainly in reaction to incidents. There is no documented framework; if something goes wrong, it is looked at ad hoc, usually by whoever happens to be present. At foundation level there is a document, often taken over from a sector standard or legal template, but it is rarely consulted before a decision is made — it exists, but does not function.

At activation, the framework starts to move along with the process: there is a fixed moment at which a new AI application is assessed for possible harm, bias or unintended use, and that moment is not skipped. At insight, that assessment is fed by what actually happens: outcomes are tracked, deviations are discussed, and the framework is adjusted based on what practice shows rather than what was expected in advance.

The intelligence level is not so much a thicker document as an organization in which ethical consideration is woven into decision-making itself: people in different places recognize a gray area before it becomes a problem, and there is a habit of making that area discussable without it requiring a formal escalation process.

Where you see the difference

The distinction between the levels does not lie in the text of a policy document but in what happens the moment something goes wrong. At the lower levels, the search is then for who made the mistake. At the higher levels, the question is whether the framework could have caught the moment, and if not, what changes as a result. Another signal is who is allowed to ask the question. If ethical doubt can only travel upward through a formal channel, it often gets stuck along the way. If everyone who works with an AI application can raise a question without having to justify a reason for doing so, that is a sign that the framework is alive.

This dimension is not separate from the other six. It touches directly, for example, on how your organization handles privacy and security in AI, because an ethical framework without clarity about who is allowed to view which data is a framework without ground. It also touches on what your people know and can do when it comes to AI: an employee who does not understand how a model arrives at an outcome can hardly assess whether that outcome is ethically responsible to use. And it touches on how performance is measured and discussed in an AI environment, because a review system that steers only on speed treats ethical considerations as a delay rather than as a necessity.

The plot round: why spread says more than an average

A single score on this dimension says little. What does show something is what happens when several people in the organization, separately, without consultation, give their own assessment. A CHRO who places the framework at foundation level while a team lead sees it at activation does not point to a measurement error but to an organization in which the framework exists on paper and is not used as intended in practice. That spread is exactly what the plot round was made for: not to calculate an average, but to show where the perceptions diverge and why.

What moving up a level requires

A step up on this dimension rarely calls for a new document. It calls for a fixed moment in the process at which the question is asked, for people who feel free to ask that question, and for a way to learn from what happens when the framework falls short. How much that costs in time and attention depends on where the organization already stands on the fundamental dimensions; whoever wants to know more about this can find an elaboration at what it costs to raise the organizational structure by one level.

The maturity measurement shows whether the organization can carry an AI application, ethically and otherwise. It says nothing about which part of the work itself is suitable to leave to AI. That question is answered by the work scan from FTE TO AI: it calculates per task which part of the work can be taken over, given the frameworks already in place. The two measurements complement each other, but they do not start from the same question; this page is about carrying, not about taking over.