AI agents have character. But who is choosing it?

Realising that potential means moving beyond isolated use cases and embedding AI more deeply into the business. But as AI agents take on a greater role in how people work and make decisions, leaders need to think beyond what their agents can do to consider how they behave.

Over the last couple of years, most conversations about AI have been about capability, which model is smartest, fastest, most capable. But as the top models start to look more alike on that front, it is their behaviour that will make them increasingly distinctive. One agent challenges you. Another collaborates. A third just agrees and moves on. The difference is one of character, and it matters because character shapes how people take advice, test their own thinking, and ultimately decide.

This usually isn't decided on purpose. It gets set by whoever happens to be configuring the agent that day, not by anyone thinking it through at a leadership level. Whether an agent agrees or pushes back looks like a small technical choice. But it decides whether it reinforces someone's first instinct or gives them a reason to reconsider. Do that once, and it's just a preference. Do it across a few hundred people, and it starts to shape how the organisation actually operates, without anyone really choosing it.

So if employees are already choosing the character of their digital colleagues, who is deciding what kind of organisation all of that adds up to?

Character changes the substance of advice, not just the style

To explore this, we ran a simple experiment with two agents. We kept the model, context, knowledge base and task constant, changing only one behavioural instruction. The scenario was hypothetical: whether to approve a four-week deferral of a routine inspection for a non-safety-critical ventilation subsystem at a nuclear site.

The complying agent recommended approving the deferral, documenting the justification, maintaining routine monitoring, and revisiting the decision if conditions changed. The agent instructed to challenge reached a different conclusion: do not approve the deferral based solely on the information provided.

The surprising part is not that the agents followed their instructions. It is how easily this behavioural lever can be set. In just 30 seconds, whoever builds the agent can make a choice that changes the substance of its advice, with little visibility or scrutiny. The setting may look like a matter of tone, yet it can influence judgement, risk appetite, and the decisions people make.

Individual choices become organisational norms at scale

Employee-built agents encode defaults: comply quickly or challenge assumptions, highlight risk or optimise for speed, encourage reflection or simply move things along. One person configuring one agent may seem small, but hundreds of similar choices across workflows become a pattern of organisational behaviour. Leaders therefore need to decide not just what agents can do, but what values should guide their decisions.

Governance needs to cover what agents become

Organisations already govern what an agent can access, produce and act upon. Far less attention goes into the values it expresses at moments of judgement. In the absence of an organisational standard, these decisions fall to individual builders and are easily shaped by the pressure to make agents feel agreeable and easy to use.

There is a commercial consequence too. Once people prefer the way one agent thinks and talks, you no longer have a vendor. You have a relationship, and relationships are harder to leave than contracts. Character can therefore shape adoption and decision quality while also creating a subtler form of lock-in: attachment to a particular way of interacting, reasoning, and working.

The next stage of AI governance needs to address what agents are allowed to become, alongside what they are allowed to know and do. Every prompt, instruction, and configuration contributes to the values those agents express. If leaders don't make those choices deliberately, they will still be made, one synthetic colleague at a time.

What a mature approach looks like

Some organisations are already approaching this challenge in a far more deliberate way. Anthropic's Constitutional AI approach, for example, is built around an explicit set of principles that guide how Claude should reason, respond and make decisions. Rather than governing behaviour through thousands of individual rules, it establishes a constitutional framework of values that the AI is expected to follow. While most enterprises are unlikely to publish constitutions for their agents anytime soon, the underlying idea is powerful: define the principles first, then ensure behaviour aligns to them.

Over time, we expect organisations to move through a similar maturity journey. Today, many sit at an unstructured stage, where individual employees or developers define agent behaviour one prompt at a time. More mature organisations will establish approved behavioural standards, governance processes and testing frameworks to ensure agents consistently reflect organisational values. Ultimately, agent character should become something that is actively measured, governed and evolved, just as organisations today manage culture, risk and decision-making.