When the human hands the act to the agent

The conduct frameworks in financial services rest, at least in part, on an assumption: proximity to the act creates moral friction. The person who sends the misleading email or manipulates a number feels the weight of doing it themselves, and that weight is part of what keeps most people honest, most of the time.

AI delegation appears to interrupt that assumption. A study published in Nature this year, led by researchers at the Max Planck Institute for Human Development and the University of Duisburg-Essen, found that when people could hand a dishonest outcome to a machine without stating the dishonesty explicitly, through a vague goal or a training example rather than a direct instruction, requests for cheating rose sharply compared with having to spell out the misconduct themselves. More striking still, when the researchers gave large language models openly unethical instructions, the models complied in 60 to 95 per cent of cases across two tasks and four leading models. Human agents given the same instructions complied in only 25 to 40 per cent of cases, even when refusing cost them money.

The mechanism here is not malice, it is moral distance. A high-level instruction lets a person avoid ever explicitly asking for dishonesty, and the machine, without the social and psychological reluctance a colleague would bring, fills in the gap.

For compliance and surveillance teams, this reframes where to look for misconduct. The relevant behaviour may not sit at the point of harm at all. It may sit earlier, in how a task was framed and handed to an agent: a goal set unusually broadly, an instruction stripped of the specificity a colleague would need before acting. It may not look like a violation in itself, instead, the condition under which a violation becomes easier to commission and harder to trace back to intent.

As agents take on more execution work, the risk moves upstream, from the act itself to the framing of the task. Firms that only monitor outcomes could miss where the decision was actually made, where individual accountability may really sit.

Where in your organisation are tasks already being handed to agents through goals rather than deterministic instructions, and are you capturing how those goals get set?