You can’t outsource your accountability: agentic AI amplifies the principal-agent problem

Andre Fredericks, COO, Sanlam Studios, Venture Builder

Picture a financial adviser who once spent an afternoon drafting a client recommendation. Now an AI system produces a polished version in ninety seconds. The adviser scans it, tweaks a paragraph, and sends it.

The output looks the same. The signature is the same. The regulatory obligation is identical.

What has changed is what sits behind the signature.

This is not a technology story. It is an accountability story. And it begins long before AI was invented.

An old problem in new clothes

Adam Smith identified it in 1776. Managers entrusted with other people’s money, he observed, could not be expected to watch over it with the same vigilance as their own. Two centuries later, Jensen and Meckling formalised it: the principal-agent problem.

The setup is simple. A principal — a shareholder, a client, a policyholder — needs something done. They cannot do it themselves. They delegate to an agent: a CEO, a financial adviser, an underwriter. The moment they do, three problems emerge: the agent knows things the principal does not; the agent has their own interests; and the principal cannot fully observe what the agent is actually doing.

This gap between what the principal wants and what the agent does is agency cost. And it is, in some form, everywhere.

The insurance industry knows this well. We built entire architectures around it. Treating Customers Fairly. Fit and proper requirements. Record of advice obligations. The FSCA market conduct frameworks from Kenya to Ghana. Every one of these frameworks is, at its root, a response to the same underlying problem: how do you ensure the agent acts in the principal’s interest when you cannot watch them at all times?

All of it rests on one assumption so obvious it was never written down.

The agent can be held to account.

The foundation

Agents are legal subjects. They can be contracted with, monitored, incentivised, debarred, sued, and sanctioned. The adviser who gives reckless advice loses their licence. The underwriter who misprices risk faces consequences. The board that fails to oversee answers to a regulator.

Accountability is the foundation. The contracts, the compliance frameworks, the professional indemnity structures. All of it is built around the fact that the agent is a human being, or at a minimum, a legal person, capable of bearing consequences.

Nobody questioned this because there was nothing to question. Of course, the agent is a person.

What else would it be?

The silicon agent

Today, anyone can deploy an agent by whispering words into a machine.

No contract. No onboarding. No fit and proper assessment. No legal relationship. In minutes, an AI agent can draft client communications, analyse policy documents, process claims data, or generate advice summaries acting on behalf of the professional who deployed it.

The word “agent” is doing new work here. An AI agent is not an agent in the regulatory or economic sense. That is precisely the problem.

When a human agent fails, we have recourse. Professional indemnity. Debarment under the FAIS Act. Civil liability. Regulatory sanction. When an AI agent fails, when a model hallucinates a policy exclusion, misrepresents a product feature, or produces subtly biased advice at scale, there is no recourse.

You cannot debar a neural network.

The accountability did not disappear. The regulator’s logic does not allow it. It has to land somewhere. And when the agent is an AI system with no legal standing, no licence to lose, no reputation to defend, and no assets to seize, it lands on you.

Friction was the feature

Here is what makes this moment genuinely new.

The principal-agent problem always involved friction. Appointing an agent was slow, costly, and contractual. That friction was inconvenient, but it was also a handbrake. It made you think before you delegated. It made the delegation legible. There was a paper trail, a relationship, a human on the other end who could be questioned.

Now the friction is gone. You delegate by speaking. The agent acts in seconds.

Ease of use creates an illusion of safety. The tool feels controllable because it responds to plain language. But plain language is the loosest possible specification. You have handed significant discretion to a system built by people who will openly tell you they do not fully understand it either.

And then something goes wrong. A client receives incorrect information. A claim is assessed on flawed logic. An advice document misrepresents risk. The question arises: who is responsible?

Not the model. Not the model provider. Not the training data.

You. The licence holder. The FSP. The professional who deployed the tool and signed the output.

The gap is yours

The frameworks will catch up. They always do. In the meantime, the gap is yours to manage.

Treat AI output the way a senior partner treats a junior associate’s draft: useful, often impressive, never signed without reading. Decide in advance which decisions you will not delegate, regardless of how capable the tool becomes. The line should be clearest where the consequences are hardest to reverse. Keep a record of what you asked, what the system produced, and what you changed. When something goes wrong, reconstruction falls to you.

Most importantly: notice when speed is doing your thinking for you. The ninety-second draft is a problem, not a solution, if it gets signed in ninety-one seconds.

You can outsource your thinking. You can outsource your doing. You cannot outsource your accountability.