Accountability: Governing and Assuring Delegated Decisions

Delegating a decision does not delegate responsibility for it. When an agent approves a claim, prices a contract or acts on a customer's behalf, the organisation still answers for the outcome: to regulators, to courts, to customers and to its own board.

This is not a new principle. Organisations have always been responsible for what their employees, contractors and systems do. What is new is the scale, speed and autonomy of the delegate, and the difficulty of supervising decisions that people no longer make directly.

Part V of The Age of Agency asks how to govern those decisions. This essay sets out the argument.

Oversight at machine speed

The traditional model of oversight assumes that people can review decisions. A manager checks a sample of an employee's work; an auditor tests a sample of transactions; a committee reviews exceptions each month.

That model strains when an agent makes thousands of decisions an hour. Reviewing after the fact finds problems only after the damage has been done. Reviewing every decision in advance removes the benefit of the agent altogether.

Good oversight of agents therefore shifts its weight from review towards three other kinds of control.

Limits set in advance. Hard boundaries on what an agent may do: values, counterparties, data, systems and types of action. Within the limits, it acts; outside them, it stops or escalates.

Monitoring as it happens. Continuous observation of behaviour, looking for patterns no single rule would catch: drift from normal behaviour, actions near the limits, unusual timing, signs of manipulated inputs.

Intervention that works. The ability for anyone accountable to pause or stop an agent immediately, tested in advance rather than discovered in a crisis, and a way to reverse or contain what it has already done.

Human review does not disappear. It moves to where it adds most: approving decisions above thresholds, investigating alerts, sampling outcomes and deciding when an agent's authority should change.

Every agent needs an identity and an owner

One of the simplest controls is also one of the most neglected. Every agent should have its own identity in the organisation's systems, so that its actions can be traced to it, and a named person who answers for it.

Without an identity, an agent's actions blur into those of the people or systems it acts through, and nobody can reconstruct what it did. Without an owner, problems fall between teams: the technology function says the business requested it, the business says the technology function built it, and the vendor says it was used outside its instructions.

An agent without an owner is an agent nobody is accountable for. No organisation should allow one to act on its behalf.

What financial audit teaches

Financial reporting faced a similar problem long ago. Organisations reported their own results, and stakeholders needed a reason to believe them. The answer was independent assurance: an auditor, with professional standards and its own liability, who tests the evidence and gives an opinion.

AI will need its own equivalent. Organisations will be asked to show that their agents behave as intended, and "the vendor says so" will not be an adequate answer. Assurance will need independent evidence: tests of performance across representative cases, reviews of controls, examination of records and checks that limits are enforced in practice.

Several principles carry over directly from audit. Those who build a system should not be the ones who assure it. Assurance should rest on evidence that can be examined. And the records needed to provide that evidence must be created at the time decisions are made, not reconstructed afterwards.

The direction of regulation

Regulation is moving in the same direction, though at different speeds in different places.

The European Union's AI Act, adopted in 2024, requires that high-risk AI systems be designed so that people can oversee them effectively, understand their limitations and intervene or stop them. In Singapore, the Model AI Governance Framework for Agentic AI, published in January 2026, emphasises bounding risks upfront, keeping humans meaningfully accountable and applying technical controls across an agent's life cycle. The Monetary Authority of Singapore has set out expectations for financial institutions covering board oversight, inventories of AI in use and controls across its life cycle.

The details differ, but the themes are consistent: know what AI you are using, decide deliberately what it may do, keep people accountable, apply controls throughout its life, and keep records that show all of this.

The practical lesson is to build those records and controls now. Organisations that wait for regulation to arrive will find themselves retrofitting controls after an incident, under scrutiny, at greater cost.

What good looks like

An organisation that governs its agents well can answer, for any material agent:

1. What is it, who owns it, and where is it on the delegation map?
2. What limits apply, and how are they enforced?
3. What is monitored, and who receives the alerts?
4. How can it be stopped, and when was that last tested?
5. What records exist of what it has done, and could an independent party examine them?

These are not technical questions. They are governance questions, and they belong on the board's agenda.

What follows

Governance answers how organisations should use agents responsibly. The final part asks a wider question: if capable intelligence becomes a basic input to work and life, who gets access to it, and on what terms?

This essay introduces Part V of the forthcoming book The Age of Agency. References to regulation are summaries of public material and should not be relied on as legal advice.

Views are the author's own. The frameworks described are applied methodologies of Kiros Intelligence Group, not regulatory standards.