Ask a financial institution today whether it uses AI responsibly, and it can usually answer with confidence – supported by model risk frameworks, fairness testing, and explainability sign-off.

Yet if you ask what it's prepared to let an AI Agent decide without a human-in-the-loop, that confidence tends to run out.

That's the harder question Agentic AI asks of compliance leaders in 2026 – not ‘can we use AI responsibly’, but ‘what decisions and actions are we prepared to delegate to an Agent, under what conditions, and how do we stay accountable if it makes a mistake?’

It's not a semantic distinction. Governance for AI that produces an output looks very different from governance for AI that takes an action. Most institutions are already moving along a spectrum – from AI that assists a human, to AI that recommends a course of action, decides, or acts on a decision without waiting for sign-off.

The further along that spectrum an Agent operates, the more specific the governance question moves from ‘is this AI safe in general?’ to ‘is this particular decision one we're prepared to hand over, and on what terms?’

That's a harder question to answer in the abstract, so it's worth breaking into pieces. Before extending an AI Agent's authority on any given decision, five questions are worth asking.

1. What decision is this, exactly?

Not ‘what does the AI do’, but what specific judgement is being made – closing an alert, filing a report, releasing a payment. A vague response makes every question after it harder to answer.

This distinction matters more in financial crime compliance than most functions realise, because ‘the AI handles alerts’ describes a process, not a decision. 

Closing a low-risk name-screening match, drafting a case narrative, and releasing a flagged payment are three entirely different judgements with three different risk profiles – yet institutions often describe all three under one banner, as if delegating ‘screening’ were a single decision, rather than dozens of distinct ones. 

Naming the decision precisely, rather than the system performing it, is a discipline everything else depends on.

2. What's the cost of getting it wrong?

An incorrectly closed low-risk alert and a missed sanctions match don't carry the same consequence, and shouldn't attract the same level of control. Materiality, not the presence of AI, should set the bar.

This is becoming one of the more demanding governance requirements across jurisdictions.

Singapore's proposed Guidelines on AI Risk Management explicitly propose that governance scales with an application's materiality and the autonomy granted to it, rather than applying uniform controls regardless of stakes [1].

The Financial Stability Board's proposed sound practices similarly support a proportionate approach, with the materiality of a particular AI use informing the type and intensity of governance and controls applied [2].

The question that matters is therefore not simply whether AI is involved, but what happens if the decision is wrong.

3. Who can overrule it, and how fast?

Escalation only works as a safeguard if it's fast enough to matter before the decision has consequences downstream.

This is where governance stops being theoretical and becomes a design question. An escalation path that exists on paper but takes days to trigger doesn't function as a safeguard for a decision an Agent can act on in seconds.

Effective delegation therefore requires escalation boundaries to be built into an Agent's operating parameters: defining when it can proceed, when additional evidence is required, and when a decision must stop and wait for human judgement. 

This is the approach built into Iris 7, where those boundaries can be calibrated to the risk and materiality of the decision at hand.

4. What evidence will exist afterwards?

If a regulator, auditor, or investigator asks why a decision was made six months later, is there a trail that answers that – not just a log of an output, but a reconstructable rationale?

This is quietly one of the more demanding requirements emerging across jurisdictions. The EU AI Act's high-risk provisions place significant requirements around technical documentation, record-keeping, transparency, and human oversight [3]. 

The FSB's proposed sound practices similarly emphasise governance and risk management throughout the AI lifecycle, including as institutions adopt more complex forms of AI such as Agentic AI [2].

An audit trail that shows what happened without showing why is not evidence – it's a record, and the two answer very different questions when concerns are raised later down the line.

5. Who owns the outcome?

Not which team built the model – who owns the outcome if the decision turns out to be wrong. If that answer is unclear, the Agent probably isn't ready for that decision yet.

The UK's Competition and Markets Authority has made this principle explicit in its guidance on AI Agents: businesses remain responsible for what their Agents do, just as they are responsible for the actions of employees [4]. 

That accountability cannot simply dissolve into ‘the system decided’. Institutions still need clear ownership of delegated decisions – and if that ownership cannot be established before authority is delegated, the governance model isn't ready.

The test, not the answer

These aren't sequential steps, and they're not a maturity model to work through once and move past. They're a working test institutions should be able to apply, decision by decision, to every judgement they consider handing to an Agent – not settled once for AI as a category and left alone.

As Agentic AI takes on more of what compliance, risk, and operational teams do, that discipline – not simply the sophistication of the model – will determine which decisions institutions can responsibly delegate, and which should remain with people.