Agentic AI in Banking: What DIFC Regulation 10 Means for Governance, Risk and Deployment

AI adoption among DIFC-regulated firms rose from 33% to 52% in a single year, according to the Dubai Financial Services Authority's 2025 AI survey, with generative AI use up 166% over the same period. The same survey found that while 70% of firms using AI reported having a formal governance framework in place, a meaningful share still lack clear accountability for how that AI is actually overseen. Rapid adoption and an accountability shortfall are sitting side by side across the same set of firms. Regulation 10 of the DIFC Data Protection Regulations speaks directly to that shortfall, and it matters more every time a bank hands an AI system a task with real consequences attached.

What Regulation 10 actually covers

Regulation 10 governs how personal data is processed by what it calls a "System": any machine-based system operating autonomously or semi-autonomously that processes personal data and generates outputs on that basis. That covers a wide range of AI, from a model that flags a suspicious transaction to an agent that drafts a credit memo.

The regulation's central mechanism is a hierarchy of purpose. A System may only process personal data for purposes a human has defined or approved, or for purposes the System generates itself within human-defined principles and constraints, and where both exist, the human-defined purpose takes precedence. If a system infers a new way to use client data that nobody explicitly authorised, that inference still has to stay within the human-defined principles and limits set for it in advance.

For processing that counts as high-risk, which can include handling significant volumes of personal data in ways that increase risk to individuals, firms must also appoint an Autonomous Systems Officer, broadly equivalent in status and role to a Data Protection Officer, and obtain certification from a DIFC-accredited body. Deployers also have to be able to explain a system's outputs to the people affected by them, in plain language and backed by evidence, rather than by pointing to a model card or a vendor's documentation.

None of this bans autonomous systems. It requires that the organisation using one can show what it was told to do, what it decided on its own within those limits and why.

Governance has to be built into the system

A policy document that says "our AI systems must operate within approved limits" satisfies almost nobody's actual risk. The limits have to be built into the system itself, checked as part of how it runs, rather than filed away as a commitment nobody re-reads once the system is live.

This is where the DFSA's own numbers are worth sitting with. The survey found governance frameworks in place at 70% of firms using AI. Fewer have clear, working accountability for what those frameworks actually catch. A framework that exists on paper but isn't wired into how a lending agent, a trading tool or a client-facing assistant actually behaves leaves a firm with very little to point to if a regulator, or a client, asks how a specific output was controlled. "The policy said not to" is not an explanation if nothing in the system enforced the policy.

Where this gets concrete: agentic banking

Consider a bank using an AI system to support a loan restructuring decision. The system reviews the client's payment history, current exposure and market conditions, and produces a recommendation. If a human reviews that recommendation before it becomes a decision, the system's error tolerance is relatively forgiving. A wrong recommendation gets caught.

Now consider the same system with more autonomy: instead of recommending a restructuring, it initiates one within a pre-approved threshold. Under Regulation 10, the bank needs to be able to show that the threshold, and the logic for staying inside it, was human-defined, rather than inferred by the system from patterns in past approvals. If the system started applying a slightly different risk tolerance to a particular client segment because that's what the historical data seemed to suggest, and nobody explicitly authorised that shift, that raises a Regulation 10 question regardless of whether the outcome happened to be reasonable.

This is the practical shape of the distinction the regulation draws. It isn't asking whether an autonomous system is a good idea. It's asking whether the organisation deploying it can point to the specific human decision behind every boundary the system operates within, and explain any output that falls inside those boundaries.

What the 2026 consultation adds

On 18 June 2026, DIFC opened a 30-day public consultation, Consultation Paper No. 3 of 2026, proposing amendments to the Data Protection Regulations. The proposals aim to sharpen the certification regime, clarify the duties of the Autonomous Systems Officer role and introduce a new Regulation 11 giving the Commissioner formal power to recognise accreditation and certification schemes. Jacques Visser, DIFC's Chief Legal Officer, described the goal as keeping the framework "practical, clear and able to respond" as AI use develops.

This is a proposal open for comment. It hasn't come into force, and it shouldn't be read as loosening the existing requirements. If anything, it signals the opposite: DIFC is treating Autonomous Systems Officer accountability as important enough to spell out further rather than leave to interpretation.

What this means for banks building agentic AI now

Waiting for the consultation to conclude before acting on Regulation 10 isn't an option, since the underlying regulation already applies. Three things are worth checking now, regardless of how the consultation lands.

First, whether the organisation can name, for each autonomous or semi-autonomous system in use, exactly which purposes and limits were human-defined versus system-generated. If nobody can answer that without opening the model's training documentation, the answer isn't ready.

Second, whether high-risk processing has an Autonomous Systems Officer attached to it with real visibility into what the system does, rather than just a name on an org chart.

Third, whether the organisation could produce a plain-language explanation of a specific output on short notice, backed by evidence, for a client or a regulator asking why the system did what it did.

Tools that generate a recommendation for a person to act on sit at a different point on this spectrum than agents empowered to act directly on the client relationship. The more autonomy a system is given, the higher the practical stakes of meeting Regulation 10's requirements become, and the governance work that matters most is the work that shows up in the system's actual behaviour, rather than the policy binder sitting beside it.

AinTools works with financial services firms on secure, sovereign AI deployment and governance. Our Secure & Sovereign AI page sets out how we approach it.

Sources

• Dubai Financial Services Authority, "New DFSA AI survey: Generative AI adoption has nearly tripled within the DIFC in last 12 months" (12 November 2025): dfsa.ae
• DIFC, Regulation 10 of the Data Protection Regulations (official source, DIFC Commissioner of Data Protection): difc.com – Regulation 10
• DIFC, official announcement of Consultation Paper No. 3 of 2026 (18 June 2026): difc.com – Consultation announcement
• Legal summaries of Regulation 10's operative requirements (Lexology / Mayer Brown)