Agentic AI is coming for financial services, here is why mortgage advice should not go first

There is a difference between an AI system that recommends something and an AI system that acts on it. Most of the AI already in UK financial services, including in mortgage sourcing and case management, sits in the first category, it proposes, and a person decides. Agentic AI is the shift toward the second category, systems that carry out multi step tasks on their own, chasing a document, adjusting a case, moving money, without a person approving each individual step along the way. That shift is genuinely underway. The question worth asking is not whether it arrives, it clearly is, but which part of a mortgage business should be the first place it gets real autonomy, and the honest answer is that mortgage advice itself is a poor candidate to go first.

Where the current momentum actually is, and it is not mortgage advice

FCA chief executive Nikhil Rathi told an industry audience in June 2026 that more than 80 per cent of financial services firms are already using or adopting AI, and that the policy conversation has moved on from whether firms adopt it to how they deploy it at scale. The retail examples he gave were telling, smarter bill management, personalised investment strategies, reduced friction in everyday transactions. These are useful, low stakes, largely reversible tasks. None of them was mortgage advice.

The Bank of England's Financial Policy Committee reached a similar conclusion from a different angle in its April 2026 record. It found that financial system participants had not yet adopted more advanced agentic AI in a way that presented systemic risk, but noted that risk was likely to grow, and specifically asked the Bank and the FCA to do further work on agentic AI focused on payments and financial markets. HM Treasury's Financial Services AI Adoption Plan, published in July 2026, follows the same pattern, one of its five priority areas is explicitly agentic payments readiness. The parts of the system regulators are moving fastest on are the ones where a single action is small, repeatable, and correctable. A payment that goes wrong can usually be reversed, reissued, or refunded. A mortgage recommendation that has already been acted on cannot be undone the same way.

Why mortgage advice sits at the wrong end of the risk profile to go first

The stakes attached to a single mortgage decision are simply larger and slower moving than most of what agentic systems are currently being trusted with. A mortgage affects a customer's home and their finances over decades, not a single transaction. Once a case has been submitted, funds released, and a purchase completed, unwinding a wrong recommendation is not a quick correction, it is a much harder and often much more damaging process for the customer involved.

The variability of individual cases compounds this. Income type, credit history, property, and personal circumstance combine differently every time, which is exactly why so much of the real judgment in mortgage advice sits in the edge cases rather than the standard scenarios. A system operating with genuine autonomy has to get that judgment right without a person checking it first, in a domain where the judgment itself is rarely straightforward.

The legal ground underneath all of this is also still being mapped. The UK Jurisdiction Taskforce published a legal statement on liability for AI harms on 7 July 2026, concluding that existing common law is broadly flexible enough to handle many AI related legal questions, but identifying real gaps that still need addressing, including whether product liability law applies to standalone AI software, and how to handle situations where harm has clearly occurred but negligence cannot easily be established. Handing a system full autonomy over decisions with this much individual consequence, while the underlying liability framework is still being worked out, is a harder position to defend than doing the same thing somewhere the consequences of a mistake are smaller and easier to correct.

The regulator's own language points the same way

Rathi was direct about this in the same June 2026 speech, stating that accountability for regulated activities and their outcomes must remain clearly assigned regardless of the degree of automation involved, and that customers may be reluctant to delegate important decisions to systems they do not understand, which makes human oversight central to maintaining confidence rather than optional to it.

Industry itself is not settled on how to apply that principle in practice. In a Bank of England roundtable with banks and insurers in February 2026, firms raised concerns about whether traditional model risk management approaches can scale to widespread agentic AI deployment, and questioned how the idea of a human in the loop can be meaningfully applied once a system starts taking on more of the actual decision making itself. That is an open question the industry has not answered yet, not a solved problem being rolled out cautiously.

The FCA's Mills Review offers useful language for thinking about where mortgage advice should sit while that question remains open. It sets out an autonomy spectrum running from operator, where a system acts largely on its own, through collaborator and consultant, to approver and observer, where a human retains the final decision. Mortgage advice is a strong candidate to stay toward the approver end of that spectrum for some time, a system that gathers, prepares, and proposes, with a person approving before anything is actually acted on, rather than a system that simply acts.

Where agentic capability actually makes sense first

None of this means agentic AI has no place in a mortgage business. It means the sequencing matters. The lower stakes, higher volume, more reversible parts of the process are the sensible place to start, and there are several of them. Document chasing and collection is one, a system that follows up an outstanding payslip or prompts a client for a missing document carries a small cost if it gets the timing wrong, a delay, not a wrong recommendation. Case administration is another, pre populating a form from data that has already been verified, or flagging a case for human review once it meets a specific criteria threshold, rather than deciding the outcome itself. Pipeline management fits the same pattern, prompting a client ahead of a maturity date is a reminder, not a decision, and getting a reminder wrong is a far smaller failure than getting a recommendation wrong. Compliance flagging works the same way too, a system that reviews a file for a missing vulnerability adaptation note or an incomplete fair value rationale and raises it for a person to look at is operating as an approver, not as the decision maker.

This is the model behind how Mortgage Magic™'s case management and compliance monitoring tools are built to use AI, surfacing, preparing, and flagging for a person to act on, rather than acting independently on the recommendation itself. That is not a limitation on what the technology could eventually do. It is a reflection of where the genuine risk in this market actually sits, and where it does not.

The order matters more than the destination

Agentic AI is coming for financial services, and mortgage businesses that ignore it will fall behind the ones that do not. But the industry's own current momentum, the regulator's own language, and the state of the underlying liability law all point toward the same sequencing. Earn the operational trust in the parts of the process where a mistake is small and reversible first. Treat the advice step itself as the last place full autonomy should be extended without a person still approving before anything gets acted on.