The biggest danger is not that AI will replace financial professionals. It is that AI will industrialise the financial industry’s existing assumptions—making incomplete decisions faster, cheaper and at enormous scale.
Finance can count people without understanding them
Finance has always had a dangerous habit: treating whatever can be measured as though it were the whole truth. Income, expenditure, credit history, loan-to-value and investment risk scores matter. But they describe a financial position; they do not fully explain whether that position is sustainable for the person living inside it.
Artificial intelligence makes this tension impossible to ignore. A model can process years of transactions in seconds, identify patterns no adviser could manually spot and calculate thousands of scenarios. That is valuable. But speed does not turn a partial picture into a complete one.
Consider two mortgage applicants with the same salary, deposit, declared expenditure and proposed monthly payment. One maintains a cash reserve, reviews spending and would reduce discretionary costs after an income shock. The other is expecting childcare costs, regularly supports relatives and knows that a restrictive budget causes anxiety and impulsive spending. A conventional affordability calculation may treat them as identical. Their lived affordability is not identical.
The question is therefore bigger than, “Can AI calculate what this person can borrow?” The real question is, “Can it help determine what this person can sustain without pretending that human behaviour is another clean data field?”
Adoption is moving faster than understanding
The industry is already well beyond the experimental stage. A 2024 joint Bank of England and Financial Conduct Authority survey of 118 firms found that 75% were using AI and another 10% planned to do so within three years. More than half of reported use cases involved some degree of automated decision-making.
The most revealing figure, however, was not adoption. Forty-six per cent of respondents said they had only a partial understanding of the AI technologies they used, compared with 34% reporting complete understanding. A financial institution cannot meaningfully explain a customer outcome if it does not fully understand the system influencing it.
Boards should worry about that gap more than they celebrate another efficiency statistic. An unexplained accurate answer can still become an indefensible decision when it affects somebody’s home, pension, protection or access to credit.
The problem is not bad mathematics
Most harmful financial decisions are not caused by a calculator adding incorrectly. They arise because the question was too narrow, the data lacked context or the model inherited yesterday’s assumptions.
Historic financial data can reflect historic exclusion. Thin credit files, career breaks, irregular income, shared family responsibilities and non-standard working patterns do not automatically indicate poor financial discipline. Yet a system trained to reward conventional financial histories may convert difference into risk without ever using an explicitly discriminatory variable.
That is how bias becomes harder to see. Nobody needs to instruct a model to disadvantage a group. They only need to define “normal” too narrowly and optimise around it.
The answer is not to reject AI. It is to stop confusing prediction with judgement.
A three-layer model for responsible financial AI
Financial businesses need to separate three questions that are too often compressed into one score.
First is financial capacity: verified income, committed expenditure, debt, assets, contractual obligations and quantified stress testing. AI is exceptionally useful here because it can extract information, detect inconsistencies and run scenarios with greater speed and consistency.
Second is behavioural resilience: how the customer manages uncertainty, responds to financial pressure and prioritises competing goals. This should shape the conversation, recommendations and support provided. It must not become a secret personality score used to deny access. If badly designed, behavioural finance becomes behavioural surveillance.
Third is explainable human judgement: the ability to show what influenced an outcome, what assumptions were made, where professional judgement changed the result and how the customer can challenge it. A human signature at the end of an automated process is not meaningful oversight. The human must have enough information, authority and time to disagree with the machine.
Use AI to widen the conversation
The best near-term use of AI in advice is not replacing the adviser at the point of recommendation. It is helping the adviser notice more, test more and communicate better.
AI can highlight unexplained spending movements, compare a client’s stated priorities with their actual cash flow, model the impact of redundancy or childcare and present the same information in different formats. That last point matters. Some clients understand a graph immediately; others need a written sequence, shorter sections or time to process information away from a meeting. Personalisation should improve access to understanding, not manipulate the customer towards a product.
The model should generate better questions, not merely faster answers.
Six tests every financial AI project should pass
Before deploying AI into a material customer journey, leadership teams should be able to answer six questions:
- What customer outcome are we improving, beyond reducing cost or handling time?
- Are we using the system to support a recommendation or quietly determine eligibility?
- Which customers are least represented in the training and testing data?
- Can a customer understand the decisive factors without technical knowledge?
- Can a qualified human genuinely override the output, and is that override monitored rather than punished?
- Who remains accountable when the model, vendor, data and human reviewer all contributed to the decision?
If a firm cannot answer those questions, it is not ready to automate the outcome regardless of how impressive the model appears in a demonstration.
The competitive advantage will be knowing where automation stops
The winners in financial AI will not necessarily be the businesses with the most automated advice. They will be the firms that know precisely where automation adds clarity, where it introduces risk and where a human conversation changes the meaning of the data.
AI can calculate affordability in milliseconds. Good advice decides whether affordable is liveable. It can identify a pattern, but it cannot assume the pattern explains the person. It can recommend an action, but responsibility for that recommendation cannot disappear into a model.
The future of finance should not be human versus machine. It should be machines doing more of the calculation so humans can do more of the understanding. Anything less is not intelligent finance. It is simply automated assumption.
Source: Bank of England and Financial Conduct Authority, “Artificial intelligence in UK financial services – 2024”, published 21 November 2024.
