AI has the potential to re-shape consumer journeys so that they become more autonomous. Over time, the Review expects that using AI consumers will create personal agents to manage their finances, rather than using firm-provided tools. This could re-shape the relationship between firms and consumers.
The Review identifies an emerging appetite for automation (although trust will be a key factor in its adoption). By 2030 consumer journeys may increasingly start with AI. Consumers may begin a journey with an AI agent that helps them understand their needs, compare the market, and act on their behalf, rather than searching and comparing manually.
Over time, consumers’ use of AI is likely to shift from using AI as a helper (operator role) → working with AI on decisions (collaborator role) → AI recommending/comparing while the consumer decides (consultant) → AI carrying out tasks like switching, debt, savings or investment management within agreed limits, with the human as approver or, rarely, only an observer.
Risks to be aware of
- Consumer capability itself may change – while AI has the potential to support consumers with complex financial decisions, repeated delegation could produce “cognitive offloading” and mean consumers become less able to judge AI outputs or act independently. Design choices will need to be made which encourage scrutiny;
- Fragmentation risks – AI could make financial services easier to use but it may also leave others behind. The Review has found that 45% of consumers see no benefit from AI in day-to-day finance and 24% say nothing would encourage them to use it — skewed towards older, lower-income consumers. In addition, the Review found that 11% of AI users already pay for access and AI providers will charge more for access to AI with better capability. This means that the access, adoption and quality of AI may impact customer outcomes. There is the potential for better-resourced consumers to get better AI, better comparisons, and better ongoing management, while others get poorer information, more friction and less ability to benefit from competition. The Review warns that this risks creating a two-tier market;
- Bias and discrimination - these risks may become harder to manage where firms use larger datasets, more complex models and variables that may act as proxies for protected characteristics or characteristics of vulnerability. Some consumers could face higher prices or reduced access where models identify patterns linked to disadvantage rather than legitimate risk;
- Hyper-personalised deceptive design - AI could let firms identify exactly which prompts, framing, or friction are most persuasive for each individual. These could be adapted in real time and deployed at scale. While this could support consumers, it may also allow firms to exploit consumers’ circumstances, behaviours and preferences – for example, adapting in real time to nudge consumers toward a sale, discourage switching, or downplay exclusions; and
- Redress/accountability gap – one of the biggest concerns with the use of AI is who is responsible if things go wrong. Where firms partner with AI model providers to develop consumer journeys, consumers may not know whether responsibility sits with the firm, model provider, platform, or their own agent when something goes wrong. Trust in the use of AI in financial services will depend on consumers being able to understand decisions, challenge outcomes and obtain redress.
What do firms need to be thinking about?
- Design for inclusion, not just capability — avoid an AI-first default that leaves consumers who can’t/won’t use AI (due to disability, low confidence, language, vulnerability) with reduced choice or worse service; non-AI/non-digital routes need to remain genuinely available.
- Distinguish legitimate risk-pricing from proxy discrimination — firms need the ability to explain and justify why a model has priced or excluded a consumer, not just point to model output.
- Scrutinise personalisation design intent — is the interface genuinely serving the consumer, or is it optimised for engagement/conversion/revenue while appearing neutral?
- Build auditability and clear consent points into personalised journeys — because AI can make each consumer’s journey different, transparency becomes harder by default unless deliberately engineered in.
- Clarify accountability across the chain — especially where a firm’s service is delivered through a third-party model or platform (e.g., an AI integration) — consumers need to know who is responsible.
- Maintain auditable records, clear complaints routes, and enough understanding of AI outputs to explain decisions when challenged.
- Prepare for a rise in AI-assisted complaint volumes, including from professional representatives, and ensure genuine harm doesn’t get lost in noise.
- Consider agent-to-agent complaint handling carefully — the Report flags this as plausible (consumer agents raising issues, firm agents triaging) but only workable “if systems are reliable, interoperable, auditable and capable of escalating complex or sensitive cases to people”.
The Review considers an emerging appetite for more automated agent-led journeys where consumers may act as consultants and approvers of recommendations and actions taken by AI agents. This will require firms and regulators to provide a reliable framework and infrastructure for these developments. The paper treats agentic finance as a significant future development by 2030 and beyond. It links it to the broader move along the autonomy spectrum from AI as a tool that assists humans, to AI that can recommend, prepare and eventually execute actions within agreed boundaries. The Mills Review makes a specific recommendation (Priority recommendation 5) for the FCA to develop a trusted framework to enable the foundations for agentic finance.
The Building blocks of agentic finance and considerations for firms
The Review discusses how agentic finance can lead to both the emergence of new products and innovation but also a general boost to productivity and economic growth if agents can be trusted to operate safely in financial services. However, this depends on “foundations” or infrastructure that do not yet fully exist. Annex VII sets out six building blocks for agentic finance to operate properly:
1. Data: This is described as fundamental. Agents will need the following types of data to work effectively:
- Goal data: what the agent is trying to achieve, for whom, and why.
- Observation data: what the agent can see about the consumer, products, prices, markets and counterparties.
- Outcome data: what happened as a result of the agent’s actions, including whether the intended objective was achieved.
As a result, data quality will be essential and firms will need systems that ensure the data used by agents is accurate, complete, timely, consistent and traceable. Otherwise, firms and consumers will not be able to rely on decisions made.
2. Identity: Agentic finance will require trusted identity arrangements so that firms know who authorised an AI agent and what authority the agent has. This includes digital identity for the human principal, so firms can verify who is granting authority, and Agent ID as a distinct, verifiable identity for the AI agent itself, so firms can tell who is acting, on whose behalf, and with what authority. The Report does not propose an approach to digital identity and recognises that the UK does not currently define agent identity, but notes that the market is beginning to set verification standards for particular scenarios (such as agentic commerce and enterprise identity). The Review considers this may lead to a fragmented approach where standards are not interoperable and could cause poor consumer outcomes.
3. Authorisation and delegation: Consumers must be able to grant agents bounded, revocable, verifiable authority to act within defined limits. The paper emphasises that current consent models are too tied to point-in-time human approval and do not yet support ongoing delegated action across multiple firms and products. Currently, the UK does not have a standardised framework through which this can happen.
4. Payments / execution: Payments are the mechanism that allows an agent actually to complete financial activity. The Review says existing payments infrastructure and rules are still geared to human approval, so agentic finance requires payment rails and execution frameworks that can support trusted agent-initiated transactions. The Strategy for Future Retail Payments Infrastructure recognises the potential benefits of programmable payments and tokenised forms of money.
5. Liability attribution: There must be clarity on who is responsible when an AI agent causes loss or harm. Without a workable liability framework, firms are likely to require human approval, which will limit autonomy.
6. Supervision and audit at scale: Regulators and consumers will need to understand what an agent did, on what authority, and with what result. The Review notes that at the individual level this can resolve consumer complaints but at the aggregate level, it provides the FCA with a supervisory mechanism to detect systemic patterns. Firms will therefore need end-to-end audit systems capable of recording the customer identity, the agent identity, the mandate in force at the time, the data relied on, the instruction or reasoning path followed, the action taken, the result and any challenge or escalation.
