TL;DR: Agentic AI does not automatically count as regulated financial advice in the UK. The FCA regulates activities rather than technologies. Whether an AI interaction crosses the regulatory perimeter depends on what the system does, how personalised its output is, whether it recommends a specific course of action and whether it goes on to arrange or execute it.
The issue is getting harder to separate in practice. AI agents can already compare products, use detailed information about an individual, make personalised recommendations and act on a consumer’s behalf.
The FCA’s July 2026 Mills Review calls this emerging category “advice-like support”: highly personalised financial support from AI models that sits outside the current regulatory perimeter but may look very similar to regulated advice. It recommends that the FCA examine how the advice and guidance boundary, arranging rules and financial promotion rules work in continuous AI-enabled journeys.
The debate now has two legitimate sides. More regulatory clarity could protect consumers and create a level playing field between regulated firms and general-purpose AI providers. Draw the boundary too widely or too early, however, and firms may struggle to use AI to provide cheaper and more accessible financial support.
For a broader view of all seven recommendations, see Aveni’s Mills Review hub and our analysis of what the Mills Review means for financial services firms.
Does agentic AI count as regulated financial advice?
The short answer is sometimes, depending on what the agent is doing.
The FCA’s regulatory framework is activity-based. Using AI does not create a separate category of financial advice, and using an AI interface does not remove an activity from regulation.
For investment advice, FCA guidance says a personal recommendation can include a recommendation about a particular investment that is presented as suitable for an individual or is based on that person’s circumstances. Those circumstances can include income, objectives, needs and risk appetite. The medium used to deliver the advice does not determine whether it falls within the definition.
That creates a fairly clear position at the extremes.
An AI tool that explains what an ISA is is providing information.
An AI system operated by an authorised firm that considers a customer’s circumstances and recommends a particular investment as suitable may be carrying out an activity already covered by financial services regulation.
The difficult area sits between those two examples.
That is where agentic AI is developing fastest.
Why agentic AI makes the advice boundary harder to apply
Most digital financial journeys used to contain reasonably identifiable stages.
A consumer searched for information. They compared products. They spoke to an adviser or made their own choice. They then completed a transaction.
An AI agent can combine those stages in a single conversation.
A consumer could tell an agent:
“I want to get a better return on my savings, but I need access to the money next year.”
The agent might then:
- analyse the consumer’s finances
- identify relevant products
- compare their features
- decide which ones appear most suitable
- recommend one
- initiate or complete the transaction
- continue monitoring the consumer’s position afterwards
The FCA describes this development as a move towards financial services that are more continuous and delegated, with AI systems progressing from providing recommendations to taking actions within agreed parameters. Its consumer research found that one in five UK adults is already open to AI making decisions within goals they have set.
The regulatory question therefore becomes more complicated than deciding whether one chatbot response constituted advice.
Firms may need to determine which regulated activities are taking place across an entire agent-led journey.
The FCA now has a term for the grey area: “advice-like support”
One of the most important concepts in the Mills Review is advice-like support.
The Review defines it as highly personalised support from frontier AI models that currently falls outside the FCA’s perimeter, but which would be treated as regulated financial advice if it fell within that perimeter.
This matters because general-purpose AI can increasingly use an individual’s own information to generate highly specific financial recommendations.
The Mills Review highlights three problems.
1. Personalisation can blur the line between guidance and advice
The FCA’s current framework already distinguishes between different levels of financial support.
Under the Advice Guidance Boundary Review, for example, targeted support allows authorised firms to make suggestions to groups of consumers with common characteristics. Simplified advice can go further where a consumer needs a recommendation based on their individual circumstances.
AI changes what can happen between those categories.
A general-purpose model may know considerably more about one person than a traditional guidance service does. It can combine their income, holdings, goals, transaction history and preferences and produce a highly individualised answer.
The Mills Review specifically warns that these hyper-personalised recommendations may deliver advice-like support outside the regulatory perimeter.
2. Similar financial influence can carry different regulatory obligations
This creates a potential level-playing-field problem.
A regulated wealth manager may face strict requirements when making personalised recommendations.
A general-purpose AI platform may influence the same consumer’s financial decision without carrying equivalent obligations if the activity falls outside the perimeter.
The Mills Review calls out the risk of regulatory arbitrage, where unregulated AI platforms exert comparable influence to regulated firms without equivalent consumer protections.
HM Treasury’s July 2026 Financial Services AI Adoption Plan reaches a similar conclusion. It recommends a review of financial guidance and advice-like outputs from general-purpose LLMs, including their impact on consumer outcomes, competition and the level playing field between regulated and unregulated providers.
3. Agents can move from recommending to acting
A recommendation is only one part of the issue.
An agent might also submit an application, move money, arrange a transaction or route a customer towards a particular provider.
The regulatory perimeter then extends into other areas.
The Mills Review therefore recommends examining:
- advice and guidance in conversational AI interfaces
- the “by way of business” test for general-purpose AI
- financial promotions
- arranging rules
- continuous AI-enabled consumer journeys
This is why the debate around agentic finance is wider than whether a chatbot has technically given advice.
The relevant question is increasingly what financial activity has the agent performed on the consumer’s behalf?
Where might an AI agent move closer to regulated advice?
There is no single sentence or feature that automatically determines the answer. Context matters.
This simplified spectrum shows where the regulatory questions become more significant.
| AI behaviour | Example | Main regulatory question |
|---|---|---|
| Information | “This ISA pays 4.2% and allows three withdrawals a year.” | Is the agent presenting factual information? |
| Comparison | “Product A has a higher rate. Product B offers more flexible access.” | Is the agent comparing products or beginning to influence a choice? |
| Personalised support | “Based on your plans to buy a house next year, access to your cash appears important.” | How far has the agent used individual circumstances to steer the consumer? |
| Recommendation | “Based on your circumstances, Product B is the best option for you.” | Does this amount to a personal recommendation or another regulated activity? |
| Execution | “I’ve submitted the application for Product B.” | Do arranging, payment or other regulatory requirements apply? |
| Continuous management | The agent monitors the consumer’s finances and changes products when conditions change. | How should regulation apply when recommendations and actions happen continuously? |
These examples are illustrative. The exact regulatory treatment depends on the product, activity, provider and circumstances.
The important point is that agentic AI allows a system to move through several of these stages without handing the consumer from one clearly defined service to another.
The argument for clearer regulation
There is a strong consumer protection case for clarifying this boundary.
Consumers may reasonably assume that two services offering highly personalised financial recommendations provide similar protections.
That may not be true.
The Mills Review found that around 26% of consumers trust general-purpose AI tools such as ChatGPT, Claude or Gemini for financial advice, while awareness of the difference in formal routes to redress can be limited.
As AI becomes more convincing and more personalised, consumers may find it increasingly difficult to distinguish regulated advice from an AI-generated recommendation that sits outside the perimeter.
Clearer rules could therefore help consumers understand:
- who is responsible for the recommendation
- whether the provider is authorised
- what standards apply
- what happens if the recommendation causes harm
- whether the consumer has access to formal redress
It could also give regulated firms more certainty about what they can build.
That certainty matters when firms are committing significant time and money to AI systems that may eventually interact directly with customers.
The argument against drawing the boundary too widely
There is an equally important argument on the other side.
AI has the potential to make financial support available to people who currently receive very little of it.
The FCA’s Advice Guidance Boundary Review was created partly because many consumers cannot access affordable financial advice. Targeted support is intended to give firms more room to help customers make decisions while retaining appropriate protections.
Agentic AI could extend that further.
A well-governed system could help consumers understand complicated products, identify relevant options and manage routine financial decisions at a much lower cost than traditional advice.
Treating every personalised AI output as regulated advice could make those services harder or more expensive to provide.
The government’s Financial Services AI Adoption Plan therefore makes the balance explicit. It supports examining advice-like outputs and the regulatory perimeter while warning against unnecessarily constraining innovation and the consumer benefits AI could provide.
The challenge for regulators is to distinguish useful financial support from activity that carries the influence and risk of regulated advice.
The Mills Review does not propose a new AI rulebook
This distinction is important.
The Mills Review does not conclude that agentic AI requires an entirely separate regulatory regime.
The FCA has repeatedly said that it intends to rely on existing frameworks, including the Consumer Duty, SM&CR and established governance requirements, while providing further clarity where AI creates practical problems.
The Mills Review’s first recommendation is therefore to secure and adapt the regulatory perimeter.
It recommends that the FCA consider conducting a review within three to six months of the Review’s publication on 6 July 2026. That work would examine general-purpose AI used for savings, investments, pensions, mortgages and debt management, then determine whether the FCA should change its guidance, recommend changes to government or retain the current approach.
That is a recommendation from the Mills Review rather than a confirmed new rule or deadline for firms.
For the full set of recommendations and their practical implications, read The Mills Review findings: what the FCA decided and what compliance teams should do now.
What regulated firms should be doing now
Firms do not need to predict exactly where the FCA will eventually draw every boundary.
They do need to understand what their AI systems are already doing.
Map the activity, rather than the technology
Start with each customer-facing use case.
Document:
- what information the AI receives
- what decisions it makes
- whether it compares specific products
- whether outputs are personalised
- whether it presents an option as suitable
- whether a person approves the recommendation
- whether the system can initiate an action
- which regulated activities may be relevant
“AI chatbot” or “agent” is too broad a classification for regulatory analysis.
Map the level of autonomy
The Mills Review describes an autonomy spectrum running from AI as a tool through to AI acting continuously while a human monitors outcomes.
Higher autonomy creates more difficult questions around accountability, consent, auditability and redress.
Firms should know where each system sits on that spectrum today and what additional permissions it may receive later.
Test the difficult cases before deployment
The highest-risk behaviour may occur near the boundary.
An agent might provide information during most conversations but generate a highly personalised recommendation under a particular combination of prompts and customer data.
Testing should therefore cover realistic customer journeys, including edge cases and situations where the agent is pushed towards recommendation or execution.
Aveni explored this problem directly through the FCA’s Supercharged Sandbox. Read what the FCA Supercharged Sandbox revealed about AI agent governance in practice.
Keep evidence of what the agent actually did
A policy describing what an AI system is supposed to do cannot show what happened in a specific customer interaction.
Firms need retrievable evidence of:
- the customer context available to the system
- the output it generated
- the action taken
- relevant approvals
- the rules and controls applied
- whether the behaviour remained within expected limits
- the resulting customer outcome
This becomes increasingly important when agents can make many decisions at speed.
Our practical guide covers five areas firms should be able to evidence when deploying AI agents.
Keep accountability attached to a person
Greater AI autonomy does not remove senior management responsibility inside regulated firms.
The Mills Review expects existing accountability frameworks to remain important as AI use grows, with more clarity needed around the reasonable steps senior managers should take as delegation increases.
For a closer look at that issue, see AI agents and the SMCR: who’s on the hook when the bot gets it wrong?.
The bigger question is what job the AI is doing
The regulatory debate around agentic AI is often framed as a debate about technology.
The more useful way to approach it is to look at the activity.
An AI system can explain.
It can compare.
It can personalise.
It can recommend.
It can arrange.
It can act.
As financial AI moves through those stages, the regulatory questions change with it.
The Mills Review recognises that existing boundaries may become harder to apply when general-purpose AI produces highly personalised financial recommendations outside regulated firms, or when one agent-led interaction combines information, recommendation and execution.
Clarifying that boundary could give consumers better protection and firms greater certainty.
The difficult part will be doing so without removing the opportunity to use AI to make useful financial support available to many more people.
That is now one of the most important regulatory questions in agentic finance.
Explore Aveni’s Mills Review resources and practical guidance for financial services firms.
FAQs
Does agentic AI automatically count as regulated financial advice?
No. Agentic AI is a type of technology, while the FCA’s perimeter is based primarily on activities. The regulatory treatment depends on what the AI does, the product involved, how personalised the interaction is and how the recommendation or action is presented.
Can AI give regulated financial advice in the UK?
AI can be used as part of a regulated advice service. Using AI does not remove the regulatory requirements that apply to the underlying activity. FCA guidance also makes clear that the medium used to communicate investment advice does not determine whether that communication falls within the relevant definition.
What is “advice-like support”?
The Mills Review uses “advice-like support” to describe highly personalised support from frontier AI models that falls outside the FCA’s regulatory perimeter but would be treated as regulated financial advice if it fell within the perimeter.
What is the difference between targeted support and regulated advice?
Targeted support allows authorised firms to make suggestions designed for groups of consumers with common characteristics. Advice can involve a recommendation based on an individual’s specific circumstances and needs. The FCA introduced targeted support as part of its wider work on the Advice Guidance Boundary.
Why are AI agents creating a new problem for the advice-guidance boundary?
AI agents can combine information, comparison, personalisation, recommendation and action in one continuous interaction. The Mills Review warns that this can make it harder to identify where advice-like support becomes a regulated activity and where responsibility sits.
Is the FCA changing the rules on AI financial advice?
No new AI-specific advice regime has been announced. The Mills Review recommends examining whether current guidance and the regulatory perimeter remain appropriate as agentic AI develops. It leaves open the possibility of revised guidance, recommendations for perimeter changes or retaining the existing approach.
When will the FCA review the regulatory perimeter for agentic AI?
The Mills Review recommends that the FCA consider carrying out a review within three to six months of its 6 July 2026 publication. The recommendation covers general-purpose AI used in areas including savings, investments, pensions, mortgages and debt management.
What should financial services firms do now?
Firms should map what each AI system actually does, identify potentially regulated activities, understand the level of autonomy, test customer journeys before deployment and retain evidence of the system’s behaviour and resulting outcomes. Existing FCA expectations around governance, Consumer Duty and senior management accountability continue to apply where relevant.
