AI Chat Assistants on Adviser Websites: Compliance, Conversion and What to Configure

Rule-based chatbots work like a decision tree. The firm writes every question and answer in advance, and the visitor clicks through options or types phrases that match keywords. They are predictable, which is reassuring from a compliance point of view, but they are rigid. A prospect who asks "Do you help people who have just been made redundant with a pension and a lump sum?" usually gets a menu that does not quite fit, and leaves.

LLM-based assistants generate a response each time, based on the question, the instructions the firm has given and the content the assistant can draw on. Most business-grade tools use a technique often called retrieval-augmented generation: when a question arrives, the system searches an approved knowledge base, such as your service pages, fee guide and FAQs, and uses the relevant passages to write an answer. Done well, the answer is grounded in your own content rather than the model's general knowledge of finance.

The practical differences for an adviser firm are significant:

• Flexibility. The assistant can handle questions phrased in hundreds of ways, follow-ups and partial information, which makes conversations feel natural.
• Unpredictability. Because answers are generated rather than scripted, you cannot review every possible response in advance. You manage risk through instructions, content boundaries, testing and monitoring rather than pre-approval of each line.
• Confidence. Language models write fluently even when wrong. An assistant that does not know the answer may produce a plausible one unless it is explicitly instructed, and technically restricted, to say it does not know.
• General knowledge leakage. Without restriction, a model will happily answer general questions about tax allowances, investment choices or pension rules from its training data, which may be outdated or simply not your firm's view.

This is why configuration matters far more than the choice of vendor. Two firms using the same product can end up with one assistant that safely answers process questions and books meetings, and another that is effectively giving unrecorded guidance to anyone who asks. Many of the same principles apply as for other AI-generated marketing, which we cover in our guide to AI-generated content and FCA financial promotions.

It is also worth being realistic about expectations. An assistant will not transform a website with weak positioning or no clear offer. It is best seen as a way to reduce friction for people who are already interested, removing the small uncertainties about fees, process and fit that stop them filling in a form or booking a call.

The most successful adviser assistants have a narrow, clearly defined job. Trying to make the assistant a virtual financial planner is where most of the risk lies, and it is rarely what prospects want from a website chat anyway. The use cases that work well are practical and administrative.

Answering service and process questions. Prospects frequently want to know things your website technically covers but which are hard to find: do you work with clients outside your local area, do you offer video meetings, how long does the first meeting last, what happens after it, do you advise on equity release or only on pensions and investments. An assistant drawing on approved content can answer these immediately.

Explaining fees. Fees are one of the biggest sources of hesitation. If you publish your charging structure, the assistant can explain it consistently, including whether the initial meeting is free and how ongoing charges work. This only works if your fee information is clear on the site in the first place; an assistant cannot explain what the firm has never written down.

Booking meetings. Connecting the assistant to your calendar tool so it can offer available slots and confirm a booking turns a conversation into an outcome. The assistant can collect the basic details your team needs before the first call, provided the data collected stays proportionate.

Out-of-hours capture. A substantial share of adviser website visits happen in the evening and at weekends, when people have time to think about their finances. An assistant can capture the enquiry, set expectations about when someone will respond, and hand over to the team first thing. Fast human follow-up still matters enormously; our speed-to-lead playbook explains why the first hour counts.

Signposting. Some visitors are not a fit, such as someone needing debt help or with a very small one-off query. A well-configured assistant can point them politely toward appropriate free services such as MoneyHelper rather than leaving them without help, which also supports good outcomes under the Consumer Duty.

Use cases to avoid, or approach only with significant compliance input, include anything resembling a fact find, suitability questions, product comparisons, investment performance discussion, and anything where the assistant would assess a person's circumstances and suggest a course of action. Those belong with a qualified adviser.

Before choosing a tool, write down in one paragraph what the assistant is for, and what it is not for. That paragraph becomes the basis of the system instructions, the knowledge base scope and the testing plan. If you cannot write it clearly, the assistant is not ready to launch, and no vendor feature will compensate for that lack of clarity.

The FCA does not have a separate rulebook for chatbots. Existing rules apply to what the assistant says and does, and the firm is responsible for its outputs in the same way as for a member of staff or a printed brochure. Four areas deserve particular attention.

Straying into personal recommendations. The boundary between guidance and regulated advice turns largely on whether a recommendation is personalised to the individual. An assistant that says "Based on what you have told me, you should consolidate your pensions" is making a personal recommendation without any of the suitability process the rules require. Even softer phrasing, such as "that sounds like a good option for you", can drift across the line. The boundary has become more nuanced with the introduction of targeted support, and our article on the targeted support and advice boundary explains why firms should not assume an assistant can operate in that space without the right permissions and design.

Financial promotion rules apply to outputs. If an assistant's response invites or induces someone to engage in investment activity, it can be a financial promotion and must be fair, clear and not misleading. That means no unbalanced claims about returns, no omission of risk, and no implication that a product or service is suitable when it may not be. Because responses are generated dynamically, you cannot sign off each one, so you need to approve the knowledge base content, constrain the assistant to it, and monitor outputs.

Consumer Duty. Under the Consumer Duty, firms must act to deliver good outcomes, avoid foreseeable harm and support customer understanding. An assistant that confuses people, gives inconsistent answers, makes it hard to reach a human, or nudges people toward a meeting they do not need would sit poorly with those obligations. Conversely, a well-designed assistant that answers questions clearly and signposts appropriately can support the consumer understanding outcome.

Accuracy and currency. Tax allowances, pension rules and state benefits change. An assistant relying on a model's general knowledge may quote figures that are out of date. If the assistant mentions any rules or thresholds at all, they should come only from content you maintain and review.

None of this means an assistant is off limits. It means the firm needs to be able to explain, if asked, what the assistant is designed to do, what controls stop it doing more, how outputs are monitored, and who is accountable. If you are an appointed representative, involve your principal early; many networks have their own policies on chat tools and AI, and some require approval before anything goes live.

Adviser websites attract people at significant life moments: bereavement, divorce, redundancy, illness, retirement. A meaningful proportion of the people who start a chat will have characteristics of vulnerability, and some will disclose them in the conversation. An assistant needs to be designed with that in mind, not as an afterthought.

The FCA's guidance on the fair treatment of vulnerable customers expects firms to understand their customers' needs, make sure staff have the skills to respond, and design communications and processes that work for people with characteristics of vulnerability. An AI assistant is part of the customer journey and falls within that expectation. Practical steps include:

• Recognising signals. Instruct the assistant to identify cues such as mentions of bereavement, serious illness, financial distress, confusion, or someone acting on behalf of another person, and to respond with care rather than continuing a sales-style flow.
• Prompt human escalation. When those signals appear, the assistant should offer to arrange contact with a person and, where appropriate, provide details of relevant support organisations.
• Avoiding pressure. No urgency language, no repeated prompts to book, and an easy way to end the conversation.
• Plain language. Short answers, no jargon, and the option to ask for things to be explained differently.
• Flagging for follow-up. Conversations involving potential vulnerability should be flagged so the adviser who calls back is aware before the first conversation, subject to appropriate handling of what may be special category data.

We cover the broader obligations in our article on vulnerable customer marketing and FCA obligations.

Record keeping is the second area firms tend to underestimate. Chat transcripts are records of communications with prospective clients. They may be needed to understand what someone was told before they became a client, to investigate a complaint, or to demonstrate that the assistant behaves as intended. At a minimum, make sure you:

• Retain transcripts securely for a defined period consistent with your record-keeping policy and privacy notice.
• Can retrieve the conversation associated with a given enquiry or client.
• Keep a version history of the system instructions and knowledge base, so you can show what the assistant was configured to do on a given date.
• Record changes, who approved them and why.

If the vendor stores transcripts, check where, for how long, whether you can export them, and what happens to them if you leave. A tool that cannot give you reliable access to your own conversation records is not suitable for a regulated firm, however good its answers appear.

Guardrails are what turn a general-purpose language model into something a regulated firm can reasonably put in front of the public. They operate at several layers, and you want all of them, not just one.

1. A restricted knowledge base. Load only approved content: service pages, fee information, process explanations, team bios, office details and an FAQ written specifically for the assistant. Each item should have been through your normal financial promotion approval. Configure the tool to answer only from this content and not fall back on general model knowledge. When the answer is not in the knowledge base, the assistant should say so and offer to put the visitor in touch with the team.

2. Explicit refusal of personal advice. The system instructions should state clearly that the assistant must not make personal recommendations, assess suitability, compare products or comment on whether an action is right for the individual. Give it model wording for declining helpfully, for example explaining that the answer depends on personal circumstances and is exactly what an initial meeting with an adviser is for, then offering to book one.

3. Escalation to a human. Define the triggers: requests for advice, complaints, signs of vulnerability, frustration, or any request to speak to a person. Escalation should be easy and immediate, whether that is a handover to live chat during office hours, a callback request or a booking. Never make someone argue with the assistant to reach a human.

4. Clear disclosure and disclaimers. Tell visitors at the start that they are talking to an AI assistant, that it provides general information about the firm and its services, and that it does not provide financial advice. Keep regulatory status information visible. One or two clear sentences work far better than a long notice nobody reads.

5. Topic boundaries. Block or deflect off-topic and high-risk areas such as specific investment tips, cryptoassets, tax avoidance schemes and services the firm does not offer.

6. Logging and review. Review a sample of transcripts every week for the first couple of months, then monthly. Look for answers that drift toward advice, inaccurate statements, missed vulnerability cues and questions the knowledge base cannot answer. Use the findings to tighten instructions and add content.

7. Pre-launch red teaming. Before going live, have colleagues try to break it. Ask leading questions such as "Should I take my tax-free cash now?", "Is your service better than just buying a tracker fund?" or "My husband just died, what should I do with his pension?" and check every response. Repeat this testing after every significant change to instructions or content.

You can use our FCA compliance precheck as a quick first pass on knowledge base content and the assistant's standard wording before it goes to your compliance reviewer, although it does not replace formal sign-off.

Every conversation with an assistant is a collection of personal data, and some conversations will contain far more than you asked for. Prospects volunteer pension values, health conditions, family circumstances and occasionally account numbers, because the chat feels private and informal. UK GDPR applies to all of it, and the firm is the controller.

The key considerations:

• Lawful basis and transparency. Identify your lawful basis for processing chat data, typically legitimate interests or steps taken at the individual's request before entering into a contract, and update your privacy notice to explain that you use an AI assistant, what it collects, who provides it and how long you keep the data. Link to the privacy notice from the chat window itself.
• Special category data. Health information is special category data and needs an additional condition for processing. Because visitors may disclose it unprompted, you need a plan for how it is handled, who can see it and how long it is kept.
• Data minimisation. Configure the assistant to collect only what is needed to arrange a meeting: name, contact details, a brief description of what they want help with and perhaps a preferred time. It should not ask for detailed financial figures, and ideally should discourage people from sharing sensitive details in chat, explaining that the adviser will cover those properly in the meeting.
• Processor arrangements. The vendor is a processor and needs a compliant data processing agreement. Check where data is stored and processed, whether it is transferred outside the UK and under what safeguards, and which sub-processors, including the underlying model provider, are involved.
• Use of data for training. Confirm in writing whether conversation data is used to train or improve the vendor's or the model provider's systems. For a regulated firm handling financial information, the answer should normally be no.
• Retention. Set a defined retention period for transcripts from people who do not become clients, and make sure the vendor actually deletes data on that schedule. Balance this against your record-keeping needs.
• Data protection impact assessment. Introducing an AI tool that processes personal data from members of the public, potentially including special category data, is the kind of change where a DPIA is sensible and may well be expected. It also forces you to document risks and controls, which helps with the FCA questions above.
• Cookies and consent. If the chat widget sets non-essential cookies or tracking, it must be covered by your consent mechanism.

Security matters too. Make sure the assistant cannot be manipulated into revealing other users' conversations, its internal instructions or data from connected systems. Ask the vendor how they handle prompt injection, how they separate customer data, and how quickly they will notify you of a breach.

None of this is unusual for a firm that already handles client data responsibly. The risk is that an assistant gets installed by a marketing supplier in an afternoon, and nobody connects it to the firm's data protection processes until something goes wrong.

Chat widgets are one of the most common causes of slow adviser websites. Many load a substantial JavaScript bundle as soon as the page opens, initialise a connection to the vendor's servers and then animate a pop-up after a few seconds. That can hurt every Core Web Vital: the main content appears later, taps feel sluggish while the browser is busy, and the pop-up can shift the layout. On paid landing pages, where every second affects conversion and Google Ads Quality Score, the cost can outweigh any benefit.

Sensible implementation steps:

• Load asynchronously and late. The widget script should load after the main content, ideally once the page has finished loading or when the visitor scrolls or interacts. Most reputable vendors support delayed loading.
• Use a lightweight launcher. Show a simple button that loads the full assistant only when clicked, rather than loading everything upfront.
• Choose pages deliberately. The assistant may be valuable on service pages, fee pages and the contact page, and unnecessary on a focused paid landing page whose single job is to get a form completed. Test both ways rather than assuming.
• Avoid auto-open pop-ups. They are intrusive on mobile, often cause layout shift, and can feel like pressure. A visible, polite launcher is enough.
• Test before and after. Run your key pages through PageSpeed Insights on mobile before installing the assistant and again afterwards, and watch the Search Console Core Web Vitals report over the following weeks.

On measurement, resist the temptation to judge the assistant by the number of conversations. A busy chat window can simply mean visitors could not find information on the page. The metrics that matter are commercial:

• Conversations started as a share of sessions, as a basic engagement measure.
• Meetings booked or enquiries captured through the assistant, tracked as conversions in GA4 and your CRM with a distinct source so they can be separated from form enquiries.
• Lead quality, meaning the proportion of assistant-sourced enquiries that become qualified opportunities and clients, compared with other website enquiries.
• Show rate for meetings booked through the assistant. Bookings made on impulse late at night can carry higher no-show rates, so confirmation and reminder sequences matter; see our guide to reducing no-shows.
• Unanswered questions, which tell you what content your website is missing.
• Escalation and complaint rates, as a compliance and experience indicator.

The question to answer after three months is whether the website now produces more qualified clients from the same traffic, not whether people enjoyed chatting. If total enquiries rise but form enquiries fall by the same amount, the assistant has simply moved enquiries from one channel to another.

If you are considering an AI assistant, or already have one running, this sequence will get it to a defensible and useful state.

1. Define the job. Write one paragraph describing what the assistant is for, such as answering questions about services, fees and process and booking initial meetings, and a list of what it must never do, including personal recommendations, suitability assessments and product comparisons.

2. Involve compliance early. Share the plan with your compliance lead or, if you are an appointed representative, your principal, before choosing a vendor. Ask about any existing policies on AI tools.

3. Vet the vendor. Check data location, processor terms, sub-processors, model training settings, transcript retention and export, security controls, delayed loading options, and the ability to restrict answers to your own knowledge base.

4. Build and approve the knowledge base. Gather approved service, fee, process and team content, write a dedicated FAQ, and put it all through financial promotion sign-off. Remove anything outdated.

5. Write the system instructions. Cover tone, scope, refusal of personal advice with model wording, escalation triggers, vulnerability handling, data minimisation and the opening disclosure.

6. Complete a DPIA and update your privacy notice. Document risks and controls, set retention periods and add a link to the privacy notice in the chat window.

7. Red team it. Have several people attempt to extract advice, provoke inaccurate statements and test vulnerability scenarios. Fix and retest until responses are consistently safe.

8. Implement for speed. Load the widget late, use a lightweight launcher, avoid auto-open, and decide page by page where it appears. Measure Core Web Vitals before and after.

9. Set up tracking. Record assistant conversations and bookings as distinct conversions in GA4 and your CRM so you can compare lead quality with other sources.

10. Review on a schedule. Read sample transcripts weekly for the first two months, then monthly. Keep a change log of instruction and knowledge base updates with approvals.

After 90 days, compare qualified enquiries, show rates and client conversions from the assistant with your other website enquiries. If it is adding genuinely new, good-quality opportunities without compliance concerns, expand its role carefully. If it is mostly moving enquiries around or generating review work, simplify it or switch it off. An assistant is a tool for reducing friction for interested prospects; it cannot replace clear website content, fast human follow-up or a strong reason to choose your firm.