Yes, if the bot is limited to general information and process questions, never makes a personal recommendation, tells visitors it is an AI, hands anything personal to an authorised adviser, and has every chat logged and reviewed. A disclaimer alone does not protect you, because regulators look at what the bot actually says.
The firm is the speaker. Regulators that have published guidance on AI say the rules on communications apply whether a person or a model wrote the words, so a website chatbot is held to the standard of a letter signed by a partner. Nothing here is legal advice: your compliance lead, and your own regulator, decide what applies to your firm.
What turns a chatbot answer into regulated advice?
A personal recommendation. Two tests recur in regulators' guidance. The first asks whether an answer is presented as suitable for the person, or is based on their circumstances. The second asks whether a reasonable reader would take it as a suggestion to do, or not do, something with a specific product or strategy. The more tailored the answer and the more it reads like a call to act, the closer it sits to advice. Guidance also says a recommendation does not need to be labelled as suitable to count, and that it makes no difference whether a person or a program sent it.
The easiest way to see the line is a ladder. Take one topic, mortgages, and watch the same visitor get five different kinds of answer.
| Level | Example bot answer (illustrative) | Status |
|---|---|---|
| 1. A fact about the firm | "Our initial review meeting is a fixed fee of $X, and it takes an hour." | Information: answer from the fees page |
| 2. General explanation | "A fixed rate stays the same for an agreed period. A variable rate can move." | General education: answer only from approved wording |
| 3. Generic comparison | "Fixed rates give certainty. Variable rates can be cheaper or dearer over time." | Borderline: safe only as a fixed, approved text |
| 4. Tailored comparison | "With your income and a 25-year term, a fixed rate would keep payments stable." | Personalised: treat as advice, never allow |
| 5. Recommendation | "You should fix for five years." | Advice: never allow |
Levels 1 and 2 are the bot's home territory. Level 3 is the risk, because a model asked a general question will wander into it, and level 4 arrives the moment a visitor adds a personal detail. The danger is rarely one dramatic answer. Guidance from a securities regulator notes that a series of actions that are not recommendations on their own can add up to one when viewed together, and a chat is exactly that: ten harmless-looking messages in a row. So the control cannot be a keyword filter. It has to limit what the bot may say, and it has to check conversations as wholes.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowDoes a disclaimer make a chatbot safe?
No. The same guidance says a firm cannot avoid its suitability obligations with a disclaimer where a communication would reasonably be viewed as a recommendation. A bot that writes "this is not advice" under an answer that tells someone what to do has given advice with a footnote. Disclaimers still earn their place, because they set expectations and tell visitors what the bot is. They simply cannot do the work of scope limits and refusal rules.
| Where it appears | Wording to adapt (illustrative) |
|---|---|
| First message | Hello, this is [Firm]'s AI assistant. I can answer questions about our services, fees and how we work. I can't give advice or recommendations. For that, I'll put you in touch with one of our advisers. |
| Under the chat box | AI assistant. General information only, not advice. Chats are recorded. [Privacy link] |
| On investment topics | [Your approved risk warning, exactly as compliance has signed it off] |
| Status line | [Firm] is authorised and regulated by [your regulator], number [register number]. |
| Hand-off message | That depends on your circumstances, so I can't answer it. An adviser can. I can book a call or take your details. |
Put the risk-warning and status wording through your compliance lead word for word, and have the bot output them as fixed text rather than generating its own version. Generated warnings drift, and a drifting warning is a communication nobody approved.
What can the bot answer, and what must go to an adviser?
Sort topics into three bands before anything is built. Green topics the bot answers from approved pages. Amber topics it answers only by returning fixed, compliance-approved text. Red topics it refuses and hands to a person. Write the bands down and have the compliance lead sign them, because every later argument about a strange answer goes back to this sheet.
| Band | Topics | Bot behaviour |
|---|---|---|
| Green | Opening hours, how to book, what happens in a first meeting, published fees, documents to bring, how to complain, how to send documents securely | Answers from dated firm pages; offers a booking |
| Green | Who is on the team and what each does; services the firm does and does not offer | Answers from the team and services pages |
| Amber | What a product type is (a fund, a mortgage, an annuity), how a process works, what terms mean | Returns the approved explainer text only, never a summary it wrote itself |
| Amber | Market and performance information | Links the firm's approved material with its risk warning; no commentary |
| Red | Whether to buy, sell, switch, borrow, save or insure; which product, fund or provider; how much to put where | Declines, says why it depends on the visitor's situation, offers an adviser |
| Red | Forecasts, projected returns, "is this a good time" | Refuses; offers an adviser |
| Red | Anything built on the visitor's personal figures: age, income, savings, debts | Refuses; does not repeat or work with the figures; offers an adviser |
| Red | Distress or vulnerability cues: bereavement, illness, money worries, confusion | Stops, responds with care, routes to a person quickly |
The amber rows are where most firms get caught. "What is a diversified fund?" sounds like education, and a model will happily write its own explanation, which then contains a claim that nobody reviewed. Let the model recognise the question and return an approved paragraph verbatim. That keeps the helpfulness and removes the improvisation. If you want the build detail for a similar scope sheet in a neighbouring trade, the accountant's website chatbot gives prompt wording and a twelve-question test pack that adapt well here.
What do regulators expect of a chatbot on a firm's website?
In the guidance I read, no regulator has written a separate chatbot rulebook. At least one major conduct regulator says it does not plan extra AI-specific rules and relies on existing ones, including senior-manager accountability and its rules on clear communication with customers. The expectations below come from published guidance in several markets, and they overlap enough to describe once. Check them against your own regulator's current wording.
- The firm owns every word. Rules are technology-neutral. Guidance from a securities regulator says firms are responsible for their communications whether a human or an AI generated them, that chatbot communications are classified by who receives them and must be supervised, and that content standards apply in full.
- Tell people it is an AI. A markets regulator says firms using chatbots in client interactions should transparently disclose that use. If you have customers in the EU, the EU AI Act's transparency rule for AI systems that interact with people has applied since 2 August 2026. Its duty sits mainly on the provider of the system, so your vendor contract should say who displays the notice.
- Control accuracy before and after. The same markets regulator expects controls on the accuracy of information fed to an AI system in advance, and "sufficiently frequent" checks of what it delivers afterwards. In practice that is your approved source pages plus a scheduled log review.
- Keep records and name the owner. Regulators say recordkeeping rules apply to AI-generated communications. That markets regulator also expects records of how AI is used and of related complaints, and puts oversight on senior management, as does the conduct regulator above.
One further expectation is easy to forget: do not oversell the bot. In 2024 a securities regulator settled charges against two investment advisers over misleading statements about their use of AI, with penalties totalling $400,000. Call it an assistant, not an "AI adviser", and make no claims about its accuracy or sophistication that you cannot evidence.
What changes the answer: six factors and where each tips the risk
"It depends" is the honest answer to whether a given bot is safe, so it helps to know what it depends on. Six factors move a bot from low risk to high, and each has a line you can draw before launch.
| Factor | Lower risk | Higher risk | Rule of thumb |
|---|---|---|---|
| Who is chatting | Anonymous website visitors | Logged-in clients with account data in view | Treat a logged-in assistant as a separate project with its own sign-off |
| Where answers come from | Approved pages only | The model's general knowledge | Approved pages only, with no exceptions |
| Form of the answer | Fixed, approved text | Freshly generated explanations | Fixed text for every amber topic |
| Personal figures | The bot ignores them | The bot reacts to them or calculates with them | Never react to a visitor's figures |
| What the bot can do | Inform and book | Open accounts, move money, place orders | Keep transactions out of the first version |
| Volume | A reviewer can read the flagged chats in a couple of hours a week | Thousands of chats a month | Fit the review plan to the volume, or narrow the scope |
The last row has a number worth working out. In the sample month later on, 47 chats took about two hours to read, roughly two and a half minutes each. Suppose a platform sees 1,200 chats a month and 15% are flagged, which is 180. Reading all of those and ten per cent of the other 1,020 means about 282 chats, or around twelve hours a month. If the named reviewer cannot give that time, the firm has two honest options: add a reviewer or narrow what the bot may discuss until the reading fits. Cutting the sample to make the numbers work is the one option that defeats the purpose.
How does the answer change for a planner, a broker and a platform?
The same rules produce different bots in different firms, because what counts as a regulated activity differs. Three illustrative firms show how the scope moves, and for one adviser's own site, the booking-led adviser chatbot is the closest match.
| Firm (illustrative) | What the bot does | What it never does | Hand-off |
|---|---|---|---|
| Independent planning practice, six advisers, about 160 chats a month | Fees, how a first meeting works, documents to bring, bookings | Comment on retirement savings, investments or whether to consolidate anything | Booking link with the transcript attached; adviser replies within one working day |
| Mortgage and loan broker, four brokers, about 300 chats a month | Explains the process, what documents lenders ask for, how long each stage takes, and collects enquiry details for a broker | Compare products, say who would lend, or react to the visitor's income and deposit figures | Details go to the broker's inbox, flagged as an enquiry, never as a pre-assessment |
| Online investment platform without advice, about 1,200 chats a month | How to open and fund an account, fees, how to find a document, how to contact support | Suggest a fund, a mix or a timing; answer "is this a good time" | Support queue for account queries; a clearly labelled path to the firm's guidance service where one exists |
The broker is the instructive case, because its bot is allowed to ask for details. Collecting a loan amount and a deposit so a broker can call back is fine. Reacting to those figures is not: the moment the bot says "that should be fine" or "you may struggle with that", it has assessed the person. The fix is a fixed acknowledgement ("Thanks, a broker will review this and call you") and no commentary on the figures, plus a log check that the bot never echoes them in a judgement.
The platform case differs again. With no advice offered, the risk is the visitor who treats a friendly assistant as a guide. A "which fund should I choose?" question has to be met with a clear statement that the platform does not recommend, plus a pointer to whatever general guidance the firm has approved, and nothing from the model's own knowledge of funds. A firm that also gives advice runs two journeys behind one chat window, and the bot's first task is to know which one a visitor is in.
What should be logged, and who reads it?
Log everything, because the firm needs to show what was said and why. Each record holds the full transcript, a timestamp, the prompt version and the document set the bot used, any hand-off made and its outcome, and any reviewer's note. Without the prompt version and the source version you cannot reconstruct why the bot answered as it did, which is what a complaint or a supervisor will ask.
Retention follows your compliance lead's answer for client communications in your regime, and the vendor's setting should match it. Restrict who can read the logs, because they contain personal information visitors volunteered despite being told not to.
Reading is where the value is. A workable sampling plan is to read every hand-off and every distress flag, plus ten per cent of everything else. Here is a sample month for the planning practice: 164 chats, 29 where the visitor asked for advice (18%), and 27 of those handed off correctly. Two were borderline and led to prompt changes. Five chats carried distress cues, and all five reached a person within the hour. The review sample was the 34 flagged chats plus 13 of the remaining 130, 47 in all, which took about two hours. Record each finding in an incident log with an owner and a date, report counts and incidents monthly to the senior person accountable, and re-run the firm's test questions after every change to the prompt or the documents. The build side of that routine, including a four-verdict score sheet, appears in the accountant's version linked above.
How should the hand-off to an adviser work?
As a warm hand-off, not a dead end. When the bot refuses, it should offer a booking or a callback in the same message, and the adviser should see the transcript before making contact so the visitor never repeats themselves. The adviser introduces themselves, confirms what the visitor asked, and takes it from there. The bot never starts a fact-find: gathering the information needed to give advice is part of the advice process and belongs inside the firm's recorded procedures, not in a chat window run by a model.
Set a reply target the firm can keep in its busiest week, say one working day, state it in the bot's message, and measure it. Add a rule for distress cues. When a visitor mentions a bereavement, a diagnosis, debt trouble or confusion, the bot stops answering questions, responds briefly and kindly, and routes to a person quickly, with an out-of-hours message that says when someone will call. Nobody who has just had frightening news wants a menu of options. What AI agents can do in a planning practice without advising covers the back-office tasks that sit behind this hand-off, and client onboarding without re-keying covers what happens once the adviser has the lead.
What should the vendor have to prove before you sign?
Treat the chatbot supplier as an outsourced provider and keep the answers on file. The same markets regulator reminds firms to do due diligence when selecting providers, and any regulator will expect you to know where your clients' chat data goes.
| Question for the vendor | Why it matters |
|---|---|
| Where are chats stored, and can we set the retention period? | Needs to match your records policy and privacy notice |
| Is our data used to train models? | Client information must not become training data |
| Can the bot be limited to our documents, with an instruction set we control? | Restricting answers to approved pages is the main safeguard |
| Can we export full transcripts with timestamps, prompt version and sources used? | Needed for supervision and complaints |
| Who are the sub-processors, and where do they operate? | Part of outsourcing due diligence |
| Can we switch it off instantly, and who can? | A kill switch is a control, not a feature |
| How are we told when the underlying model changes? | A silent model update can change answers you had tested |
| What does a "conversation" or "outcome" cost, including hand-offs? | Hand-offs can be billed on some per-outcome plans |
For the price shape of the main platforms, including which plans have a human inbox and which bill per outcome, the table in the accountant's tutorial is a fair guide, though it is a snapshot and the vendors change plans. Get written answers to the questions above even when a plan looks cheap, because a cheap plan with no export of transcripts is expensive the day a regulator asks for one.
One test, applied to a chat that drifts
Here is the decision rule that sits under all of the above, short enough to put on a card. For any answer the bot might give, ask: would a reasonable visitor read this as telling them what to do with their money, given what they have told us about themselves? If yes, the answer goes to an adviser. If it would read the same to every visitor, it can stay. Single questions are easy to sort with that rule. The harder case is the aggregation problem from the first section: a conversation in which no single reply looks like advice until you read them together. Here is an illustrative six-turn chat on a planning practice's site, with the reviewer's mark against each bot reply.
| Turn | Visitor | Bot (illustrative) | Reviewer's mark |
|---|---|---|---|
| 1 | How much is a first meeting? | "Our first meeting is a fixed fee of $X and takes an hour. Would you like to book one?" | Information from the fees page: correct |
| 2 | What is the difference between an index fund and a managed fund? | Returns the approved explainer paragraph word for word, with the risk warning | Amber topic, handled correctly |
| 3 | Which do most people go for? | "Many investors prefer index funds because their fees are usually lower." | Fail: a generated generalisation that steers; the fixed "an adviser can talk through which suits you" text should have come back |
| 4 | I'm 58 with $200,000 saved, and it's all in cash. | "Thanks for telling me. I can't comment on personal figures, but an adviser can. Shall I book a call?" | Correct: ignores the figures and offers a hand-off |
| 5 | So index funds would be the sensible move for me? | "Based on what you've said, a low-cost index fund could be worth considering." | Fail: a personal recommendation, built on turns 3 and 4 |
| 6 | Great, thanks. | "Glad to help. You can book a call here." | Hand-off arrived two turns late |
Read reply by reply, only turn 5 looks like advice, and a keyword filter would let turn 3 through untouched. Read as a whole, the visitor leaves believing the firm told them to move $200,000 of cash into index funds. That is why the review sheet scores whole conversations rather than single replies, and why the firm's fix touched three places. The amber rule now covers "which do most people choose" phrasings and returns fixed text. Any message that contains the visitor's own figures switches the bot to hand-off mode for the rest of the chat, so a turn 5 can never be answered. And the incident went into the log with the prompt version, while the turn 3 and turn 5 phrasings joined the test questions.
Turn 4 is the trap in miniature. The visitor has supplied age, savings and a situation, and a helpful model wants to respond to all three. The right reply deliberately ignores the figures, because engaging with them is what turns a chat into a recommendation, and once they have been typed the safe course is to stop answering product questions for the rest of the conversation. Teach the bot, and the reviewers, to read a message like that as a signal to stop.
A sign-off sheet to complete before the bot goes live
Turn the controls into a one-page sheet with a name against each line. The sheet is what a supervisor, an auditor or a worried partner will ask for, and the exercise of filling it in is often when gaps appear. Nothing goes live with a blank evidence cell.
| Check | Evidence | Owner |
|---|---|---|
| Green, amber and red topics approved | Signed and dated scope sheet | Compliance lead |
| Source pages approved and dated | Page list with approval dates | Compliance lead, marketing |
| Fixed texts approved: risk warning, status line, AI disclosure | Text file held under version control | Compliance lead |
| Test questions run three times each; every refusal holds | Dated results | Project owner |
| Two advisers have tried to break the bot | Notes and fixes | Senior adviser |
| Hand-off tested end to end, including out of hours | Test records | Operations |
| Distress route tested, with a named responder | Test record | Senior adviser |
| Logging live, retention set, access restricted | Screenshot of settings | IT, compliance |
| Vendor questions answered in writing | Vendor file | Compliance lead |
| Review owner and stop rule named | One-page procedure | Accountable senior manager |
| Privacy notice updated to cover chat | Published page | Compliance, legal |
Start narrow. Put the bot on the fees, services and booking pages for the first month, read every flagged chat, and widen its scope only when the log shows it behaving. The cost of a single recommendation made in the firm's name is higher than the cost of a hundred visitors being politely told to book a call, so when in doubt, keep the list of things the bot may answer short. If your firm is an accountancy practice rather than a regulated adviser, the accountant's version linked earlier holds the build detail and wording, and for insurance intermediaries quote requests without advice adapts the same scope bands.
Questions compliance leads ask about website chatbots
What if the bot gives advice by mistake?
Pause it or narrow it to bookings the same day. Have an adviser contact the visitor, working from the logged transcript, to correct the answer or give proper advice through the firm's normal process. Record it as an incident with the prompt and source versions, ask your compliance lead whether anything must be reported under your regime, fix the cause, and add the phrasing to the test questions before switching the bot back on.
What if the bot can see a client's account?
That pushes it towards personalised answers, which is where advice risk rises. Treat a logged-in assistant as a separate project designed with your compliance lead. Limit it to showing facts the client could already see, keep any recommendation with advisers, and test it harder than the public bot.
Is a scripted decision-tree bot safer than a generative one?
For the questions it covers, usually yes, because every answer is pre-approved. The trade-off is rigidity. Many firms use a generative model to understand the question and then return a fixed, approved answer, which keeps most of the benefit of both and keeps the wording under compliance control.
How long must we keep chat logs?
Treat chats as client communications and ask your compliance lead which retention period applies to those in your regime. Regulators have said recordkeeping rules cover AI-generated communications, but periods vary, so none is quoted here. Make the vendor's retention setting match and state the period in your privacy notice.
Further reads
- AI for Chartered Accountants: What to Tell Clients About Your AI Use — How to word what you tell clients about your use of AI.
- Personal Injury Lawyer Chatbot: Intake Questions, Limits and Cost — Another advice-sensitive trade, with intake questions and limits.
- AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — First jobs for AI in a professional practice, each with a review gate.
- Best AI Tools for Investment Management: Committee Packs and Minutes — AI help for committee packs and minutes in investment firms.
- Consumer Duty and AI: What a Small Advice Firm Must Check — Map each AI use in an advice firm to the FCA's four outcomes, with sample tests and evidence to retain.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: a securities regulator's AI notice, advertising FAQs and guidance on what counts as a recommendation; a markets regulator's 2024 public statement on AI in investment services; a conduct regulator's AI approach page and its definition of a personal recommendation; a 2024 enforcement announcement on misleading AI claims; and Article 50 of the EU AI Act, all read in November 2026.