Yes. An FCA-regulated advice firm must apply the Consumer Duty to consumer outcomes affected by AI, whether AI drafts a suitability letter, triages an enquiry or supports a review. The FCA has not introduced a separate AI rulebook. Check product fit, price and value, client understanding and support, then record who owns the decision and how you tested it.
The important question is not whether a model makes the final decision. An internal assistant can still affect which clients receive a prompt response, what risks an adviser notices and how a recommendation is explained. A person’s signature at the end does not reveal a poor outcome earlier in the journey.
Start with a use-case card, not a technology policy
The FCA’s AI approach, updated on 2 October 2026, says it will rely on existing frameworks rather than introduce extra AI regulations. Its later AI practice examples make the responsibility concrete: firms remain responsible for consumer outcomes produced through their products, communications and support. Those examples are illustrative, not a new set of prescriptive rules.
For each proposed use, write a one-page card: the client need; the exact step AI touches; input data; output; named owner; who may approve or override it; the customer group affected; what failure would look like; and evidence that the intended outcome occurred. Avoid “we use AI for efficiency” as an objective. “A review-preparation tool flags missing income updates before the adviser calls” is testable.
A small practice might have three cards: a draft meeting summary, a website question router and a review reminder. The risks differ. A summary can omit a risk warning; a router can misclassify a vulnerable person; a reminder can be late or misleading. Put each on a separate card so a successful low-risk pilot does not become blanket approval for every other use.
| Consumer Duty outcome | Question for the AI use case | Evidence to keep |
|---|---|---|
| Products and services | Does AI change who is offered, screened into or steered towards a service? | Target market, test cases, exceptions and adviser overrides |
| Price and value | Has delivery or access changed while the fee stayed the same? | Before-and-after service map, costs and value assessment |
| Consumer understanding | Can the intended client explain the message and its limits? | Sample communications, comprehension tests and corrections |
| Consumer support | Can a person get timely, effective help or challenge an error? | Escalation tests, response times and complaints review |
Review the three cross-cutting rules as well: act in good faith, avoid foreseeable harm, and enable customers to pursue their financial objectives. A route that is quick for the firm but quietly screens out people who need a phone conversation would fail a good-outcomes test. The FCA’s Consumer Duty explanation sets out the four outcomes and the small-firm context.
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 nowProduct fit: test who the assistant lets through
An advice firm may use AI to summarise a discovery form or sort enquiries. That becomes a product-and-services question when the tool influences access to an adviser or the kind of service offered. List the target market, the customer characteristics relevant to the route and the circumstances that should always reach a person. Include incomplete forms, unusual income, limited digital confidence and a client who says they do not understand the question.
An illustrative failure: a mortgage advice enquiry form asks for monthly income. A client types “varies, about 4,000” and the AI marks the record “low quality, defer”. The adviser never sees the context of self-employment. The firm’s test should ask whether a non-standard answer is surfaced for human review, not treated as a refusal. If the tool cannot explain the routing decision in a useful way, narrow its role to organising the enquiry rather than ranking people.
Write an abstain rule: “If information is missing, inconsistent or suggests a client needs an accessible alternative, route to an adviser. Do not infer ineligibility.” Test ten fictional enquiries designed around edge cases and record the route each received. The exact number is an illustrative starting sample, not an FCA threshold. Expand testing as the risk and volume warrant.
The related advice-firm AI governance checklist covers the wider register and ownership framework. Here the focus is the additional evidence that the client journey still delivers the four outcomes.
Fair value: map the service the client now receives
AI may reduce an adviser’s preparation time without reducing client value. It can also replace a human review that the client reasonably thought was included. Compare the fee, the promised service and the delivered service before and after the change. Ask whether a saving is reinvested in better support, whether response quality improves and whether any group loses access. The mere fact that software costs money is not proof of fair value.
Illustrative worked example: a three-adviser firm completes 40 annual reviews a month. Staff previously spent 25 minutes per case gathering documents and 20 minutes drafting a meeting outline. A controlled AI tool reduces gathering to 15 minutes and drafting to 10, but each adviser spends an extra 8 minutes checking facts and suitability cues. Net saving is 12 minutes per case, or 8 hours across 40 reviews. At an internal planning rate of $70 an hour, that is $560 of released capacity before tool cost and setup.
Suppose the licence and maintenance cost $240 a month and initial setup takes 12 staff hours. In the first month, the $560 planning value does not become $560 cash saved; the firm may have no reduction in payroll and must absorb setup time. The value question is whether reviews remain complete, clients understand the resulting communication and the service still warrants its fee. A prudent use of the eight hours might be to call clients whose circumstances changed rather than send every client an automated note.
Document the before-and-after service map. “Before: adviser reviews all source statements and writes the outline. After: tool extracts items, adviser checks original statements and writes the recommendation.” If the adviser stops checking statements, the product may have changed more than the firm admits. A fair-value review should look at the work actually done, not the workflow diagram approved at procurement.
Understanding: make the message testable by a client
A generated letter can be fluent and still be obscure. Select a few representative passages, ask someone unfamiliar with the case what action the client is being asked to take, and check whether the answer matches the adviser’s intent. Test numerical comparisons separately from plain-English explanation. If a model turns “annual charge” into “monthly charge”, an apparently minor wording change becomes a material misunderstanding.
One illustrative before-and-after: an AI draft says, “Your protection provision has been optimised for resilience.” A clearer adviser-approved version says, “We recommend keeping your existing cover for now. The monthly premium is unchanged. We will review the amount again after you confirm your new income.” The second version names the action, the reason and the information still needed. Have the client or a colleague explain it back in their own words.
Disclosure should be useful, not ceremonial. If AI summarised the meeting, explain that an adviser checked the summary and how the client can correct it. If a website assistant only provides general information, make that limit clear at the point of use and direct advice questions to a qualified person. The firm’s compliance lead should review wording against its own obligations before release.
For the drafting controls around a recommendation, see AI-assisted suitability reports and compliance review. This Consumer Duty check asks a separate question: whether the communication actually supports the customer’s understanding, including someone who reads it without an adviser beside them.
Support: try to break the route to a person
A chatbot that answers opening hours may reduce waiting. A chatbot that keeps a worried client in a loop can obstruct support. Test the path for “I need to speak to someone”, a complaint, a missed deadline, a request for a different format and a client who says they cannot use the online form. Record whether the request reaches the correct person, how long it takes and whether the automated response creates a false expectation.
Consider an illustrative client who types, “I cannot read the document on my phone, please call me.” The assistant replies with a link to the same document three times. That is not merely a poor chatbot answer; it is a failed support journey. A safer rule routes accessibility requests and repeated failed answers to a person, with a message that gives an honest response window. Never invent a “within one hour” promise unless the firm staffs that response.
Vulnerability should not be a crude exclusion flag. The FCA’s AI practice examples warn that automatically excluding people with vulnerability indicators from an AI journey may create another poor outcome if no suitable alternative exists. Offer a route to effective human support, and test that it works. A client may choose an accessible digital path; the firm should not assume every such client wants the same channel.
A second example is a client asking whether to stop a contribution after redundancy. A general-purpose assistant should not improvise personalised advice. It can acknowledge the question, record that an adviser will review it and explain the available contact route. For limits on customer-facing financial questions, see the finance-firm chatbot boundary.
Make sign-off a control with evidence behind it
The firm should know who owns the use case from selection through retirement. The FCA’s AI approach page points to the Senior Managers and Certification Regime (SM&CR) as one of the existing rules that already make senior managers accountable for safe AI use, so the owner is not a formality. Record the senior manager or accountable owner, the person who can pause it, and the adviser who approves client-specific material. Keep versioned prompts or rules, test results, changes, incidents and supplier information. A small spreadsheet and folder can be enough for a narrow use case if they are maintained and accessible.
Human review must have a defined job. “Adviser checks the draft” is too vague. A stronger instruction is: verify names and figures against the source documents; check the recommendation against the file; remove unsupported claims; confirm the client’s requested format; and record the reviewer and date. If advisers are too busy to perform those checks, the control exists only on paper.
The FCA’s recent examples distinguish support, automation and abstention. Build technical and process controls around the risky point: block an unattended recommendation, require source-document comparison, route ambiguous cases to a person and log overrides. For an internal drafting tool, compare the error rate of AI-assisted drafts against a baseline of manual drafts. Do not assume the manual baseline is perfect either.
Give the owner a stop trigger. An illustrative rule might pause a pilot if any live output gives personalised advice without approval, if a complaint reveals inaccessible support, or if the document extraction repeatedly misses a field essential to suitability. These are firm-designed thresholds for a pilot, not regulator-set limits. The owner should be able to suspend the flow quickly, tell staff what to do manually and preserve the evidence for review.
Use a small test pack that resembles actual advice work
The FCA accepts that testing at scale is not always proportionate: its examples describe smaller firms using sample model outputs, adding vulnerability markers to test cases and comparing outputs across scenarios. A practical pack has at least one ordinary case, one incomplete case, one conflicting-document case, one client with a different communication need and one complaint or urgent support request. Use fictional or properly controlled test data. For each, write the expected route and the unacceptable result before running the tool. That prevents the team from redefining success after seeing a polished answer.
For example, a review-preparation assistant receives two fictional statements with different pension values: $81,000 on an older statement and $78,500 on a newer one. Expected behaviour: identify both values and their dates, flag the conflict, and ask the adviser to verify the current value. Unacceptable behaviour: silently use $81,000 in a client letter. The test probes source handling, not prose quality.
A separate support test sends “I want to complain about the advice I received.” Expected behaviour: give the firm’s approved complaints route and escalate. Unacceptable behaviour: reassure the client that the advice was correct or try to close the conversation. A communications test supplies a jargon-heavy paragraph and checks whether the assistant simplifies it without changing the stated fee, risk or condition.
Record the input, output, expected outcome, reviewer and decision. Retest after a supplier changes the product, a prompt is edited or the firm expands the tool’s access. The human review tutorial helps design an approval step, but this firm must judge it against the client outcomes of its regulated activity.
Review live outcomes, not only launch tests
At the end of a pilot period, compare a small sample of AI-touched cases with comparable cases that were handled without the tool. Look for missed details, time to reach an adviser, corrections, complaints, abandoned journeys and client questions showing confusion. Segment where appropriate so an overall average does not conceal worse outcomes for clients who use an alternative channel.
An illustrative monthly review for 40 AI-assisted annual reviews might show 6 source corrections, 2 requests for a clearer explanation and 1 missed escalation. The fact that 39 cases appeared smooth does not cancel the escalation failure. The owner should inspect its cause, fix the route, retest the relevant scenario and decide whether to pause the use case. Keep the decision and its rationale.
Ask an adviser to identify one benefit the client received, not merely minutes the firm saved: a faster review after a changed circumstance, a clearer summary or earlier discovery of missing information. If the only measurable outcome is internal efficiency, the Consumer Duty assessment is unfinished. A qualified compliance adviser should review the specific obligations of your firm and regulated activity, particularly before AI affects recommendations, eligibility or complaints.
Keep a decision log a principal can actually use
A log should explain why the firm accepted a risk, who had authority and what would reopen the decision. One filled-in illustrative entry might say: “Use case: annual-review document extraction. Owner: practice principal. Approved scope: extract dates, balances and unanswered questions into an internal checklist. Prohibited: recommendations, client-facing letters and automatic CRM changes. Test: five fictional files plus five controlled historic files. Finding: one date omitted when a scanned page was rotated. Action: require original-page comparison and flag unreadable pages. Pilot: ten cases, adviser sign-off on each. Review date: after ten cases or any material error.”
This entry is more useful than “AI pilot approved”. It tells the next adviser what the tool may do, tells the principal what to monitor and gives an auditor a path from risk to control. It also makes a later scope change visible. If the firm adds a letter-writing function, that is a new use or a material extension; it needs fresh communication and understanding tests.
Supplier questions belong in the same file. Ask what data the tool receives, where output and source material are retained, who can access them, how a customer record can be corrected or removed, what changes the supplier may make without notice and what happens during an outage. Obtain written answers where the risk warrants it. If a supplier cannot explain a feature that influences a regulated decision, restrict that feature or choose a simpler route.
Assign two calendar checks. The weekly check reviews exceptions and complaints while the pilot is live. The monthly check asks whether the intended client benefit still appears, whether staff are following the approval rule and whether any client group has an unexpectedly poor route. Neither check needs a large dashboard at the start; a short case sample and a signed decision can be more revealing than a colourful chart. The evidence should support a clear decision: continue, amend, pause or retire the use case.
Consumer Duty questions for an AI pilot
Has the FCA introduced separate AI rules for advice firms?
The FCA's AI approach page, updated in October 2026, says it does not plan extra AI regulations and relies on existing frameworks. That does not exempt an AI-supported client journey. Consumer Duty, relevant governance and senior manager accountability still need to be applied to the use case and its risks.
Does a human adviser signing off AI output settle the Consumer Duty issue?
No. Sign-off can catch a mistaken draft, but the firm must also test the whole journey: who gets through, whether communications are understood, whether support works and whether value remains fair. The FCA's AI practice examples call for proportionate controls and monitoring, not a signature in place of outcome evidence.
Must a small firm use the same AI governance as a large bank?
The FCA says its examples are illustrative and firms should respond proportionately to each use case. A small firm can keep a short use-case register, named owner, approval record, test samples and monthly outcome review. The level of control should rise with client impact, especially where AI affects access, suitability or a vulnerable client's route to support.
Further reads
- Advice-Specific AI or ChatGPT: Which Should an Advice Firm Use? — Choose a tool suited to regulated advice records.
- AI Note Takers for Financial Advisers: Compliant Meeting Records — Assess meeting records and consent.
- How Advisers Use AI to Prepare for Annual Client Reviews — Apply controls to review preparation.
- AI Financial Planning Software for Advisers: What to Test First — Question suppliers before procurement.
- AI Receptionist for Financial Advisers: Cost, Scripts and Limits — Inspect the first contact a client receives.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: FCA, AI and the FCA: our approach (updated 2 October 2026); FCA, About the Consumer Duty (updated 23 October 2026); FCA, AI in financial services: good practice and areas for improvement (published 19 October 2026).