// FOUNDRY5
Ai & Tech

AI Inside an FCA-Regulated Firm: What You Can Ship and What Compliance Will Stop

This guide to FCA regulated AI development shows which AI features UK financial firms can ship, which compliance will stop, and which named rule decides each, from COBS 4 and the Consumer Duty to SM&CR and SYSC 9, with a ship-versus-stop table and a pre-build checklist.

Table of Contents

  • No AI rulebook exists, so your existing rules decide
  • What can you ship in FCA regulated AI development, and what will compliance stop?
  • Financial promotions and Consumer Duty: where AI output gets stopped
  • Who is accountable for AI under SM&CR?
  • Audit trails and record keeping: what you must be able to show
  • Explaining an AI decision to a customer and to the FCA
  • Why FCA regulated AI development breaks when compliance arrives after launch
  • A pre-build checklist for product and compliance to sign off together
  • Frequently Asked Questions
  • Next step: settle the audit trail before the first sprint

 

For FCA regulated AI development, Foundry 5 designs to existing rules, because the FCA says it does not plan extra AI regulations and will rely on frameworks such as the Consumer Duty and the Senior Managers and Certification Regime (SM&CR) (FCA, 2025). Existing rules decide what ships. Build the audit trail in before launch, not after.

 

Your product lead has a working AI prototype. Your compliance lead has doubts. On FCA regulated fintech buildsFoundry 5 closes that gap before the first sprint, because the fear of a blocked launch is rational but aimed at the wrong target.

 

FCA regulated AI development is rarely stopped by a missing AI rulebook, because no such rulebook exists. It is stopped by existing rules applied to a feature nobody has mapped. This guide sets out which AI features ship, which get stopped, and which named rule decides each one.

 

 

No AI rulebook exists, so your existing rules decide

The FCA does not plan extra regulations for AI. It relies on existing frameworks, including the Consumer Duty and SM&CR, so any AI feature is judged by the rules that apply to the firm. In the 2024 Bank of England and FCA survey, 75% of firms already used AI, so the live question is how to ship it compliantly.

 

Waiting for an AI-specific rulebook is a common mistake. Teams pause the build, expecting guidance that never arrives, while the launch date slips. As of November 2026, the FCA’s published position on AIissued in September 2025, is direct: “We do not plan to introduce extra regulations for AI. Instead, we’ll rely on existing frameworks, which mitigate many of the risks associated with AI.” The page names the Consumer Duty and SM&CR as relevant frameworks.

 

That shifts the question from whether the FCA permits AI to whether this feature, as designed, meets the rules the firm already answers to: the Principles for Businesses, the Consumer Duty, SM&CR, the financial promotions rules in the Conduct of Business Sourcebook (COBS 4) and the record-keeping rules in the Senior Management Arrangements, Systems and Controls sourcebook (SYSC 9). Each of those has an owner inside your firm. None of them is new.

 

The scale of adoption makes this urgent. The Bank of England and FCA surveyed 118 firms in 2024, and a further 10% planned to start using AI within three years. Regulated AI is the norm. The firms that ship cleanly treat each AI feature as something the existing rulebook already covers, from day one.

 

 

What can you ship in FCA regulated AI development, and what will compliance stop?

In FCA regulated AI development, Foundry 5 treats compliance as the first problem and engineering as the second. Features that ship share three traits: a named senior owner, human review where customers are affected, and a complete record. Features that get stopped are customer-facing outputs that nobody can explain, evidence or attribute to an accountable person.

 

Every AI feature in a regulated firm lands in one of two buckets. The table below sorts the common ones by the named rule that decides the outcome, so product and compliance can read the same page.

 

The ship-versus-stop table

AI feature What ships What gets stopped Rule that decides
AI-drafted financial promotions Drafts approved by a named person, with the approval recorded Drafts sent to customers without recorded human sign-off COBS 4: fair, clear and not misleading
Customer-facing outputs such as chat, summaries and nudges Outputs tested against customer outcomes, with a route to correct errors Outputs nobody can evidence as good for the customer Consumer Duty
Automated decisions Decisions with a named senior manager and human oversight where customers are affected Decisions with no accountable owner SM&CR
Logging Input, model version, output, reviewer and action recorded for each decision Logs that cannot reconstruct a decision after the event SYSC 9 record-keeping
Explanations of decisions Plain-language reasons a customer and the FCA can follow Outcomes the firm cannot explain Explainability under the Consumer Duty and UK GDPR automated decision-making safeguards

Read the table by its last column. The rule, not the technology, decides the result: a chatbot and a spreadsheet macro face the same test if both reach a customer. A feature passes when it has an owner, a reviewer and a record. It fails when any one of the three is missing.

 

The table also leaves out any ban on a model or vendor. Nothing in it stops a firm from using a large language model. Compliance stops the unexplained, the unowned and the unrecorded, which is a narrower and more fixable list than most teams fear.

 

 

Financial promotions and Consumer Duty: where AI output gets stopped

AI-written content is still a financial promotion. COBS 4 requires communications to clients to be fair, clear and not misleading, whoever or whatever drafted them. Under the Consumer Duty, the firm answers for the customer outcome, so an AI tool that drafts promotions needs human sign-off and a record of that sign-off.

 

Consider a hypothetical investment firm. An AI tool drafts a weekly market email for retail customers. A reviewer skims it, replies that it looks fine, and the email goes out. Nothing records who checked it, against which standard, or which version was sent. A customer later complains that the email played down a risk, and the firm has a complaint and no evidence to answer it with. This is an illustration, not a case.

 

The email itself may well have been fine. That is not the test. COBS 4 asks whether the communication was fair, clear and not misleading, and the firm has to be able to show that the check happened. AI can write financial promotions under FCA rules, provided a person owns the review and the review leaves a record.

 

The Consumer Duty raises the bar further. It tests customer outcomes rather than intentions, so a well-meant AI feature that leaves customers misunderstanding a product still fails. Strong teams test AI-drafted content against what a customer understood after reading it, rather than against what the author meant to say.

 

The practical consequence is a default. Treat every AI-assisted customer message as a promotion until compliance says otherwise. Build the approval step into the tool itself, with the reviewer’s name, the time and the version held on the record, so the evidence exists before anyone asks for it.

 

 

Who is accountable for AI under SM&CR?

A named senior manager stays accountable for AI under the Senior Managers and Certification Regime (SM&CR). The model, the vendor and the build studio do not carry that accountability. In the 2024 Bank of England and FCA survey84% of firms reported an accountable person for their AI framework.

 

The common mistake is assuming the supplier carries the risk. A firm buys an AI tool, the vendor’s contract describes the output as advisory, and everyone relaxes. Accountability does not move. Under SM&CR a named senior manager answers for the activities in their area, and AI is a method of doing those activities, not an exemption from them.

 

The survey figure matters for a second reason. If 84% of firms report an accountable person, naming an owner is the norm, and a firm without one is the outlier a supervisor notices first. The owner is the first control the rest of the framework hangs from.

 

Apply a practical test before build starts. Can compliance name the senior manager for each AI feature? That means a person whose responsibilities cover the feature, not a team or a vendor. If finding the answer takes a meeting, the feature is not ready to build.

 

Consider a firm planning three AI features: a promotions drafter, a triage assistant and a reporting summariser. Two map cleanly to one senior manager. The third falls between two. That gap is where projects stall, so settle it in week one rather than at the release review.

 

 

Audit trails and record keeping: what you must be able to show

An AI audit trail should record the input, the model and version, the output, the human reviewer and the final action for every decision. SYSC 9 sets the FCA’s record-keeping rules. A log that cannot reconstruct a decision after the event will not satisfy a supervisor.

 

Most teams meet the audit trail problem when a complaint or a supervisor asks a simple question: why did the system do that? If the only answer is a screenshot and a recollection, the record has failed. The question is always asked after the event, which is exactly when a weak log cannot be repaired.

 

A usable audit trail for AI decisions in financial services holds five things: the input the system saw, the model and its version, the output it produced, the human who reviewed it, and the final action taken. Miss one and the chain breaks. Without the version, nobody can say which behaviour produced the output. Without the reviewer, human oversight is a claim rather than evidence.

 

Model governance depends on the same record. Version control, change approvals and a note of what changed between releases all live in the log. Confirm the exact SYSC 9 wording with compliance before quoting it in a policy document, then design the record so it can reconstruct any decision, whatever the final wording says.

 

Retrofitting is the expensive route. Logging touches every write to the data model and every screen where a reviewer acts, so adding it later means reopening the build. Consider a firm that launches, then has to add reviewer identity to ten workflows. Each one needs a schema change, a screen change and a retest. Designing it in costs a few extra decisions in the first week.

 

 

Explaining an AI decision to a customer and to the FCA

Firms should be able to explain an AI-assisted decision in plain language to the customer and to the FCA. In the 2024 Bank of England and FCA survey, only 2% of AI use cases were fully autonomous, so most AI-assisted decisions still involve a person. Where a decision is solely automated, UK GDPR safeguards also apply.

 

Explanation is where the human in the loop pays for itself. Where a reviewer approves an output, that person can say why. Where nobody reviews, the firm has to explain a decision that no person made. The test is simple: can the firm tell a customer, in plain language, why they received this outcome? If compliance cannot say it, a customer will not follow it either.

 

The same survey found that 55% of AI use cases have some degree of automated decision-making, and 24% of those are semi-autonomous. Most firms therefore sit in the middle: automation helps and a person decides. That middle ground is the easiest to explain and to defend.

 

Solely automated decisions are the edge case. UK GDPR automated decision-making rules add safeguards when no person is involved in the decision. The ICO published updated draft guidance on automated decision-making, including profiling, for consultation on 8 April 2026. As of November 2026, treat it as consultation guidance rather than settled, and watch for the final version.

 

Firms unsure where a feature sits on that scale can start with AI consulting before they commit to a build, so the explanation route is designed rather than discovered at the first complaint.

 

 

Why FCA regulated AI development breaks when compliance arrives after launch

In FCA regulated AI development, Foundry 5 sees compliance and the audit trail arrive after launch too often, and the product gets rebuilt. Logging, approvals and review steps touch the data model and every user flow, so they cannot be bolted on. Gather, an FCA-regulated investment platform, was built the other way round.

 

What forces the rebuild

The pattern is consistent. A team builds the feature first and hands it to compliance in the final weeks. Compliance asks for reviewer sign-off, version history and a decision log. Each request touches the data model. Each approval step changes a user flow. The release moves, and part of the product is rebuilt.

 

The failure is one of sequencing: compliance asked for what the rules always required, and the build was never shaped to provide it. A better fix is a change of order rather than a change of people: owner, review route and record first, features second.

 

What Gather shows

The team at Foundry 5 built the Gather investment platforman FCA-regulated product, with regulation shaping the design from the start rather than reviewing it at the end. Retail investing is the setting the Consumer Duty’s outcome tests exist for, so the regulatory questions had to be answered in the product itself.

 

The same sequencing applies to AI features. The accountable owner, the review route and the record are design inputs, not release tasks. A firm building AI in an FCA-regulated business gets the same benefit: one build, reviewed once, rather than a build and a rebuild.

 

Test your AI roadmap against a compliance review. A 30-minute scoping call can check it against the rules above, with no pitch deck and no commitment. Book a free discovery call and bring your compliance lead if you can.

 

 

A pre-build checklist for product and compliance to sign off together

Before build starts, product and compliance should agree five things: the accountable senior manager, the financial promotions review route, the record each AI decision leaves, the plain-language explanation a customer receives, and due diligence on any third-party AI supplier. Agreeing them up front avoids a rebuild later.

 

Use the list as a shared sign-off. Both sides initial it before sprint one, so nobody learns the answers from the release review.

  • Accountable senior manager: a named person under SM&CR for each AI feature, recorded before build.
  • Promotions review route: who reviews AI-drafted content against the FCA Handbook rules in COBS 4and where the approval is stored.
  • Decision record: the input, model version, output, reviewer and action held for each AI-assisted decision.
  • Customer explanation: the plain-language reason a customer receives, written before the feature is built.
  • Supplier due diligence: what the firm knows about any third-party AI supplier, including where data goes, how model changes are notified and what the contract says about records.

 

Find out early which of the five is hardest to answer, because that item is where the rebuild risk sits. Raise the supplier question too. If the model changes behaviour next month, how will the firm know, and who approves the change? Model governance lives in that answer.

 

For regulated clients, Foundry 5 scopes its AI development services around this list, with product and compliance signing it off together. Teams who want the briefing side of the same conversation can read the guide to briefing an AI development project before any studio call.

 

 

Frequently Asked Questions

Does the FCA have specific rules for AI in financial services?

No, the FCA has no AI-specific rulebook; it relies on existing frameworks such as the Consumer Duty and SM&CR, and Foundry 5 designs every build to those rules. Existing rules still apply in full, including the Principles for Businesses, COBS 4 and SYSC 9. The firm answers for outcomes, whichever technology produced them.

 

Who is accountable if our AI gets a decision wrong?

The firm and its named senior manager under SM&CR remain accountable for the decision, not the model, the vendor or the studio that built the feature. For every feature, Foundry 5 builds so that an accountable person can be named. In the 2024 Bank of England and FCA survey, 84% of firms reported an accountable person for their AI framework.

 

Do AI-generated financial promotions need FCA approval?

AI-generated promotions must meet COBS 4, meaning fair, clear and not misleading, whoever drafts them, and whether FCA approval is needed turns on the firm’s status. Compliance should sign off every AI-assisted promotion, and the review should leave a record. Foundry 5 builds that sign-off step into the workflow rather than leaving it to email.

 

Can an FCA-regulated firm use ChatGPT?

Yes, an FCA-regulated firm can use OpenAI’s ChatGPT, but existing rules apply to that use, so it needs governance, records and supplier due diligence first. The test is how the firm uses the tool, not which tool it picks. Inside that governance, Foundry 5 integrates large language models with a named owner and logged outputs.

 

What records must a regulated firm keep for AI-assisted decisions?

Records follow SYSC 9 record-keeping, and in practice that means keeping the input, model version, output, reviewer and final action for every AI-assisted decision. Check the exact rule wording with compliance before quoting it. The log Foundry 5 designs lets any decision be reconstructed after the event.

 

 

Next step: settle the audit trail before the first sprint

In FCA regulated AI development, Foundry 5 puts the audit trail and accountability record in the first release, because the rules do not change after launch and a rebuild costs more than a delay. Product and compliance agreeing one checklist before sprint one is what turns a blocked launch into a shipped one.

 

If you are building AI for a regulated financial firm and want a partner who stays after launch, book a free 30-minute discovery call with Foundry 5. There is no pitch deck and no obligation, only a direct conversation about whether your project is a fit.

← Back to Blog
Share This LinkedIn → Twitter →
More from the blog

Keep reading.

View all articles →
London Based · Founder Focused

Enough reading. Let us build something together.

Thirty minutes. No deck required. Just your idea and what it needs to do.