Industries17 min read

AI Adoption in Small UK Financial Firms: 2026 Guide

DATS is the AI consulting system from DILR.AI that helps regulated firms decide where AI belongs and how to govern it. The 2026 Bank of England and FCA survey found 73% of short-form respondents under 100 staff use AI, 58% of those mostly use general-purpose large language models, and operations leads current or planned use.

On this page 11 sections
AI Adoption in Small UK Financial Firms: 2026 Guide DATS · FINTECH AI Adoption in Small UK Financial Firms: 2026 Guide 58 % of AI-using small firms mostly run general-purpose LLMs Source: Bank of England and FCA, 2026 dilr.ai/blog

A small regulated financial firm rarely builds an AI model. It buys a seat on a chat assistant, opens a browser tab, and starts drafting client letters, summarising call notes and checking policy wording. The 2026 survey by the Bank of England and the FCA now puts numbers on small-firm adoption. Firms with fewer than 100 employees could answer a shorter questionnaire, and among those short-form respondents, 73% reported using AI. Of the firms using it, 58% said they predominantly use general-purpose large language models.

That changes the governance question. The model risk frameworks written for banks assume a firm that trains, validates and monitors its own models. A twelve-person payments firm or a forty-person wealth manager has none of that, and does not need most of it. What it needs is a control set sized to the thing it actually runs: a third-party chat assistant used by staff, mostly in operations, sometimes near customers. This guide is written for the COO or compliance lead of that firm.

Scope, stated plainly. This post covers the small-firm adoption mix and a proportionate control set for general-purpose LLM use. It cedes the survey's small-firm figures on regulatory and data protection constraints, and the question of how far data protection law really limits AI, to our guide to AI regulatory constraints in UK financial services. It cedes the sector-wide case for where AI pays in payments and lending to the fintech industry guide.

This guide is shipped by the team behind DATS, the five-stage AI consulting system from DILR.AI delivered by senior practitioners who ship code, not decks. Or see the AI operating model engagement, which covers governance, RACI and lifecycle, audit-ready by design.

What does the 2026 survey say about AI use in small UK financial firms?

The Bank of England and FCA 2026 survey found that 73% of small-firm respondents, meaning firms under 100 employees that chose the short-form questionnaire, reported using AI. Of those AI users, 58% predominantly use general-purpose large language models, 32% combine those models with more complex AI workflows, and 9% use machine learning alongside LLM-based tools. For most small-firm AI users, the core tool is a general-purpose language model.

The survey received 196 responses in total, from 101 larger firms and 95 smaller ones, and 105 respondents completed the full questionnaire. In 2026, firms with fewer than 100 employees were given the option of a shorter form that asked for less detail about their AI systems, and section 10 of the report summarises those answers separately. Whether a respondent counted as a small firm was self-certified. The report states that its other figures are based on the full questionnaire unless otherwise stated, so those describe a different set of answers and should not be set beside the short-form percentages as if they measured the same thing.

Two denominators matter when you quote these numbers in a board paper. The 73% is a share of short-form respondents. The 58%, 32% and 9% are shares of the short-form respondents who use AI, not of all small firms, so multiplying them together produces a figure the report never states. Keep each number attached to its own base, exactly as the report writes it, and the evidence holds up in front of a regulator or an investor.

The practical reading is simple. The largest group of small-firm AI users, 58%, predominantly run the same class of tool: a general-purpose model, whether reached through a vendor's chat interface or an assistant built into licensed software. The survey does not report which. That is the starting point for every control decision below, and it is why our DATS approach to AI placement begins by listing what a firm actually runs before it recommends anything.

Why do small financial firms mostly run general-purpose LLMs?

Small financial firms mostly run general-purpose LLMs because those tools need no data science team, no training data and no model build. A firm under 100 staff can buy access, write an acceptable use rule and start using a capable model the same week. The Bank of England and FCA survey confirms the pattern rather than explaining it, so the reasons below are our reading of the economics, not survey findings.

The first reason is staffing. A firm of this size usually has a compliance lead, an operations lead and a small technology function, often outsourced. Building and validating a model needs skills that firm does not carry, and hiring them for one use case rarely makes sense. Buying a general-purpose model moves most of the engineering to the vendor.

The second reason is the shape of the work. Small firms spend a large part of their week on documents and correspondence: client letters, file notes, policy drafting, onboarding packs, complaint responses. That is language work, and it is exactly what general-purpose models do well enough to be useful on day one. A scoring model for credit or fraud is a different proposition, and the report records only a minority of AI-using respondents going beyond LLM tools alone: 32% combine them with more complex workflows and 9% add machine learning.

The third reason is that the tool arrives inside software the firm already pays for. Assistants now sit inside office suites, CRM systems and practice management tools, which means adoption can happen without a procurement decision anyone recognises as an AI decision. Our Microsoft 365 Copilot rollout evidence review covers what the UK government trial found when that happened at scale. For a small firm, the governance consequence is the same: the firm may be using AI before anyone has decided that it should.

None of this makes the tools low risk by default. It makes the risk different in kind. The firm does not own the model, cannot validate it in the way model risk guidance describes, and depends on a vendor for every change. The controls have to work around those facts rather than pretend the firm is a bank, a point the DATS pillar guide to enterprise AI consulting makes for larger organisations too.

Where are small financial firms putting AI to work?

Small financial firms are putting AI to work mainly in operations. In the Bank of England and FCA 2026 short-form results, 64% of respondents named operations as a domain for current or planned AI use. Other domains followed at 47%, then customer services at 36%, risk management at 34%, financial crime at 27% and HR at 21%. Respondents could name several domains, so the shares overlap.

Domains for current or planned AI use, small-firm respondents
64%Operations47%Other36%Customer services34%Risk management27%Financial crime21%HR
Share of short-form respondents (firms under 100 employees, self-certified) naming each domain for current or planned AI use, 2026. Multiple domains could be named. Source: Bank of England and FCA, Artificial intelligence in UK financial services 2026, section 10

Read the chart as a map of where controls need to sit, not as a ranking of value. The survey does not say what operations use consists of. In our reading it is likely to be mostly internal work such as summarising, drafting and checking, where output passes through a member of staff before it reaches anyone outside the firm, which is the natural control point. The AI operating model work we do starts by marking which uses keep a person between the tool and the customer, because that one fact decides most of the rest.

Customer services and financial crime are different. A customer services use sits close to the Consumer Duty, because the output can shape what a retail customer is told. A financial crime use sits close to the firm's systems and controls obligations, because a missed alert is a control failure the firm owns regardless of which tool missed it. The 27% naming financial crime for current or planned use is the figure a compliance lead should watch, because a general-purpose model has not been validated for that work.

The report does not break down what sits inside the 47% of other domains, so treat that bar as unexplained rather than guessing at it. The useful point is that operations leads by a wide margin. Where your own firm's map looks different from this chart, that difference is worth writing down, because it tells you which of the controls below to build first. The AI solutions we have documented follow the same logic, starting from the workflow rather than the model.

What do the FCA's existing rules mean for a firm using a chat assistant?

The FCA's existing rules apply to a chat assistant in the same way they apply to any other tool an authorised firm uses. The FCA has said it does not plan AI-specific regulation and will rely on existing frameworks, naming the Consumer Duty and the Senior Managers and Certification Regime among them. The obligations bind the authorised firm itself, not the AI vendor whose model the firm uses.

The FCA states the position directly on its AI and the FCA: our approach page, last updated on 18 November 2026:

"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 same page describes the FCA's regulatory approach as principles-based and focused on outcomes, and points firms to the AI Update it published on 22 April 2024, which sets out how existing rules apply to AI. For a small firm, the absence of a dedicated AI rulebook is not an absence of obligations. It means the obligations are the ones the firm already has, applied to a new tool.

Three of those frameworks do most of the work. The Consumer Duty applies to firms in their dealings with retail customers, so a firm with retail customers has to make sure that anything an assistant helps produce for those customers still communicates in a way that meets their information needs. A firm whose business is wholly wholesale may sit partly outside it, which is a scoping question for the compliance lead rather than an assumption. The Senior Managers and Certification Regime puts accountability on named individuals, and the exact obligations depend on the firm's category under the regime, so the practical step is to make sure the AI tool falls inside a named person's existing responsibilities rather than nobody's.

The third framework is data protection. The UK GDPR binds the firm as controller whenever staff put personal data into a tool. Article 35 of the UK GDPR requires the controller to carry out a data protection impact assessment before processing that is likely to result in a high risk, and it names new technologies in particular. The ICO's guidance on AI and data protection covers what to consider in that assessment. Whether a given use reaches the high-risk threshold is a judgement the firm has to make and record, not one a vendor can make for it.

If a firm is building or testing something more ambitious than a staff assistant, the FCA also runs testing routes. Our guide to FCA AI live testing covers what evidence that route asks a firm to produce. A firm using a chat assistant for staff drafting is unlikely to need that route.

What does proportionate AI governance look like for a firm under 100 staff?

Proportionate AI governance for a small financial firm is a short control set that matches a third-party chat assistant, not a bank's model risk framework. In our view it needs four parts: a register of the AI tools in use, a written rule on what data and tasks each tool may touch, a named owner under the firm's existing accountability map, and a review step wherever output reaches a customer.

This is DILR's view of a sensible minimum, not an FCA expectation. The FCA's own word for its approach is principles-based, and nothing in its AI material sets a checklist for small firms. The survey does show the issue is live: 23% of short-form respondents identified governance, accountability and human oversight as constraints on adopting AI. A control set that a small team can actually run is meant to address that constraint without adding a programme the firm cannot staff.

A proportionate AI control set for a small regulated firm
01Tool registerWhat runs, where, for …02Use ruleData and tasks allowed03Named ownerInside existing accoun…04Review pointBefore output reaches …
Four controls sized to a third-party general-purpose model used by staff. DILR's view, not a regulatory checklist.

The tool register. One page listing every AI tool staff use, including assistants switched on inside existing software, with the vendor, the plan the firm is on and the business areas using it. Unlicensed personal accounts belong on it too, which usually means asking staff directly. Our piece on shadow AI governance covers how discovery works when tools arrive without a decision.

The use rule. A short written rule, read and acknowledged by staff, saying which data may go into which tool and which tasks are out of scope. The most important line is usually about personal and client data: whether it may be entered at all, and on which plan. The second most important line names the tasks where AI output may never be the final word, such as advice, credit decisions and suitability.

The named owner. The tool sits inside an existing person's responsibilities, usually the COO or the compliance lead, and that person can say what the firm uses, why and under which rule. This is the cheapest control on the list and the one most often missing. It is also the step that connects AI use to the firm's existing SM&CR map, which a written AI operating model is designed to make explicit.

The review point. Wherever output leaves the firm, a member of staff checks it before it goes. That sounds obvious, but it has to be written down and sampled, or it quietly erodes once the tool is trusted. The same principle drives our analysis of AI agents and decision support in UK financial services, where respondents to the full questionnaire mostly reported agents that provide decision support, with a person deciding and acting.

The same diagnostic logic underpins our AI execution office, embedded delivery that turns a control set like this into production placements the client owns.

When does a chat assistant stop being a low-risk tool?

A chat assistant stops being a low-risk tool when its output starts deciding things about people, reaching customers without review, or handling financial crime work it was never tested for. At that point a small financial firm needs more than the four-part control set: a recorded risk assessment, a data protection impact assessment where the threshold is met, and testing before the use goes live.

Four triggers move a use up the scale, and a COO can check each one in a few minutes.

  1. A significant decision about an individual. If the tool's output decides a credit application, an account closure or a claim without meaningful human involvement, UK automated decision-making rules apply. Since 5 February 2026, those rules sit in Articles 22A to 22D of the UK GDPR, and Article 22C sets out the safeguards that must accompany a significant decision taken solely by automated means. Our guide to automated decision-making under Article 22A explains the test.
  2. Output that reaches a retail customer unreviewed. A drafted email that goes out without a check is a Consumer Duty communication the firm owns, written partly by a tool it does not control.
  3. Financial crime work. Screening, alert triage and suspicious activity narratives sit inside the firm's own financial crime controls, and a general-purpose model has not been validated for any of them.
  4. Personal data in the tool. The Article 35 question applies whenever processing is likely to be high risk, which can happen at any volume, and moving to routine processing of client files makes it impossible to leave unanswered.

None of these triggers prohibits the use. Each one changes what evidence the firm should hold before relying on it. A firm that crosses one of them is no longer governing a staff tool; it is placing AI into a regulated process, which is the kind of decision our AI placement work through DATS exists to make carefully.

How should a small firm manage the vendor behind its AI tool?

A small financial firm should manage the vendor behind a general-purpose AI tool as a third party it depends on, with the same basic checks it applies to any important supplier: where data goes, what the contract allows the vendor to do with it, how the firm hears about model changes, and what happens if the service stops. The firm cannot validate the model, so it validates the arrangement.

Start with the plan, not the model. Consumer and business plans from the same vendor can carry different data terms, and those terms decide whether client data may be entered at all. The use rule in the control set should name the plan explicitly, because staff rarely know which one they are on.

Then ask how changes reach you. A general-purpose model can be updated by its vendor without any action by the firm, and the behaviour your staff relied on last month may not be the behaviour they get next month. Our guide to controlling vendor AI model updates covers this for larger firms; for a small firm, a named person who reads the vendor's change notes and re-runs a handful of saved test prompts is often enough.

Security is the third check. The Department for Science, Innovation and Technology published a Code of Practice for the Cyber Security of AI on 31 January 2025, setting baseline security principles for organisations that develop and deploy AI systems. A small firm will not implement it line by line, but it is a useful source of questions to put to a vendor before client data goes anywhere near the tool.

Finally, plan for exit. If the vendor withdraws a plan, changes terms or suffers an outage, the firm needs to know which processes stop. For most small firms the honest answer is that staff fall back to doing the work by hand, which is fine as long as someone has checked that the fallback still works. The mid-market AI readiness assessment guide covers how to test that dependency as part of a wider review.

What is the best way for a small financial firm to govern AI in 2026?

The best way for a small UK financial firm to govern AI in 2026 is to start from what it actually runs, usually a general-purpose LLM used by staff, and build the smallest control set that covers it. That means a tool register, a use rule, a named owner and a review point, with heavier assessment only where a use crosses into decisions, customers or financial crime.

Who should help depends on where the firm sits. A firm whose AI use is limited to staff drafting and summarising may need no outside help at all: a compliance lead with a clear afternoon can write the register and the use rule, and an existing compliance consultancy can review them as part of normal monitoring. That is the right answer for many firms reading this, and it costs far less management time than a formal programme.

The larger consultancies, including Deloitte, PwC, EY, KPMG and Accenture, offer AI governance work, and they are a natural choice where a small firm sits inside a larger group that already uses one of them. Specialist AI firms such as Faculty focus on data science, which matters when a firm is building its own models rather than using a general-purpose one. Where a firm wants senior practitioners to place AI into a regulated process and then ship it, that is what DATS is built for, with a stated delivery discipline of three shippable placements a year and a named owner per ship.

The common mistake is buying the framework before knowing the footprint. A firm that cannot yet list every AI tool its staff use has nothing for a framework to govern, because no framework can govern a list that does not exist. Write the register first; the rest of the decision follows from it. If you want to talk it through, contact the DILR team.

Frequently asked questions

What can the small-firm survey figures not tell a COO?

The small-firm figures in the Bank of England and FCA 2026 survey cannot tell a COO how their own firm compares with peers in the same sector, because the short-form results are reported for small firms as a group. Small-firm status was self-certified, and only firms that chose the short form are included, so the figures describe those respondents rather than every regulated firm under 100 staff.

That is not a reason to ignore them. It is a reason to quote them with their base attached and to use them as a prompt for your own register rather than as a benchmark. The survey's small-firm figures on regulatory and data protection constraints, and the question of how far regulation really limits AI adoption, are covered in our guide to AI regulatory constraints in UK financial services.

Does a small FCA-regulated firm need a separate AI policy?

A small FCA-regulated firm does not need a separate AI policy because of any AI-specific FCA rule, since the FCA has said it will rely on existing frameworks rather than add AI regulation. In practice, a short written use rule for AI tools is still the simplest way to show that existing obligations, such as the Consumer Duty and data protection law, are being met.

Whether that rule sits in a standalone document or inside an existing information security or conduct policy matters far less than whether staff have read it and someone owns it. Our about page sets out who we are if you want a second view on what yours should cover.

Want to work through this on your own tools? Start with DATS AI consulting, read our approach to placing AI inside regulated systems, browse the industries guides, or see the enterprise AI solutions we have documented.

Talk to the operators

Govern the AI you already run.

30-min scoping call · No deck · Confidential. We will tell you which uses need more than a staff rule, and which do not.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.

  • ai adoption small financial firms uk
  • fintech
  • small regulated firm ai governance
  • general purpose llm financial services
  • ai for firms under 100 employees
  • ai consulting reddit
  • best ai governance for small firms 2026
  • dats

What does this article answer?

The questions this article answers, in short

  • The Bank of England and FCA 2026 survey found that 73% of small-firm respondents, meaning firms under 100 employees that chose the short-form questionnaire, reported using AI. Of those AI users, 58% predominantly use general-purpose large language models, 32% combine those models with more complex AI workflows, and 9% use machine learning alongside LLM-based tools. For most small-firm AI users, the core tool is a general-purpose language model.
  • Small financial firms mostly run general-purpose LLMs because those tools need no data science team, no training data and no model build. A firm under 100 staff can buy access, write an acceptable use rule and start using a capable model the same week. The Bank of England and FCA survey confirms the pattern rather than explaining it, so the reasons below are our reading of the economics, not survey findings.
  • Small financial firms are putting AI to work mainly in operations. In the Bank of England and FCA 2026 short-form results, 64% of respondents named operations as a domain for current or planned AI use. Other domains followed at 47%, then customer services at 36%, risk management at 34%, financial crime at 27% and HR at 21%. Respondents could name several domains, so the shares overlap.
  • The FCA's existing rules apply to a chat assistant in the same way they apply to any other tool an authorised firm uses. The FCA has said it does not plan AI-specific regulation and will rely on existing frameworks, naming the Consumer Duty and the Senior Managers and Certification Regime among them. The obligations bind the authorised firm itself, not the AI vendor whose model the firm uses.
  • Proportionate AI governance for a small financial firm is a short control set that matches a third-party chat assistant, not a bank's model risk framework. In our view it needs four parts: a register of the AI tools in use, a written rule on what data and tasks each tool may touch, a named owner under the firm's existing accountability map, and a review step wherever output reaches a customer.
  • A chat assistant stops being a low-risk tool when its output starts deciding things about people, reaching customers without review, or handling financial crime work it was never tested for. At that point a small financial firm needs more than the four-part control set: a recorded risk assessment, a data protection impact assessment where the threshold is met, and testing before the use goes live.
  • A small financial firm should manage the vendor behind a general-purpose AI tool as a third party it depends on, with the same basic checks it applies to any important supplier: where data goes, what the contract allows the vendor to do with it, how the firm hears about model changes, and what happens if the service stops. The firm cannot validate the model, so it validates the arrangement.
  • The best way for a small UK financial firm to govern AI in 2026 is to start from what it actually runs, usually a general-purpose LLM used by staff, and build the smallest control set that covers it. That means a tool register, a use rule, a named owner and a review point, with heavier assessment only where a use crosses into decisions, customers or financial crime.
Written by Dilr.ai EngineeringEngineering team

Can voice AI handle the calls in my sector?

Dilr Voice answers and places calls for regulated sectors

From clinics and estate agents to financial services, Dilr Voice handles calls 24/7 with the consent, recording and audit rules each sector needs.