Blog Industries

AI in financial services: KYC, AML and compliance (2026).

Where AI helps banks, lenders and fintechs with KYC files, AML alerts, credit memos and service, what the FCA, FSB and EU AI Act expect of the system, and where a person must decide.

Fig. 0Every answer checked before it leaves

For banks, lenders and fintechs, the best first AI systems are internal: preparing KYC and AML files, drafting alert narratives, answering staff questions from your own policies, and tracking regulatory change. A trained person reviews every output, the volume is high and the time saved is easy to measure. Customer-facing assistants and anything touching credit decisions come later, under tighter controls.

This guide covers where AI helps in compliance and operations, which kind of AI fits which job, what regulators now expect of the system, and how to start. It reflects the rules as of October 2026.

Where does AI help in KYC and AML work?

Financial services adopted machine learning early for fraud detection and credit scoring. The newer opportunity is the reading, checking and writing around every customer, loan and regulation. The Bank for International Settlements’ Financial Stability Institute noted in December 2024 that institutions were experimenting with generative AI for efficiency and productivity, while customer-facing and high-risk uses stayed relatively limited. Compliance work fits that pattern: it is language-heavy, high-volume and reviewable.

KYC and onboarding files

A file arrives as passports, certificates of incorporation, ownership charts, bank statements and source-of-funds letters, often scanned and in several languages. Document AI extracts the fields, a rules layer cross-checks them against the application and the registry data you already hold, and the system produces a structured summary with the gaps and inconsistencies marked. The analyst starts from that summary rather than from a pile of PDFs, then decides. The result is written back to your core banking system, CRM or onboarding tool.

AML alerts and case narratives

Transaction-monitoring and screening tools already raise the alerts. What takes analysts time is the investigation: pulling the customer’s history, reading the related documents and writing up why an alert was closed or escalated. An AI system can gather that material, draft the narrative with references to each source, and leave the closure or escalation decision to the investigator. Detecting the suspicious pattern itself stays with your monitoring models and rules.

Document and identity fraud

FinCEN, the US Treasury’s financial crimes bureau, said in its November 2024 alert on deepfake media that it had seen an increase in suspicious activity reports describing suspected deepfake media in fraud schemes, which often involve altered or invented identity documents used to get past identity verification. Treat every uploaded document as untrusted input: check it against independent data, use dedicated liveness and document-forensics tools rather than a general language model, and route doubtful cases to a person.

Regulatory change and policy questions

Systems that track regulatory publications, summarise what changed and map it to the policies and procedures affected. A compliance officer decides what to do; the system makes sure nothing is missed. The same retrieval approach powers an internal assistant that answers staff from approved policies, with references, and says so when the policy is silent.

Credit memos and customer service

For credit, AI can draft memos from financial statements, bank data and application documents for an underwriter to review. The decision stays with a person, and the memo makes the reasoning auditable. For service, assistants answer account, product and process questions from approved content, in the customer’s language and channel, and hand over anything sensitive, complex or a complaint. Our AI customer service guide covers what to automate and what to keep human.

Which kind of AI fits which job?

Not every financial-services problem needs a language model. Fraud detection and credit scoring were machine-learning problems long before generative AI, and often still are. The honest design question is which tool does each job.

TaskRight kind of AIData it needsRisk and who decides
KYC file preparationDocument AI plus rules, with an LLM summaryID documents, filings, registry and screening resultsMedium. Analyst approves the file
AML alert narrativesLLM drafting from case data, with source referencesAlert, transaction history, customer fileMedium. Investigator closes or escalates
Transaction fraud and suspicious-pattern detectionMachine-learning models and rules, not an LLMTransaction streams and labelled outcomesHigh. Thresholds are tuned and tested by the owning team
Deepfake and document forgery checksSpecialist liveness and document-forensics modelsImages, video and device signalsHigh. Doubtful cases go to a person
Credit scoringExplainable statistical or ML scorecardsApplicant and bureau dataHigh. Regulated as high-risk in the EU
Credit memo draftingLLM drafting from application documentsFinancial statements, bank data, policyMedium. Underwriter signs
Regulatory change tracking and policy Q&ARetrieval over approved documentsRegulator publications, your policiesLow to medium. Compliance officer decides
Customer serviceLLM agent answering from approved contentProduct terms, FAQs, account lookupsMedium to high. Complaints and vulnerable customers go to a person

What do regulators now expect of AI in a financial firm?

As of the BIS paper, most financial authorities hadn’t written separate AI regulations. It found that the themes in AI guidance (reliability, accountability, transparency, fairness) are the same ones already covered by rules on governance, risk management and consumer protection. Two publications from October 2026 turn that into practical checklists.

The FCA’s good and poor practice review

The UK Financial Conduct Authority published AI in financial services: good practice and areas for improvement on 19 October 2026. It calls its examples illustrative and says they do not create new requirements. It is organised around Consumer Duty and vulnerable consumers, governance and accountability, and operational resilience. For builders, the useful points are specific:

  • Firms own the outcomes. The FCA says firms are responsible for AI systems they develop, procure and deploy, including outputs and outcomes, which covers systems bought from a vendor. Its examples include clear inventories of AI use cases, contractual notice of significant model changes, and exit options rather than single-platform dependence.
  • Human review needs design. It flags over-reliance on human review and box-ticking reviewers, and warns that human review may mask weaknesses in the model. One firm it highlights had reviewers assess each case without being shown the AI’s proposed outcome.
  • Vulnerable customers need another route. The FCA lists excluding vulnerable consumers from AI journeys by default as a weakness, as are inflexible automated journeys that make it harder to get support or move to another channel.
  • Keep records and monitor ahead of harm. Its examples include firms keeping records of prompts, outputs, model versions, applied filters and escalation decisions, and monitoring with thresholds and predefined actions rather than reactive approaches that only identify harm after it arises.

The FSB’s twelve sound practices

On 15 October 2026 the Financial Stability Board finalised its sound practices for the responsible adoption of AI, after a June 2026 consultation. There are twelve, in two groups: organisation-wide governance (board oversight, defined roles, AI in risk management, continuous learning) and risk across the AI lifecycle (materiality assessment, choosing the right technology, data governance, explainability, performance evaluation, human oversight, cyber and ICT risk, and third-party dependencies). The FSB calls them a toolkit to tailor, and urges boards and senior management to use them. They make a sensible checklist for any build.

Which AI rules apply, by market?

  • EU. The AI Act’s Annex III lists AI systems used to evaluate the creditworthiness of natural persons or establish their credit score as high-risk, with an exception for systems used to detect financial fraud. The Council’s final approval of the Digital Omnibus on 29 June 2026 set new application dates for high-risk rules: 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for those embedded in products. Customer-facing assistants also fall under the transparency duty in Article 50: from 2 August 2026, providers must design them so people are told they’re interacting with an AI system, unless that’s obvious to a reasonably well-informed person. See our EU AI Act guide.
  • UK. The FCA’s review above applies existing expectations, such as the Consumer Duty and governance and accountability, to AI. On data, section 80 of the Data (Use and Access) Act 2025 replaced the UK GDPR’s automated decision-making article and came fully into force on 5 February 2026. For significant decisions made solely by automated means, it requires safeguards that include a way to obtain human intervention. Our UK AI rules guide has the detail.
  • US. Colorado’s SB 26-189 covers automated decision-making technology in consequential decisions, which include financial or lending services. It requires notice at the point of interaction, a plain-language description of the system’s role after an adverse outcome, correction of wrong personal data and meaningful human review and reconsideration after an adverse outcome. Developer obligations start on 1 January 2027. California’s CCPA regulations require businesses using automated decision-making technology for significant decisions to comply from 1 January 2027.

These rules change, and applying them to a particular product is a job for your legal and compliance advisers.

Where should a person stay in charge?

A model should prepare, draft and flag. A person should decide in these places:

  • Approving or rejecting a customer or a loan. The AI can assemble the evidence and draft the memo. An accountable person signs.
  • Filing or closing a suspicious activity report. The investigator owns the judgement and the record of it.
  • Complaints and vulnerable customers. The assistant hands over with the conversation attached, and never loops them back into automation.
  • Exceptions the system hasn’t seen. Low confidence, conflicting documents or a failed cross-check should stop the flow and go to a queue, not produce a best guess. Our note on evals before features covers how to measure where that line sits.

Make the review real. If the reviewer sees only “approve” beside a confident summary, they will approve. Show the source documents and the checks that passed or failed, and measure how often reviewers change the draft.

In a regulated firm, the audit trail isn’t overhead. It’s what lets you use AI at all.

What controls should the system have?

ControlWhy it matters
Answers only from approved sources, with citationsCustomers and staff can’t be given invented terms, rates or advice
A full audit trail of inputs, outputs, model versions and approvalsModel risk management, regulators and complaints
Evaluation sets, including adversarial cases, run before every releaseEvidence that changes don’t degrade accuracy or safety
Strict data boundariesCustomer data stays in approved regions and systems, with minimal data in prompts and logs
Defence against manipulationUploaded documents and messages treated as untrusted input (prompt injection)
A replaceable model and a vendor exit planFirms remain responsible for outsourced AI, so no single provider should be a point of failure
A route to a personComplaints, vulnerable customers and any adverse decision

How do you start, and what do you measure?

  1. Pick one internal flow. KYC file preparation or alert narratives are the usual choices: high volume, reviewed by a trained person, with a clear before and after.
  2. Build the evaluation set first. Collect past files and alerts, including the awkward ones, with the outcome a senior analyst would have reached.
  3. Connect the systems. Integration with the core banking, onboarding and case-management tools is usually the biggest piece of work, and the AI is only as good as the data it can reach. It is the kind of work our AI systems services cover.
  4. Run in shadow mode. The system drafts alongside the team while analysts work as normal. Compare, fix, repeat.
  5. Go live with review, then widen. Add customer-facing work only after the controls have been proven internally.

Track a few numbers: analyst time per file or alert, time from application to a complete file, how often reviewers change the AI’s draft, factual errors found on the evaluation set, and the share of cases escalated to a person. After launch, keep measuring, because models and vendors change.

Questions financial firms ask

Can we use public models with customer data?

That is a decision for your security, data-protection and compliance teams. Common setups are enterprise APIs with contractual terms on data residency and no training on your data, or open-weight models run in your own environment for the most sensitive work. Keep the model replaceable either way.

How do we satisfy model risk management?

Treat the AI system like any other model: document its purpose and limits, validate it on a representative evaluation set, monitor it in production, and review it when it changes. The FCA’s review gives monitoring with thresholds and predefined actions as an example of good practice, and a well-built system produces that evidence as it runs.

Will regulators accept AI-drafted credit memos or alert narratives?

A memo drafted by AI and reviewed, edited and signed by an accountable person has a named owner, and the audit trail shows it. Whether a regulator is satisfied depends on your jurisdiction and process, so confirm it with your compliance team. The FCA also describes box-ticking review as a weakness, so the review has to be real.

Can AI decide whether to approve a loan or a customer?

It can be built, but in the EU credit scoring of individuals is classed as high-risk, and Colorado gives customers rights around automated decisions, with California’s rules to follow. Most firms keep a person accountable for the decision and use AI to prepare it. A qualified adviser should confirm what applies to your product.

Does AI make KYC more exposed to deepfakes?

Fraudsters now use generative AI to produce convincing documents and images, which FinCEN has warned about. The defence is layered: specialist liveness and document checks, cross-checks against independent data, and a person for doubtful cases. A language model reading the file shouldn’t be your only control.

Working in financial services?

Modulus Labs AI builds compliance, onboarding and service systems for banks, lenders and fintechs, with audit trails and human review built in. Tell us about the workflow, and we’ll tell you what we’d build first.