Industries
AI for banking and financial services
Policy assistants, document extraction and alert triage for banks, lenders and payment firms. Built to fit model risk governance, with an analyst or underwriter signing off anything that touches a customer or a credit decision.
The sector
What's different here
Banks already run a lot of models and have the governance to match: inventories, validation, effective challenge. A language model or an agent goes through the same process, so the design has to produce the evidence a model risk team will ask for from the first week.
Most of the useful work is internal. Operations staff searching policy, analysts assembling AML case files, underwriters re-keying pay slips into the LOS. These are document-heavy, high-volume tasks where a draft with citations saves time and a person still makes the call.
Customer-facing decisions are a different category. Credit scoring of individuals is high-risk under the EU AI Act, and a service bot that misstates a product term creates conduct risk. We start where the model informs staff and move outward only once accuracy has been measured.
75%
of UK financial firms surveyed already use AI, and a further 10% plan to within three years. Survey of 118 firms by the Bank of England and FCA.[1]
2%
of AI use cases in the same survey are fully autonomous, although 55% involve some degree of automated decision-making.[1]
3 billion
client interactions with Bank of America's Erica assistant to date, averaging more than 58 million a month. Company-reported figures.[2]
Workflows
Where AI does useful work
Each of these runs inside systems the sector already uses. Steps that carry real risk wait for a person.
Policy and procedure assistant
Staff ask questions in plain language and get answers drawn from current policy, product and procedure documents, with the source attached.
- 1A staff member asks a question in the assistant or the CRM sidebar
- 2Retrieval runs over indexed policy documents, filtered by the user's entitlements
- 3The answer cites the section and version it used
- 4The assistant declines when the documents don't cover the question
- 5Every query and the documents returned are logged for compliance review
Compliance approves which document sets are indexed. Answers are read-only and change nothing.
Document management (SharePoint, OpenText) · CRM (Salesforce Financial Services Cloud) · Identity (Entra ID) · Audit log store
KYC and AML alert triage
An agent assembles the case file for each monitoring alert and drafts a narrative, so analysts spend their time on judgement.
- 1Transaction monitoring raises an alert
- 2The agent pulls customer KYC records, transaction history and adverse-media hits
- 3It drafts an investigation summary and a proposed disposition
- 4An analyst reviews the evidence and approves, changes or escalates
- 5Any SAR is written and filed by a person
An AML analyst decides every disposition. The agent never closes an alert or files a report.
AML case management (NICE Actimize, Oracle FCCM) · Core banking (Temenos, FIS) · KYC records · Adverse-media screening
Loan document extraction
Pay slips, bank statements and financial statements are read, spread and checked before they reach the underwriter.
- 1Ingest the applicant's documents from the portal or email
- 2Extract fields and spread the financials
- 3Flag inconsistencies, such as stated income that doesn't match deposits
- 4Pre-fill the LOS with each value linked to its source page
- 5The underwriter reviews and corrects before submission
The underwriter confirms extracted values. The credit decision stays with the bank's approved model or a person.
LOS (nCino, Finastra) · Document capture (Azure AI Document Intelligence) · Core banking · Data warehouse
Regulatory change mapping
New rules are compared with internal policies and controls, and compliance gets a draft gap analysis to work from.
- 1Ingest new or amended rules and guidance
- 2Map each obligation to existing policies and controls
- 3Draft a gap analysis with references to both texts
- 4Compliance reviews, edits and assigns owners
Compliance owns the gap analysis, and nothing reaches the policy library without their sign-off.
Policy library · GRC platform (ServiceNow GRC) · Document management · Regulator publications
Service assistant with hand-off
Authenticated customers get answers and read-only account information, and harder cases move to a person with the context attached.
- 1The customer authenticates in the app or with the contact centre
- 2Intent is detected and routed
- 3Read-only lookups answer balance and status questions
- 4Transactions run through fixed, scripted flows
- 5Everything else goes to a person with a summary of the conversation
A service agent handles disputes, complaints and anything that resembles advice.
Contact-centre platform (Genesys) · Core banking · CRM (Microsoft Dynamics) · Knowledge base
Systems and rules
What the design has to respect
The systems we expect to connect to, and the regulation that decides what an AI system may do with the data in them.
- Core banking (Temenos, FIS, Fiserv, Finastra, Oracle FLEXCUBE)
- LOS (nCino, Finastra)
- CRM (Salesforce Financial Services Cloud, Microsoft Dynamics)
- AML and case management (NICE Actimize, SAS, Oracle FCCM)
- Document management
- Data warehouse or lakehouse
- Contact-centre platform
- Identity and entitlements
- GRC and policy library
- EU AI Act, Annex III
- AI used to assess the creditworthiness or set the credit score of individuals is high-risk, as is life and health insurance pricing. Under the Digital Omnibus these obligations are scheduled to apply from 2 Dec 2027.
- SR 26-2 (formerly SR 11-7)
- Revised US model risk guidance from the Fed, OCC and FDIC, issued 17 Apr 2026. It covers inventory, validation, effective challenge and third-party models, and the Fed says it is most relevant to banks with over $30B in assets. The letter places generative and agentic AI outside its scope, so banks govern those under internal policy.
- GDPR Art. 22
- Limits decisions based solely on automated processing that significantly affect a person. We keep a person in any flow that ends in a decision about a customer.
- PCI DSS v4.x and DORA
- PCI DSS applies to any flow that touches card data. DORA governs ICT third-party risk for EU financial entities, which includes model and hosting providers.
- Record-keeping and data residency
- Adviser communications fall under record-keeping and surveillance rules, and local outsourcing and residency rules often decide where the model and the data may run.
Risks
What goes wrong, and how we design for it
- Wrong policy or product terms given to a customer
- Answers come only from retrieved, versioned documents, with the citation shown. When retrieval confidence is low the assistant declines and routes to a person.
- Credit outcomes nobody can explain
- We don't put a language model in the credit decision. It extracts and summarises; the approved scoring model or an underwriter decides.
- Data crossing information barriers
- Entitlements are synced from source systems and enforced at retrieval time, and we test with low-privilege accounts that should see nothing.
- Drift in fraud and AML support models
- A fixed evaluation set is re-run on every model or prompt change, and agreement with analyst dispositions is tracked over time.
- Cutting service staff on optimistic numbers
- Commonwealth Bank reversed planned job cuts tied to a voice bot in 2025. We report repeat contact and complaints next to containment before anyone plans headcount around it.
Where to start
A read-only policy assistant for operations staff
It makes no customer-facing decision and sits outside the Annex III credit category, yet it builds the data access, permissions and evaluation that KYC and loan-document work need later. The value is easy to measure in time to answer and escalations avoided.
- Index one policy and procedure set, such as retail operations, from SharePoint or the DMS with its access controls
- Answer with citations to section and version, and decline when the documents don't cover the question
- Log every query, answer and document returned for compliance and model risk review
- Build an evaluation set of 100 to 300 real staff questions with answers agreed by the policy owners
- Measure time to answer, escalations to the policy team and correctness against the evaluation set
- Next: extraction for loan or KYC packs, with an underwriter approving before anything is written to the LOS
In the sector
Publicly reported examples
What others in the industry have announced. These are not our clients.
Morgan Stanley adviser assistant[3]
Over 98% of financial-adviser teams actively use the AI @ Morgan Stanley Assistant to retrieve internal information. Vendor-reported by OpenAI.
JPMorgan LLM Suite[4]
The bank's in-house LLM Suite reached about 200,000 employees within roughly eight months of its 2024 launch, with later 2025 reports putting access at about 250,000. Reported by trade press.
Commonwealth Bank reverses voice-bot job cuts[5]
CBA reversed a decision to cut 45 customer-service roles tied to an AI voice bot after the union disputed its call-volume claims. The bank admitted its assessment had not adequately considered all relevant business considerations.
Use cases
Related use cases
- Company knowledge searchStaff ask a question in plain language and get an answer drawn from your own documents, with citations, limited to what each person is already allowed to see.
- Document processingInvoices, forms and contracts are read, checked against your own records and posted to the right system, with low-confidence fields and failed checks sent to a person.
- Back-office workflowsReconciliations, approvals and onboarding are prepared by agents with the evidence attached, then routed to the right approver. Nothing is posted or paid without a person signing off.
- Customer support automationAn agent answers routine questions from your own policies and account data, takes a small set of allowed actions, and hands everything else to a person with the case already summarised.
Sources
- 1.Bank of England and FCA, Artificial intelligence in UK financial services - 2024 (Nov 2024)
- 2.Bank of America, A decade of AI innovation: BofA's virtual assistant Erica surpasses 3 billion client interactions (Aug 2025)
- 3.OpenAI, Morgan Stanley customer story
- 4.AI News, JPMorgan Chase AI strategy 2025
- 5.Bloomberg, Commonwealth Bank reverses job cuts decision over AI chatbots (Aug 2025)
Figures are as reported by each source on its publication date. Examples are public industry cases, not AI Sewa clients.
Next step