Enterprise AI in Banking

01The state of AI in banking

Banking should be the natural home of enterprise AI. Retail lenders scored credit statistically long before anyone said data scientist, card networks built fraud models decades ago, and the whole business rests on turning information into decisions about risk, price and eligibility.

That is not what has happened with the current wave. Most banks now run promising pilots: a document-extraction tool in onboarding here, a relationship-manager copilot there, a fraud model in a sandbox, a call-summarisation trial in the contact centre. Far fewer have those systems in production, assisting real decisions on real customers and real money. That gap between demonstration and deployment is where most banking AI initiatives live, and where many quietly end.

The pressure to close that gap is not abstract. Net interest margins move with the rate cycle and rarely in the bank’s favour for long, cost-to-income ratios stay stubborn because core processes still depend on people reading documents and rekeying data between systems, digital challengers keep resetting customer expectations, and fraud and financial-crime losses climb every year. Each of those pressures is an argument for AI, and each is being made in board papers right now.

If you run retail, business, institutional, risk or data for a bank, you do not need convincing on the opportunity. The harder questions are why AI stalls in banking specifically, and what the organisations reaching production do differently.

02The banking data problem

A universal bank is really several banks sharing a brand. Retail, business and institutional banking each grew up on their own core systems, and each wave of acquisition added another core the promised consolidation programme never quite retired. Cards, mortgages, deposits, payments and lending often sit on separate platforms again, bought or built in different decades.

The customer means different things in each of them. In retail a customer is a person; in commercial it is an entity inside a group hierarchy with beneficial owners, guarantors and related parties; in institutional it is a counterparty with limits and exposures. The same human being can be a retail customer, a director of a business borrower and a signatory on a third account, described three different ways in three systems with three identifiers.

Transaction data lives somewhere else again, in payment rails, ledgers and the general ledger, at a volume and velocity nothing else in the bank matches. Risk, treasury and finance each keep their own warehouses and their own version of the numbers. Channel systems, branch, contact centre, app, broker and partner, each generate their own view of the customer and the interaction, in their own formats, at their own quality.

Beneath it all sits the unstructured layer, which in banking carries a surprising amount of the meaning: know-your-customer documents, credit memos and committee papers, loan and facility agreements, collateral valuations, correspondence, complaints and years of call notes hold the substance of what was agreed and why. The structured fields are often just the index.

The consequence: the same customer and the same relationship are described differently in every system that touches them. The same business appears under two registration numbers and three address formats. The same person exists under four identifiers across the group. Point an AI system at that estate directly and it will answer instantly and confidently, and it will be inconsistently wrong in ways nobody can trace.

03What governed AI could deliver

Set the data problem aside for a moment and the potential is easy to state; the same use cases appear in every bank’s strategy paper:

• A unified Customer and entity 360. One governed view of a customer or a business group across every product, account and interaction, so relationship managers and service teams stop stitching the picture together by hand.
• Credit-risk decisioning support. Applications, financials and supporting documents read, extracted and pre-assessed against policy and appetite, so credit officers spend their time on judgement rather than assembly.
• Fraud and financial-crime detection. Patterns across accounts, payments and counterparties surfaced for investigation, and alerts enriched with context so analysts triage faster and escalate the right cases.
• Next-best-action and personalisation. Relevant, compliant offers and interventions shaped by a customer’s whole relationship rather than one product silo’s fragment of it.
• Collections and hardship support. Early identification of financial difficulty and appropriate, fair intervention, grounded in the full record rather than a single overdue balance.
• Regulatory and risk reporting. Prudential returns, financial-crime reporting and risk packs assembled from consistent, traceable data rather than hand-stitched spreadsheets.

Nothing on that list is speculative; every capability has been demonstrated inside banks. The interesting question is why so little of it is in production, and the answer is rarely the models.

04Why initiatives stall

Four forces hold banking AI between pilot and production, and they compound.

Prudential supervision. Banks are prudentially supervised, so the bar for any system touching customer data or influencing decisions is set by regulation, not enthusiasm. A pilot that cannot demonstrate control over information security, operational resilience and accountability does not get promoted out of the sandbox.

Model risk management. Models that inform credit, capital and pricing sit inside a governance discipline banking supervisors across major markets now expect: documented, validated, monitored and owned. A generative tool that cannot show what data it used or why it produced an output cannot pass that gate, and validation teams are right to refuse it.

Fairness, conduct and explainability. When credit is declined, a limit is cut or a customer is treated as higher risk, regulators, ombudsman schemes and courts can ask why, and in many markets the customer is entitled to a reason. “The model said so” is not an answer any general counsel will defend, and discriminatory outcomes are a legal problem, not a tuning problem. Every AI-assisted decision needs a traceable line from source data to output.

Legacy integration cost. Each initiative that connects directly to core banking, cards, payments and channel systems pays the full integration tax alone: bespoke connections, bespoke mappings, bespoke security review. The second project pays it all again. Integration spend crowds out the value the project was meant to deliver.

Underneath all four sits the data problem. Models acting on inconsistent data produce inconsistent decisions, and at banking scale that means tens of thousands of credit, pricing and service decisions a day drifting apart until a regulator, an auditor or a remediation review asks the question.

05The regulatory landscape

None of that caution is optional; the regulatory perimeter is already in place. Five kinds of obligation matter most for banks, and while the named regulators differ by market, every developed jurisdiction enforces an equivalent of each.

Prudential standards. Prudential regulation in the Basel tradition, set internationally by the Basel Committee on Banking Supervision and enforced by each market’s prudential regulator, makes information security, operational resilience and sound risk management board-level obligations, and extends that discipline to third parties and material service providers. An AI system that reads customer data and feeds credit or capital processes sits squarely inside that perimeter.

Conduct and consumer protection. Conduct and consumer regulators expect fair treatment, responsible lending and honest dealing, and those expectations apply to outcomes however they are produced. If AI assists a decision that harms a customer, the fact that software was involved is no defence.

Financial crime. Anti-money-laundering and counter-terrorism-financing regimes, aligned to the international standards of the Financial Action Task Force, require banks to know their customers, monitor transactions and report suspicious activity, and to be able to show how they do it. AI that touches this work inherits the same evidential burden.

Data-protection law. Privacy regimes such as the EU’s General Data Protection Regulation and equivalent national laws govern the collection, use and disclosure of the personal and financial information banking runs on. Sending that data to external AI services raises questions many privacy teams cannot answer comfortably.

The direction of travel. The EU AI Act treats AI used to evaluate the creditworthiness of individuals and to set their credit scores as high-risk, with obligations covering data governance, logging, transparency and human oversight. Wherever your book sits, that is a preview of where banking AI regulation is heading.

Read together, these frameworks converge on a single requirement: a bank must be able to demonstrate control over what its AI systems saw, did and decided, with evidence. That requirement is architectural, and it is exactly what a control plane exists to satisfy.

06The control plane approach in banking

An enterprise AI control plane is the infrastructure layer that sits between an organisation’s systems and its AI tools, resolving the estate’s data into one governed, stored semantic layer and governing every interaction between the two. The full architecture is set out in the reference guide; what follows is how the pattern lands in banking specifically.

One semantic layer across the estate. The control plane connects to core banking, cards, payments, lending and channel systems where they are, resolves the same customer, the same business group and the same relationship once, and stores the result as one governed, unified semantic layer inside the bank’s own tenancy, continuously hydrated from the sources. The systems of record keep doing their jobs and existing platform investments carry forward: weeks-scale deployment rather than a multi-year core-consolidation programme as the precondition for AI.

Least-privilege access on every interaction. Access is governed with least-privilege controls over the underlying data sets, enforced at the moment AI consumes data and inherited from the bank’s existing identity and access management. A retail service copilot sees only retail data appropriate to its role. An institutional coverage assistant cannot retrieve a retail customer’s records.

Complete lineage on every assisted decision. Every credit, service or financial-crime decision assisted by AI carries a traceable record from source systems through transformation to output: which data, which model, which policy, whose authority. When a regulator, an ombudsman scheme or model validation asks why, the evidence already exists.

Deployment inside your own tenancy. The control plane runs inside the bank’s own cloud environment, with the cloud and model providers of your choice. Customer, account and transaction data is not transmitted to or processed on external systems, which keeps sensitive information under prudential-grade control and keeps the privacy assessment tractable.

07Use cases in depth

With the control plane in place, that list stops being aspirational. Here is how each use case works, and where the human authority sits.

Unified Customer and entity 360

Service teams, relationship managers and their copilots see a single governed view of a customer or a business group across every product, account and interaction, filtered by what each role is entitled to see. Related parties, group hierarchies and beneficial ownership resolve to one picture rather than three system fragments. Fewer transfers, fewer repeated questions, and answers grounded in the whole relationship.

Credit-risk decisioning support

Applications, financial statements and supporting documents are read in full, structured and unstructured content together, and pre-assessed against policy, appetite and prior history. The credit officer starts from a complete, evidenced picture instead of a document pile. Approval, terms and pricing authority stay with the officer and the delegated credit authority, and every input to the recommendation is traceable for later challenge.

Fraud and financial-crime detection

Once customer, account, payment and counterparty data share one semantic model, patterns invisible across system boundaries become visible: mule-account networks, structured payment behaviour, identities that recur across relationships. Alerts route to investigators as enriched leads with supporting evidence attached, not as verdicts, and the reasoning behind each is recorded for regulatory reporting.

Next-best-action and personalisation

Offers and interventions are shaped by a customer’s whole governed relationship and constrained by conduct and suitability rules, so what reaches the customer is relevant and compliant rather than a product push. The rules that gate an action are enforced at the same point as data access, so personalisation cannot quietly step outside policy.

Collections, hardship and remediation

Early signs of financial difficulty are identified across the full record and routed to the right support, with fairness and hardship policy applied consistently. When remediation is required, the affected cohort can be identified from governed data, with lineage to prove the population is complete and the treatment consistent.

Regulatory, risk and financial-crime reporting

Prudential returns, risk packs and financial-crime reports are assembled from the same governed semantic layer the rest of the business uses, with every figure traceable to source. The pack that took weeks of spreadsheet assembly becomes a governed output the risk and finance functions can sign off faster, with evidence attached.

08Implementation considerations for banks

Banks that reach production tend to follow the same sequence, and it is deliberately unheroic.

Start with one segment and one workflow. Onboarding document extraction in one business lending line, or a service 360 for one retail segment. A bounded scope makes the risk assessment finite, gives the initiative a named owner, and produces a result the rest of the organisation can inspect.

Connect read-only first. The control plane’s first weeks are observation: scanning and mapping the estate, resolving customers and entities, surfacing the real rather than documented state of the data. Nothing writes back to core systems, which keeps the initial security review proportionate and fast.

Bring risk, compliance and model validation in early. The functions that can veto an initiative at sign-off become its sponsors once they see lineage and audit output working. Their requirements, including human-in-the-loop checkpoints wherever an output affects a customer, should shape the design rather than review it afterwards.

• One segment, one workflow, one named owner. Expand from evidence, not ambition.
• Read-only connection first. Observe and unify before anything acts.
• Risk, compliance and model validation at the table from week one, not at sign-off.
• Human-in-the-loop checkpoints wherever decisions affect customers: credit, pricing, collections, remediation.
• Expand only once the semantic layer exists. The second workflow inherits the first one’s foundations.

The economics follow. The first workflow carries the cost of standing up the semantic layer and governance machinery; every workflow after it inherits them. That is why second and third use cases land in a fraction of the time, and why banks that reach production tend to accelerate rather than stall again.

09Frequently asked questions

10Sources and further reading

• Bank for International Settlements, Basel Committee on Banking Supervision and the Basel Framework, the international basis for prudential standards enforced by each market’s regulator.
• European Union, Regulation (EU) 2024/1689 (the EU Artificial Intelligence Act), including its high-risk classification of AI used to evaluate the creditworthiness and credit scores of individuals.
• Financial Action Task Force, the international standards on combating money laundering and the financing of terrorism.
• European Union, Regulation (EU) 2016/679 (the General Data Protection Regulation), as a representative data-protection regime with equivalents in most markets.
• DataReadyAI, Enterprise AI Control Plane: Definition, Architecture and Buyer’s Guide.