No accountable workflow owner
The innovation team runs the pilot, but no business owner accepts the outcome, process change, residual risk, or operating budget.
Banking AI · 12-stage checklist · free CSV
AI implementation in banking: from pilot to controlled production.
A practical operating guide for banks and fintech companies implementing AI workflows and agents under model, vendor, security, compliance, and customer controls.
To implement AI in banking, start with one owned workflow and a measured baseline. Define the intended use before selecting technology. Register the models, vendors, data, decisions, and owners. Test representative and high-impact cases. Integrate human review, exceptions, audit evidence, incident response, and a stop path. Run a bounded production pilot. Scale only after business results and control evidence support the decision.
Scope boundary: this guide is an operating framework, not legal, regulatory, audit, cybersecurity, model-validation, or financial advice. A bank must apply current requirements with its qualified owners and supervisors.
Pilot-to-production gap
Why banking AI pilots get stuck.
A demonstration proves that a model can produce an output. It does not prove that the bank can operate the complete workflow.
No baseline or acceptance gate
The team measures model quality but cannot show the change in cost, time, accuracy, customer effect, control performance, or incidents.
Risk joins after design
Model risk, compliance, legal, privacy, security, audit, and third-party risk find missing requirements after the system and contract exist.
The workflow stops at the API
The pilot omits identity, permissions, source systems, queues, human review, exceptions, monitoring, support, and recovery.
12-stage implementation checklist
Make each stage produce a decision and evidence.
| Stage | Required decision | Minimum evidence |
|---|---|---|
| 1. Mandate | Name the workflow and executive objective. | Current process, baseline, users, volume, cost, quality, risk, target, and stop conditions. |
| 2. Intended use | Define what the AI system will and will not do. | Affected decisions, customer impact, materiality, jurisdictions, users, and accountable owners. |
| 3. Inventory | Register every model, vendor, component, and dependency. | Versions, data sources, third parties, business owner, control owner, status, and change path. |
| 4. Data and access | Set the minimum data and tool access. | Rights, lineage, classification, security, privacy, retention, secrets, permissions, and logs. |
| 5. Build or buy | Select an architecture and delivery model. | Requirements, due diligence, contract rights, concentration, change notice, continuity, and exit. |
| 6. Controls | Assign decision rights and human oversight. | Reviewer, approval, override, escalation, separation of duties, audit trail, and prohibited actions. |
| 7. Evaluation | Set thresholds before release. | Representative cases, high-impact failures, quality, bias where relevant, security, latency, and cost. |
| 8. Integration | Connect the system to the real workflow. | Interfaces, permissions, exceptions, fallback, incident process, customer contact, and support. |
| 9. Pilot | Approve a bounded production release. | Named users, limited scope, monitoring, issue log, adoption data, and scale or stop gate. |
| 10. Monitor | Operate within defined thresholds. | Performance, drift, incidents, complaints, overrides, vendor changes, and outcome analysis. |
| 11. Scale | Expand only from verified evidence. | Benefit record, residual risk, control performance, capacity, training, funding, and approval. |
| 12. Exit | Retire or replace without losing control. | Continuity, data return or deletion, customer effect, evidence archive, lessons, and next owner. |
Agentic AI in banking
Treat every agent action as a controlled capability.
An agent can retrieve data, call a tool, change a record, send a message, or trigger another process. Control the action, not only the text response.
Use least privilege
Separate read, draft, recommend, approve, and execute rights. Bind access to an identity, role, purpose, environment, and time.
Set transaction limits
Require human approval or dual control for material, irreversible, customer-facing, financial, or regulated actions.
Test adversarial paths
Test prompt injection, poisoned retrieval, unauthorized tool calls, data leakage, secret exposure, excessive agency, and unsafe fallback.
Reconstruct each decision
Retain the version, input source, retrieved evidence, tool call, output, reviewer, override, action, time, and result where policy permits.
Control model and vendor updates
Define change notice, regression tests, approval thresholds, rollback, emergency disablement, and vendor incident communication.
Measure more than containment
Track task quality, customer effect, cost-to-serve, latency, adoption, overrides, complaints, incidents, and financial outcome.
Vendor diligence
Evaluate the complete third-party relationship.
The model is one dependency. Hosting, retrieval, monitoring, support, data processors, tools, integrators, and contract terms also affect the bank.
- Workflow fit: intended users, actions, data, languages, channels, volume, latency, exception rate, and customer effect.
- Technical fit: architecture, interfaces, identity, logging, model options, data location, portability, observability, and failure behavior.
- Risk evidence: security, privacy, validation access, model information, subcontractors, business continuity, incidents, and independent reports.
- Operating rights: use of bank data, training rights, change notice, audit, service levels, support, remediation, termination, deletion, and transition.
- Concentration and exit: critical dependencies, replacement time, retained knowledge, export formats, fallback, and orderly termination.
Paul Okhrem's AI due-diligence framework and AI governance checklist provide reusable decision records for this work.
Success measures
Measure business value and control performance together.
| Measure group | Examples | Decision it supports |
|---|---|---|
| Business | Cycle time, cost-to-serve, throughput, revenue, loss, customer retention | Did the workflow improve the intended outcome? |
| Task quality | Accuracy, completeness, consistency, exception rate, false positive and false negative cost | Is the system fit for its intended use? |
| Customer | Resolution, complaints, wait time, access, fairness measures where relevant | Did service improve without unacceptable harm? |
| Control | Overrides, unauthorized actions, incidents, response time, audit gaps, vendor changes | Are controls operating at the required level? |
| Adoption | Eligible users, active use, abandonment, workarounds, training, support demand | Has the operating workflow actually changed? |
| Economics | Model, infrastructure, integration, review, support, control, rework, and exit cost | Is the full operating case still positive? |
Current primary sources
Use banking guidance that matches the system and date.
- Federal Reserve SR 26-2, Revised Guidance on Model Risk ManagementIssued April 17, 2026. It supersedes SR 11-7 and emphasizes a risk-based approach tailored to a banking organization's model risk profile, size, complexity, and use.
- Federal Reserve supervisory guidance on model risk managementCovers development and use, validation and monitoring, governance and controls, and vendor or third-party products. The guidance states that generative and agentic AI are outside its model scope, while bank governance should still determine appropriate controls.
- OCC Bulletin 2023-17, Third-Party RelationshipsInteragency guidance for planning, due diligence and selection, contract negotiation, ongoing monitoring, and termination across the third-party lifecycle.
- NIST AI Risk Management FrameworkVoluntary cross-sector framework for trustworthiness considerations across AI design, development, use, and evaluation.
Freshness note: older banking articles often cite SR 11-7 as current. The Federal Reserve states that SR 26-2 superseded SR 11-7 on April 17, 2026.
FAQ
Banking AI implementation questions.
How should a bank implement AI?
A bank should start with one owned workflow and a measured baseline. It should define the intended use, classify risk, register the system and vendors, control data and access, test representative and high-impact cases, integrate human review and exceptions, run a bounded production pilot, monitor performance and incidents, and scale only after evidence supports the decision.
How do you secure AI agents in banking?
Give each agent the minimum data and tool access needed for its task. Separate read and write permissions. Require approval for high-impact actions. Validate inputs and outputs, restrict tools, protect secrets, log decisions and actions, test prompt and data attacks, monitor abnormal behavior, and provide a tested stop and recovery path.
How should a bank evaluate an agentic AI vendor?
Evaluate workflow fit, intended use, architecture, data rights, security, model and subcontractor dependencies, evaluation access, audit logs, human controls, change notice, monitoring, incident support, financial condition, concentration risk, continuity, contract rights, and exit. Test the vendor in the bank's own representative cases before production.
Why do banking AI pilots fail to scale?
Common causes are an unowned workflow, no baseline, unclear intended use, inaccessible data, late risk involvement, weak integration, vendor dependence, no representative evaluation, missing exception handling, low user adoption, and no scale or stop gate. A successful demo does not prove safe or valuable production operation.
How should a bank measure AI implementation success?
Measure business outcome, task quality, failure impact, customer effect, control performance, adoption, latency, operating cost, incidents, overrides, complaints, and model or vendor change. Compare results with the pre-implementation baseline and record an explicit scale, revise, or stop decision.
About Paul Okhrem
Paul Okhrem is an AI Transformation Consultant and Fractional Chief AI Officer. He helps executive teams connect AI use cases, governance, vendors, implementation, adoption, and measurable acceptance. This checklist is free to reuse under CC BY 4.0 with attribution.
Financial-services AI consulting · bank AI company research · financial-services AI statistics