Written for Risk, fraud, compliance, operations and technology leaders in banks and other financial institutions, and the sponsors who have to fund a first case.
What's in this playbook
Financial institutions are not short of AI ideas. They are short of first cases that can travel: a named process, a signed baseline, a system on the inventory the institution already knows how to run, a person who can stand behind a close, a filing or a customer-facing decision, and a way to stop. This playbook is the single financial-services document behind the finance industry pages. It is complementary to the financial-services practitioner's framework already in the Resource Library, with fraud detection, account takeover and financial-crime operations in depth, on purpose. It is not a near-duplicate of that paper, not a survey, and not a close of any named regime.
It maps six families of work. Each family has an honest first case and a case that is too wide. A triage assist that scores existing rule-based alerts, attaches the evidence it used, and orders a queue for a named investigator — with a person on every close, every escalation, and every customer-facing hold — is an honest first case in fraud detection. A model that automatically closes low-scoring alerts, or that files a report without a person, is not. The same distinction is drawn for transaction monitoring, credit and lending (not recommended as a first case; an honest later case is decision support for a named credit officer), onboarding documents, servicing and internal process, and case-narrative drafts: AI as a triage and evidence layer first; an unsupervised close, filing, or credit decision is not a first case this playbook will recommend.
How the tests and the method are meant to be used
Fourteen design tests are applied to every case. The answers will differ. The standard does not. A missing counsel answer, an undefined "review," or a tool already running and off the inventory, is a reason to pause, not permission to proceed. Counsel and the risk function decide what a supervisory statement, a reporting duty, a fair-lending rule, or a creditworthiness provision requires, and when. This playbook will not make those determinations, and it will not recite dates. Hosting is decided before a case is funded. Where data cannot leave the environment, AsymiLink AI runs on your own hardware, with an optional air gap. Built to SOC 2 standards is the exact wording we use; it is not a certification of your system.
The illustrative organization is the same hypothetical community bank as in the 2026 diagnostic, the readiness and governance frameworks, and the financial-services paper: overall 22, Developing, with two generative-AI assistants not on the model inventory, and a next-step target of 25, still Developing — a next step, not a timeline, and not a promise. Its honest first work is to put those assistants on the inventory, with an operator and a stop, and only then to consider a fraud-triage assist for investigators — not an auto-closing model, and not a model that files a report. The financial method is The Complete Guide to AI Automation ROI. This playbook will not invent a financial-services-sector result. The 3.8x average first-year ROI across client engagements is an average, not a forecast or a commitment for any individual project. ROI is measured across client engagements against a baseline agreed before work starts. It is an average, not a forecast for any financial-services project.
Six families of work
each with an honest first case, a case that is too wide, and a defined gate. Each item below is the shape of the map; the full wording, honest first cases and “too wide” warnings are in the PDF.
F1
Fraud detection and account takeover
F2
Transaction monitoring and financial-crime operations
F3
Credit and lending decisioning
F4
Onboarding and document operations
F5
Servicing, operations and internal process
F6
Case narratives and investigator support
The PDF gives, for each family, an honest first case, a case that is too wide, and a defined gate. The summaries above are the shape of the map, not a substitute for that wording.
Fourteen design tests (excerpt)
Fourteen design tests are applied to every case. The answers will differ. The standard does not. The full wording is in the PDF and the companion workbook.
T1–early
inventory and owner; proportionate consequence; a person on any close, filing or customer-facing decision, with review defined; counsel asked in time; explanations, lineage and a frozen evaluation set.
Mid-tests
hosting decided up front; a write that closes or files as a different design from a read; an audit trail that can reconstruct a day; a named operator and a switch-off; training and reuse written down.
Later tests
fairness examined for this use, not scored against a threshold we invented; honest notice; a bounded first case that raises the path; a signed baseline on the ROI method, and monitoring after launch that the organization that owns the system writes down.
A completed test sheet is not a certificate, not a compliance score, and not a finding that a system is safe or aligned with a named regime.
Get the guide and workbook
Enter your name, work email and company. The files appear on this page immediately. A backup download link is also emailed when email delivery is available.