AI in Financial Services: A Practitioner’s Framework A use-case map with design tests, and a way to treat model risk and regulation as design input — not a survey of the market, and not a ranking of use cases by return. Edition: November 2026. 21 pages. Written for: Risk, compliance, operations and technology leaders in banks, insurers and other financial institutions, and the sponsors who have to fund a first case. Canonical URL: https://asymilinkmeta.com/resources/financial-services-ai-report Written for the search query: AI in financial services practitioner framework Key takeaways - This is a practitioner’s framework, not a 2026 state-of-the-market survey, and not a ranking of use cases by return. - Nine use cases, U1 through U9, each with an honest first case and a “too wide” warning. - Ten design tests, including model-risk inventory, a person who can stand behind a decision, and counsel asked what applies. - Two illustrative institutions — the same mid-size insurer and community bank as in the 2026 diagnostic — are pictures, not client case studies. What's in this framework This 21-page document is a practitioner’s framework for risk, compliance, operations and technology leaders in banks, insurers and other financial institutions. The library has sometimes described it as a 2026 state-of-the-market survey, with a ranking of use cases by return. It is not a survey. It will not publish a sector adoption rate, and it will not rank use cases by return. What it is, instead, is a use-case map — nine use cases, U1 through U9 — with an honest first case and a “too wide” warning for each; ten design tests to apply before a case is funded; and a way to treat model risk and regulation as design input. Counsel and the risk function are asked what applies. Their answer is a design input. This framework is not that answer, and it is not legal advice. It uses two illustrative institutions: the same mid-size insurer and the same community bank as in the 2026 diagnostic and the governance framework. They are labeled illustrative. They are not client case studies, and they are not a sample of the sector. A companion workbook holds the nine use cases, the ten tests, and the two institutions. A completed workbook is not a certificate, and it is not a finding that a system meets a named regime. How the map and the tests are meant to be used The job is a process result someone already owns, not a platform, and not “explore AI in banking.” A signed baseline is agreed with the process owner before design is funded; the method is in the ROI guide in this library. A source of truth is sampled: the file, the system of record, and a recent quality sample have been seen. The mechanism fits — a rule, a workflow, a document process, a model, an agent or a person — chosen in writing because a simpler mechanism cannot do the job. If the institution already practices model-risk management, the system is on that inventory, or a written decision says why it is not a model for those purposes. A “pilot” is not an exemption. A person can stand behind the decision, and a customer can be told why, to the standard the institution already owes. An operator who is not only the builder, a log, a sampling plan, and a way to stop the system without a project, are in the first design. Production cost is in the first case. The first case can travel to production under these tests, or it is explicitly a learning case that will not be called production. A demo is not a deployment. Internal operations — finance, reconciliation, intake — are often the more honest first case. A new credit model, or a life-and-health pricing system, going to production off the inventory, is not. The governance framework, the HIPAA design framework (where health information may be in scope), the 2026 diagnostic and the readiness assessment in this library are companions. This framework will not stand in for them. Nine use cases (U1–U9) A map, not a ranking by return. Each use case in the PDF has an honest first case and a “too wide” warning. The summaries below are the shape of the map. U1. Customer onboarding, KYC and AML operations. Identity, due diligence, screening, and the file that has to stand behind a decision to onboard, refuse or escalate. U2. Credit decisioning for natural persons. Underwriting, scoring, a credit memo, a decision, and the explanation the institution already owes. U3. Fraud operations. Detection, investigation, a decision to hold, release or escalate, and the customer cost of being wrong in either direction. U4. Servicing and inquiries. Questions by phone, chat and email; a look-up; a hand-off; sometimes a small, permitted action. An honest first case is often read-only, one channel, a closed set of inquiry types. U5. Regulatory reporting and compliance operations. Extracting, assembling, checking and filing evidence the institution already has to produce, with lineage and a person accountable for the filing. U6. Insurance claims. First notice, the file, straight-through or investigation, a decision, a payment or a denial, and a person who can stand behind it. U7. Insurance underwriting, including life and health. A submission, a risk view, a price or a refer, and an explanation. On the inventory, with a person on the refer. U8. Internal operations — finance, reconciliation, intake. Often the more honest first case: a named internal process, used to raise measurement and governance, not only to save hours. U9. Wealth, guidance and anything that can sound like advice. An honest first case is a read-only assistant on a closed set of factual questions, with a recorded hand-off the moment the conversation leaves that set. “End-to-end claims, without an adjuster,” a conversational product that recommends or persuades, or a pricing model off the inventory, are “too wide” for a first case in this framework. Ten design tests (excerpt) The same standard for a document assistant and for a system that would inform a credit, claims or underwriting decision. A failed T5, T6 or T7 is usually a reason to pause. T1–T4. A named process result; a signed baseline, using the method in the ROI guide; a source of truth, sampled; and a mechanism that fits, chosen in writing because a simpler one cannot do the job. T5. On the inventory, or a written reason why not. If the institution already practices model-risk management, a “pilot” is not an exemption. T6–T7. A person can stand behind the decision, and a customer can be told why, to the standard the institution already owes. An operator, a log, a sampling plan, and a switch-off. T9–T10. Obligations named, with counsel: what applies is a design input, and this framework is not that answer. The first case is bounded, or it is honestly a learning case. A demo is not a deployment. Frequently asked questions Q: Is this a 2026 state-of-the-market survey? A: No. It is a practitioner’s framework: nine use cases and ten design tests. It will not publish a sector adoption rate, and it will not rank use cases by return. The library has sometimes described it as a market survey. That description was wrong. Q: Is this legal or regulatory advice? A: No. Counsel and the risk function are asked what applies. Their answer is a design input. This framework is not that answer, and it will not recite a date. Q: Are the institutions in the document client case studies? A: No. The mid-size insurer and the community bank are the same illustrative institutions used in the 2026 diagnostic. They are labeled illustrative, and they are not a sample of the sector. Q: What is in the companion workbook? A: The nine use cases, the ten tests, and the two illustrative institutions. A completed workbook is not a certificate, and it is not a finding that a system meets a named regime. This page is the public summary. The full PDF and companion workbook are available through the request form on the page. There is no public file URL that bypasses that form.