← Security & Governance
Foundational

AI in Regulated Industries: What Banking, Health and Legal Regulators Expect, as of November 2026

Banks, health products and law firms adopt AI under rules written for models, devices and professional duties. Learn the 2026 state of US model risk guidance (SR 26-2), adverse-action rules, India's RBI framework and draft, FDA change-control plans and ABA Opinion 512, and the product features they imply.

CONCEPT / SECURITY & GOVERNANCE

TL;DR: Regulated industries do not need a dedicated AI law to constrain AI products: existing rules for models, devices and professional conduct already do. In US banking, SR 26-2 replaced SR 11-7 in April 2026 and puts generative and agentic AI outside its scope, leaving each bank to decide how far to extend its principles. Adverse-action notices still need specific reasons. India's RBI has a committee framework and a draft on model risk that would make banks accountable for every model they use, including third-party ones. The FDA lets device makers pre-authorise planned model changes. The ABA tells lawyers what their duties mean when they use generative AI. Each becomes a feature: an inventory, a risk tier, a change plan, a disclosure, a human option, a kill switch. This is not legal advice.

The launch path regulators look for

EXPLORE THE FLOW
Launching AI where a regulator will ask questions BEFORE LAUNCH IN SERVICE re-tier after a material change 1 Classify the use each regime: sector and AI law 2 Record it in the inventory including third-party models 3 Set the risk tier purpose and exposure 4 Validate independently effective challenge 5 Plan the changes what may change, how it is tested 6 Disclose and offer a human at the point of use 7 Monitor with a kill switch outcomes, drift, complaints 8 Report incidents under an incident framework RBI draft (June 2026): a regulated entity is accountable for all models it uses. SR 26-2 (April 2026): purpose and exposure determine materiality. FDA's PCCP guidance lets authorised changes ship without a new submission. RBI's draft: tell users it is AI and offer a switch to a human.

Follow the flow, one stage at a time.

Illustrated sequence; animation speed does not represent processing time.
Read diagram description

Classify the use under each regime that applies. Record the model in an inventory, including third-party models. Set its materiality or risk tier. Validate it with independent, effective challenge. Plan how it will change before it ships. Tell people when they deal with AI and offer a human. Monitor outcomes with a kill switch. Report incidents, and re-tier after any material change.

Across sectors the questions repeat: what is the model for, how much harm can it do, who checked it independently, what happens when it changes, who is told it is AI, and who can switch it off.

Banking and lending

United States. The Federal Reserve's SR 26-2, issued on 17 April 2026 with the OCC and aimed mainly at banks with over $30 billion in assets, "supersedes and replaces SR letter 11-7". Its principles centre on materiality ("Model purpose, together with model exposure, determines model materiality") and "effective challenge" by objective experts. It states that generative and agentic AI models "are not within the scope of this guidance", while its principles apply to traditional and non-generative models. For a PM at a bank, that leaves a gap the bank fills with its own policy: the inventory, tiering and validation habits still apply, adapted to models whose behaviour depends on prompts and vendors.

Credit decisions keep their old rule. The CFPB withdrew its circulars on algorithmic adverse-action notices in May 2025, a withdrawal the Bureau says is not necessarily final, but Regulation B still requires the statement of reasons to be "specific and indicate the principal reason(s)", and says a reason such as failing to reach a qualifying score is insufficient. A credit model whose outputs cannot be turned into specific principal reasons leaves the lender unable to meet that requirement, so reason codes are a launch requirement.

India. The RBI's FREE-AI committee report (August 2025) sets out seven principles and 26 recommendations, including a board-approved AI policy, an internal AI inventory, AI incident reporting and AI disclosures in annual reports; it is a report, not a binding direction. The RBI's draft guidance on model risk management (June 2026) goes further: a regulated entity "is accountable for the outcomes of all models used by it", including third-party models; models that cannot be fully explained need compensating controls; users should be told they are dealing with AI and offered a switch to a human; and "kill-switch arrangements" are expected. As of November 2026 it remains a draft.

United Kingdom and EU. The FCA says it does "not plan to introduce extra regulations for AI", relying on the Consumer Duty and the senior managers regime instead. In the EU, creditworthiness assessment of individuals is a high-risk use under Annex III of the AI Act, with those obligations applying from 2 December 2027; see the EU AI Act for PMs.

Health

The FDA has authorised over 1,600 AI-enabled medical devices as of September 2026, by its own non-comprehensive list. Its final guidance on Predetermined Change Control Plans (December 2024, current version August 2025) lets a manufacturer describe planned model modifications, the method to develop and validate them, and their impact up front; once authorised, those changes can ship "without triggering the need for a new marketing submission". The broader lifecycle guidance from January 2025 is still a draft, and FDA's August 2026 paper on generative AI devices is for discussion only.

For a PM, the change plan is the product roadmap in regulatory form: which retraining, thresholds and data changes you will need over the plan's horizon, decided before launch, with the tests that prove each is safe.

The ABA's Formal Opinion 512 (July 2024) applies existing professional duties to generative AI. Lawyers need "a reasonable understanding of the capabilities and limitations" of the tools they use. Putting client information into a tool that learns from it needs the client's informed consent, and boilerplate in an engagement letter "is not sufficient". Hourly work is billed for actual time, and a lawyer "may not charge a client to learn about how to use a GAI tool" used regularly. A legal AI product helps lawyers meet these duties by not training on client data by default, recording sources for every citation, and making time and usage reports available for billing.

From rule to feature

ExpectationProduct feature
Inventory of models, including vendors'A registry every deployment must pass through
Materiality or risk tierA tier assigned at intake that sets the review required
Independent validationEvaluation evidence a separate team can rerun
Planned changeA change plan with test protocols, versioned with the model
Disclosure and human recourseLabels at the point of use and a handoff path
Kill switch and incident reportingA feature flag per use, with an incident runbook
Specific reasons for adverse decisionsReason codes produced with each decision

In the interview

Practise prompts such as "set up AI governance at a bank", "launch an AI feature in a medical device" or "design a legal research assistant law firms will trust". A weak answer says "comply with regulations". A Hire answer names the relevant regime. A Strong Hire answer cites the current rule with its date and status, notices where it leaves gaps (such as SR 26-2's exclusion of generative AI), and turns each expectation into a feature, a tier and an owner, with counsel involved early. This is product guidance, not legal advice.

Practise with set up AI governance at a bank and ship model updates to a cleared medical device.

Next: the EU AI Act for PMs and US and India AI rules.

RELATED CONCEPTS
PRACTISE THIS CONCEPT