For a consumer lender building on a hosted model, the places where model output reaches an applicant, a borrower, an underwriter or a decision engine start at prequalification messaging and end at collections outreach. Most of them sit in the middle, in servicing and account management, where nobody calls the work underwriting.
Those places carry more weight once the Anthropic Usage Policy names the decision. The version effective November 12, 2026 lists Credit as one of its High-risk Areas and defines it as:
"approving, denying, or setting the amount, rate, or limit of a loan, mortgage, credit line, or similar product, or reducing or closing an existing one, or producing a credit score or creditworthiness assessment used for such decisions."
Read the definition against that lifecycle. Approving and denying are the obvious entries. Rate and limit reach pricing and line management, reducing or closing reaches account management long after origination, and the last clause reaches scoring work that many programs file under analytics.
What follows is the inventory I would build and the questions I would send with it. It is method, not legal advice, and it gives no verdict on which touchpoints the policy covers. That verdict belongs to counsel, with the architecture in front of them.
Credit used to sit under Finance
Anthropic announced the update on October 8, 2026 (opens in new tab), and on the high-risk material it says plainly that the requirements did not change. The section that states them did. Credit decisions were already covered. The September 15, 2025 version (opens in new tab) listed loan approvals and creditworthiness under a broader Finance use case. The November 12 version gives Credit its own area and spells out rate, limit, reducing and closing. Users had often asked whether a given use case carried the requirements, so Anthropic rewrote the section to answer that more directly. It "now lists which kinds of recommendations are covered and which are not."
The requirements are short. For a High-risk AI Recommendation, the Usage Policy (opens in new tab) requires review and disclosure. The review sentence reads: "A qualified person must meaningfully review a High-risk AI Recommendation, with the authority to change or modify it, before it is delivered as advice to an individual who might rely on it or before it is used to implement a decision." A qualified person has the training or experience to evaluate the output in that field, holds a license where applicable law requires one, and stays responsible for what is delivered or decided. Every Usage Policy quotation in this piece comes from the version with the November 12, 2026 effective date.
Two phrases in that sentence carry most of the engineering. "With the authority to change or modify it" means the reviewer needs a control that edits or overrides the output, and an acknowledge button changes nothing. The two "before" clauses put the review upstream of the send and the write, which reads as ruling out a sampling queue someone works through after the letters have gone out.
The policy also defers to the law around it: the high-risk requirements "do not replace local legal requirements and restrictions," and users should default to local law where it conflicts with them or adds requirements. Its scope names customers who reach Claude through cloud providers and authorized resellers, so a lender on a cloud platform is reading the same text as one on the API.
Map the credit lifecycle before anyone argues about coverage
Each touchpoint gets the same architecture questions. Does the model deliver the recommendation, or feed one that a person or a rule delivers? Does it exercise judgment about this individual? Who can change the output before it reaches the person or the decision, and does that person meet the policy definition of qualified? The answers go to counsel with the data flow attached, because a coverage opinion written against a prose summary is an opinion about a system nobody built.
- Prequalification messaging. The model writes what a prospect sees before applying: which product fits, what they may qualify for, a likely rate range. The design question is whether the model applies the circumstances of the prospect to an amount, rate or limit, or phrases product information inside bounds a pricing engine already set. The policy excludes general information "without applying it to the individual's circumstances," so where the personalization happens is most of the question for counsel.
- Underwriting assist. Bank statements go to the model, which extracts income, flags inconsistencies and drafts the memo an underwriter reads. Once that memo ends in a recommended disposition or a suggested limit, it reads a lot like a creditworthiness assessment used for the decision. Counsel has to place the line between internal analysis and "the ultimate High-risk AI Recommendation delivered."
- Line increases and decreases. Volume changes the review problem here. A model proposing new limits across a portfolio can produce thousands of outputs in one run, and approving the batch is a different act from reviewing an account. The architecture fork: a fixed, versioned policy sets each limit with the model only assembling inputs, or the model weighs the situation of the account holder. Ask counsel whether batch approval is meaningful review. Decreases deserve extra care, since the definition names "reducing or closing an existing one" outright.
- Collections outreach. The work starts with reminders to borrowers who are behind, drafted or sent by the model. Some designs go further and have it propose payment plans or settlement amounts. Collections is also where an agent often holds tools that send messages and update the account. Counsel decides which collections actions set an amount, rate or limit, or reduce or close the line.
- Adverse action drafting. The model drafts the notice and the statement of reasons. This touchpoint closes more than one entry above, because Regulation B requires notice of adverse action on an existing account as well as on an application. The design question is whether the model phrases reasons the decision system already produced or selects which reasons to state. The policy excludes explaining a decision a qualified person has already given, "as long as it is not an additional High-risk AI Recommendation." Ask counsel where phrasing ends and selection begins.
Every entry also carries a contract question: what the master agreement says about the Usage Policy, whether a cloud provider or reseller adds its own paper, and who on the deployer side attests that review and disclosure exist. File the answers with the diagram.
The exclusions are design choices
The policy frames its recommendation exclusions as "non-exhaustive examples" of workflows "that are related to, but do not actually provide, High-risk AI Recommendations." Three of them shape most credit architectures:
- "Internal drafting, research, summarization, or analysis that is not the ultimate High-risk AI Recommendation delivered."
- "Carrying out a decision a person has already made, or applying a fixed rule, formula, or algorithm where the model exercises no judgment about the individual."
- "Business operations that do not advise or decide about a specific individual."
Each one describes a property of a system, and a deployer sets those properties when it designs the flow.
Where the output stops. Internal analysis stays internal when a person reads it, forms the recommendation and records it as their own. The evidence lives in the system: the reviewer writes the disposition in a field the model cannot fill, and overrides are logged next to the output they replaced. If the screen shows a model recommendation beside an approve button, the design document is describing a system the reviewer never sees. Counsel should see the screen.
Where the judgment lives. A fixed rule puts the judgment in the rule. A limit grid approved by credit policy, versioned and tested, with the model extracting the inputs, is a different architecture from a model asked what limit this customer should get. Whether extraction is itself judgment about the individual is a question for counsel. Whether the line between extraction and decision is enforced in code is a question for engineering, and it is the act-or-interpret boundary I would draw for any AI system, now with a vendor term riding on it.
Whether the output is about one person. Portfolio loss forecasting, policy drafting and model documentation do not advise or decide about a specific individual. The design risk is the feature request that adds a per-account action column to the portfolio view.
An exclusion that rests on architecture lasts as long as the architecture does. A prompt revision that starts suggesting a limit can move a touchpoint without anyone filing a change request. So any change to the prompt, the output schema, the tool set or the model version behind a credit touchpoint goes back through the review that produced the original coverage question.
Credit has no favorable-decision carve-out
The policy offers one place where review can be skipped. Where permitted by law, human-in-the-loop review is not required before acting on a High-risk AI Recommendation wholly favorable to the individual, in four listed circumstances:
- Paying an insurance claim.
- Approving coverage or prior authorization for care a clinician has ordered.
- Granting eligibility for a healthcare service or program.
- Granting or continuing a public benefit or service.
Credit approval is not on the list. A straight-through approval flow that skips review because the answer was yes finds no carve-out in the text, and the same passage closes the next argument: "A partial approval, reduced amount, or approval with conditions is not considered 'wholly favorable.'"
In credit, conditional approvals are ordinary. A line below the amount requested and a counteroffer at a higher rate are each an approval with something attached. Regulation B already treats them as events, with notice due within 30 days of a completed application whether the creditor approves, counteroffers or takes adverse action. That leaves an engineering question: where the qualified person sits in a flow built to answer in seconds, and what that person sees before the decision is implemented.
Telling the person AI was used
The disclosure requirement is one obligation and one relief. An individual who receives advice or who is the subject of a decision based on a High-risk AI Recommendation "must be clearly told that AI was used to produce it." The relief: "Users do not need to identify Anthropic, Claude, or the model used."
A separate rule governs conversation. Consumer-facing chatbots and AI agents that communicate directly with external users must disclose that the user is dealing with AI rather than a human, "at a minimum at the beginning of each chat session or in the product interface." That rule does not depend on a High-risk AI Recommendation, so a collections assistant that chats with borrowers raises it whether or not any credit decision is in play.
For the inventory, disclosure is a placement problem. Where does the statement sit in an approval letter, a counteroffer and an adverse action notice, and which channel carries it when the decision arrives by text message? Those documents are already regulated, so I would route the AI disclosure wording through the compliance review that approves the notice it sits in, and version the two together. "Clearly told" is a standard someone will eventually measure against the actual letter.
Agents act, and the lender still answers
The agent clause is two sentences: "Agentic use cases must comply with the Usage Policy. Users are responsible for ensuring that agents they build or deploy, including actions those agents take through tools, browsers, or connected systems, comply with this Usage Policy."
For a credit program, that clause makes the tool list part of the coverage map. The entries that matter most are agents holding a tool whose effect the Credit definition names, such as changing a limit or closing an account. So the inventory needs a second column: for each touchpoint, every tool the agent holds and every system it can write to.
The qualified-person checkpoint belongs at that tool boundary. A system prompt telling the agent to seek review before changing a limit is a reminder, and a reminder in prose binds nothing. A tool that refuses to execute until a named reviewer with authority has approved that specific action, and logs the approval beside the output it approved, is something an examiner can test.
Then put the vendor term in its place. Anthropic writes the Usage Policy and revises it on its own schedule, which the announcement describes as yearly. The obligations with statutory weight stay with the lender. ECOA and Regulation B require notice of action taken, and under 12 CFR 1002.9 (opens in new tab) the statement of reasons "must be specific and indicate the principal reason(s) for the adverse action." The same section calls a statement that the applicant missed a qualifying score insufficient.
When adverse action rests in whole or in part on a consumer report, FCRA section 615 (opens in new tab) adds its own notice duties, including a statement that the reporting agency did not make the decision. Fair lending work I am part of is still in progress with external legal firms, and nothing here should be read as its result. I made the case for ECOA, Regulation B and FCRA as the governing frame in the fair lending piece, and no vendor policy moves them.