The OCC's AI examiner framework (Bulletin 2026-54) covers data governance in section 2.6, privacy in section 2.7 and compliance in section 2.8. It describes risks such as unclear data lineage, models that memorize or infer customer information, and AI that gives customers wrong answers about fees, rates, account status or policies. It creates no new rules: section 2.8 says AI does not change a bank's underlying legal obligations, and the only laws it names by heading are the Bank Secrecy Act, sanctions and consumer compliance.
TRACKED DOCUMENTS: OCC Bulletin 2026-54
What does the OCC framework say about data, privacy and compliance for AI?
Three of the framework's sixteen components deal with these topics: data governance (section 2.6), privacy (section 2.7) and compliance (section 2.8). Together they describe what can go wrong with the data an AI system uses, the customer information it touches, and the laws a bank already has to follow.
The framework is the OCC's November 2026 document for its examiners, sent out with OCC Bulletin 2026-54. The OCC uses the term “Super Intelligence (SI)” for what everyone else calls AI, so this page writes AI. The bulletin says the framework “does not establish supervisory expectations for banks”. It is a resource examiners may use. Its verbs are “may” and “could”, and this page keeps them.
The framework puts all three components in its “Control” pillar (section 1.2), next to human oversight, model risk management and cybersecurity. For the full text of every component, see the section-by-section reference guide or the OCC regulation page.
What does data governance mean for an AI system?
Put simply, data governance is knowing where an AI system's data came from, whether it is good enough, and whether the system is allowed to use it. Section 2.6 opens with the reason: “The reliability of an SI system depends on the quality, lineage, representativeness, and permitted use of the data used to train, fine-tune, and operate it.”
A few of those words need a plain meaning:
- Lineage is the history of a piece of data: where it started and what was done to it on the way. The framework's glossary calls it “the history of the processing of a data element”.
- Representativeness means the data looks like the real customers and cases the system will meet.
- Fine-tuning is extra training a bank or vendor gives a ready-made model so it fits a task.
- Synthetic data is data made up by a computer to stand in for real records.
Section 2.6 says large data volumes, third-party sources and synthetic data could make these qualities harder to assess. It says data governance may cover quality, lineage, representativeness, access, retention and use “throughout the SI lifecycle”.
What does the framework say about training data, sources, prompts and outputs?
It treats all of them as data to govern, including the new data an AI system creates as it runs. Section 2.6 says the bank may consider both data given to the system and “new data created through prompts, outputs, embeddings, logs, feedback, or synthetic generation”.
A prompt is the question or instruction typed into the system. An embedding is a numeric version of a text that the system uses to search for related documents. The examples in section 2.6 are:
- Data lineage. The bank may assess the origin, licensing, changes and limits of “training, fine-tuning, source, and production data, including data embedded in third-party foundation models”. A foundation model is a large general-purpose model, often from a vendor, that other tools are built on.
- Drift and representativeness. Monitoring may spot changes in the people or data going into the system before results get noticeably worse. Drift means the inputs slowly stop looking like the data the system was built on.
- Synthetic data. It may be reviewed for quality, approval, representativeness and bias, because flaws “can be difficult to identify and may be amplified through training”.
- Permitted sources and access. Technical controls may keep the system to approved data sources and block information outside its purpose.
- Data architecture. Data models could need to change as AI creates new kinds of data.
Search-based AI appears in the framework's glossary as “retrieval augmented generation (RAG)”. Sections 2.6 and 2.14 mention “source” data and “retrieved sources”. In practice, a chatbot that answers from the bank's policy library falls under the “permitted sources” example: the question is whether it can only read the approved library.
What does the framework say about customer information going into AI?
Section 2.7 says AI “can create privacy risks beyond conventional data-handling concerns”. It names three: models may keep or repeat training data, guess sensitive facts about people, and use bank or customer information in vendor systems the bank cannot fully see into.
It lists privacy considerations that may apply to each use case: applicable law, permissible use, data minimization (using only the data the task needs), access, retention, deletion, notice, consent, vendor use and cross-border processing. These may be revisited when the model, provider, data or purpose changes. The four examples:
- Memorization and inference. A model might reproduce specific training examples, or infer sensitive traits it was never given, creating risks “distinct from a conventional data breach”.
- Prompt and output retention. Prompts, outputs, embeddings and logs could become stores of sensitive information and may warrant access controls and risk-based retention periods.
- Vendor data use. Contracts and settings may make clear whether bank data can be used to train or improve models offered to other customers of the vendor.
- Cross-border processing. Where data is stored and processed could affect residency, transfer, notice, consent and third-party obligations.
For example, a servicing agent pastes a customer's account history into an AI assistant to draft a letter. The prompt itself now holds customer data. Section 2.7 points to who can read that log, how long it is kept, and whether the vendor may learn from it. The framework does not name a specific privacy law. Vendor questions continue on the vendor and third-party AI page.
Does the OCC framework create new consumer-protection rules for AI?
No. Section 2.8 says: “SI does not change the bank's underlying legal obligations and has the potential to facilitate a bank’s compliance with them.” The bulletin adds that nothing in the framework is meant to set new regulatory expectations or enforceable standards.
What section 2.8 describes is why AI can make existing compliance harder. Compliance risk could rise when a system gives wrong or inconsistent results, is used outside its purpose, or lacks controls for the laws that apply. AI's “complexity, opacity, speed, and capacity to act” may make it harder to see how a rule applies, explain an outcome, catch a breach, or keep the system inside its approved limits. A bank may consider compliance at every stage: design, development, buying, launch, monitoring, change and retirement.
Which laws and risk areas does the framework name?
Section 2.8 names two areas by heading: the Bank Secrecy Act and sanctions, and consumer compliance. It does not name fair lending or UDAAP anywhere in the text.
The Bank Secrecy Act (BSA) is the US law that requires banks to help detect money laundering, for example by monitoring transactions and reporting suspicious activity. Sanctions rules bar dealings with listed people and countries. Section 2.8 says banks may use AI in transaction monitoring, risk assessment, case management, reporting and sanctions screening. Bad or unrepresentative data, or poorly set thresholds, “could increase false negatives, false positives, or other compliance risks”. A false negative is a real case the system misses. A false positive is an innocent case it flags. More on these uses: AI in AML and AI in KYC.
Consumer compliance means following the laws that protect individual customers. Section 2.8 says AI used in customer service, marketing, servicing, collections, disclosures or other consumer-facing work could create this risk, “especially if outputs are materially inaccurate, misleading, or inconsistent with product terms”. A disclosure is the required information a bank gives a customer about a product.
Fair lending laws ban discrimination in credit, and UDAAP is the ban on unfair, deceptive or abusive acts or practices. The framework does not mention either by name. Its closest points are the glossary definition of bias, the bias review of synthetic data (2.6), and section 3.1, which asks whether an output can be traced to a reason good enough for disclosure duties, giving “adverse action notices” as the example. An adverse action notice is the letter telling an applicant why credit was refused. For credit models, see AI credit underwriting.
| Risk area | AI-specific concern in the text | Passage |
|---|---|---|
| Bank Secrecy Act and sanctions | Bad or unrepresentative data, or poorly set thresholds, raising missed or false alerts | 2.8 |
| Consumer compliance | Outputs that are materially inaccurate, misleading or inconsistent with product terms | 2.8 |
| Hallucinations to customers | Wrong statements about fees, rates, account status or policies | 2.8 |
| Privacy | Memorization, inference, retention, vendor use, cross-border processing | 2.7 |
| Data quality and bias | Unclear lineage, unrepresentative or synthetic data whose flaws are amplified by training | 2.6 |
| Disclosure obligations | Outputs that cannot be traced to a reason good enough for notices such as adverse action notices | 3.1 |
| Fair lending, UDAAP | Not named in the framework | none |
What if an AI system tells a customer something wrong?
Section 2.8 names this risk directly: “Hallucinations concerning fees, rates, account status, or policies could contribute to legal or customer-harm exposure.” A hallucination is an answer that sounds convincing but is false.
The controls it describes are three, all phrased as things a bank may consider:
- controls that “limit automated representations”, meaning limits on what the system may state to a customer on its own,
- verification “in higher-impact contexts”, meaning a check before the answer counts,
- support for “prompt correction and remediation”, meaning fixing the error and putting the customer right.
For example, a chatbot tells a customer that an overdraft fee will be waived when the product terms say it will not. Under section 2.8 that is a customer-facing representation the bank may need to limit, check or correct. The framework sets no numbers, no error rate and no deadline for this. Testing for hallucinations is covered in section 2.13, and fixing harm to a whole group of affected customers in section 2.10. See the model risk and testing and resilience pages, and AI chatbots in banking for how banks use them.
What could a bank have ready on data, privacy and compliance?
In practice, a bank could keep a short record per AI use case that answers the questions sections 2.6 to 2.8 raise. This list is ours, not the OCC's, and the framework does not require it.
- Where the training, fine-tuning and source data came from, and which sources the system may read.
- Whether prompts, outputs and logs hold customer data, who can see them, and how long they are kept.
- What the vendor contract says about training on bank data, and where data is processed.
- Which laws apply to the use, and how testing and human overrides support them.
- What the system may tell customers on its own about fees, rates, account status and policies, and how a wrong answer gets corrected.
Next steps: the exam preparation page and the risk tiers, where data sensitivity is one of the scoring factors (section 3.1).
Does the OCC framework ban putting customer data into AI tools?
No. Section 2.7 lists privacy points examiners may look at, such as permissible use, data minimization, retention and vendor use. It does not ban any use of customer data. It describes risks, not prohibitions.
Does the framework mention fair lending or UDAAP?
No, neither term appears in the text. Section 2.8 names Bank Secrecy Act and sanctions compliance and consumer compliance. Bias is defined in the glossary, and section 3.1 mentions adverse action notices as an example of disclosure obligations.
Who is responsible if a bank's chatbot gives a customer a wrong fee?
The framework does not assign blame for a single case. Section 2.8 says hallucinations about fees, rates, account status or policies could create legal or customer-harm exposure. Section 2.4 says accountability for managing the bank's risk generally stays with the bank even when vendors build or host the system.
Can a vendor train its AI on our bank's data?
The framework does not answer yes or no. Section 2.7 says contracts and technical settings may make clear whether bank data can be used to train or improve models offered to the vendor's other customers.
Are prompts and chat logs treated as sensitive data?
They can be. Section 2.7 says prompts, outputs, embeddings and logs could become stores of sensitive information. It says they may warrant access controls and retention periods based on risk.
Is synthetic data a safe way around privacy concerns?
The framework does not call it safe. Section 2.6 says synthetic data may be reviewed for quality, approval, representativeness and bias. Its flaws can be hard to find and may be amplified through training.
Does the framework set deadlines for any of this?
No. The framework has no effective date and no compliance deadline. The bulletin says it is a resource for examiners and does not set supervisory expectations for banks.
- Super Intelligence: Supervisory Framework (full text, PDF) — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
- OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
When this rule moves, the next morning's brief says so
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning