Financial services guidelines

This page is guidance, not legal or regulatory advice. Your compliance team decides which rules apply to your workload and what they require. Akka does not interpret regulations or state retention periods.

AI workloads at a financial firm often meet requirements that other workloads do not. This page lists the topics that come up most often, as questions for you to ask your own risk and compliance teams. It is not a gate and it adds no sign-offs to the production readiness checklist or the agentic AI checklist. The questions are based on rules (noted at the start of each section) that often apply to financial firms, depending on the business and the jurisdiction. Other regions have similar rules.

Start by writing down which requirements apply to your firm and your workload: the regulators you answer to, the internal policies you follow, and the products and customers the workload touches. The questions below assume that you have this list.

At a glance, the topics to cover
  1. Model risk. Decide whether the workload is a model under your model risk policy, and who validates it.

  2. Evidence and retention. Decide what you must keep, for how long, and in what form.

  3. Fair treatment. Decide how you test for bias where the workload affects customer outcomes.

  4. Disclosure. Decide what you tell customers about the use of AI and about automated decisions.

  5. Provider concentration. Decide how dependent you are on one model provider.

  6. Operational resilience. Decide how the workload fits your recovery and incident requirements.

Model risk management and validation

Related rule: SR 26-2 from the Federal Reserve and OCC Bulletin 2026-13 from the OCC, the US guidance on how banks manage the risk of models. They replaced SR 11-7 in April 2026. The guidance states that generative and agentic AI are outside its scope, so your model risk team decides how your policy treats an agent.

  • Does your model risk policy treat the agent, the model it calls, or both as a model that needs inventory and validation?

  • Does your model risk policy require that someone independent of the team that builds the agent performs or reviews the validation, and who is that person?

  • What counts as a material change that triggers revalidation: a new model version, a prompt change, a new tool, or a change to retrieval data?

  • Which evidence can you show a validator: the evaluation suite, the thresholds, and the release record of the model, prompts, and tools?

  • How do you monitor the workload after release, and who reviews the results?

Evidence and recordkeeping

Related rules: the SEC and FINRA recordkeeping rules for broker-dealers, which set which business records a firm must keep and for how long.

  • Which interactions, decisions, and communications does your compliance team treat as records?

  • What retention period and storage format does your compliance team require for each kind of record?

  • Does the evidence you keep today cover those records: the prompt, the model, the tool calls, the guardrail outcomes, and the final result?

  • Where does the evidence live, and does its retention match the required period? Confirm the retention of the Akka ledger and logs with your Akka team.

  • Can you retrieve and produce the records for an auditor or a regulator within the time that they allow?

Fair lending and bias testing

Related rule: the US Equal Credit Opportunity Act, which prohibits discrimination in credit decisions and requires a lender to give the reasons for an adverse decision.

  • Does the workload influence a decision that affects a customer, such as credit, pricing, eligibility, or service level?

  • Which groups and outcomes does your compliance team require you to test, and with which data?

  • Does the evaluation suite include those tests, and does a failing result block promotion in your pipeline?

  • Can you explain a decision to the customer or a regulator in the terms that the rules require?

  • Who reviews the results after a model, a prompt, or retrieval data changes?

  • Which uses of AI must you disclose to customers, and where does the disclosure appear?

  • Do customers have a right to a human review or an explanation, and does the workflow provide it?

  • Does the workload use customer data in a way that needs consent or notice, including data sent to a model provider?

  • Which communications that the agent produces need review before a customer sees them?

Third-party concentration risk for model providers

Related rule: DORA, the EU regulation on digital operational resilience. It applies to financial entities in the EU and sets requirements for their contracts with ICT providers.

  • Does your third-party risk policy treat the model provider as supporting a critical or important function?

  • Which contractual terms does your policy require from the provider, for example audit rights, incident notice, and exit support?

  • How many workloads depend on the same provider, model, or region?

  • Do you have a fallback provider or model, and have you rehearsed the switch? See Provider failure and change.

  • What happens to the workload if the provider retires the model that you use?

Operational resilience

Related rule: DORA, which requires EU financial entities to manage technology risk, report major incidents, and recover from disruptions.

  • Does the workload support a critical or important function, and which recovery time and recovery point objectives apply to it?

  • Do the recovery objectives of the workload match those of the services that depend on it? See Disaster recovery.

  • Does your incident process cover an AI failure that raises no availability alert, such as a confident but wrong answer? See Incident playbooks.

  • Which incidents must you report to a regulator, and within what time?

  • Have you tested the stop path for the agent, and who can use it without a code change? See Stopping and reducing an agent.

Common pitfalls

  1. The firm treats the agent as ordinary software. A model risk policy may require inventory and validation before release.

  2. The evidence exists but the retention does not match. Confirm the retention of every record source against your own requirement.

  3. Bias tests run once. A change to a model, a prompt, or retrieval data can change the results.

  4. One provider serves many workloads. The dependence is visible only when you count the workloads together.

  5. Nobody collects the list of applicable requirements. The questions on this page cannot be answered without it.