Compliance Guide | Financial Services

    PCI DSS 4.0 Compliance for LLM Applications and AI Agents

    An LLM application or AI agent can fall within PCI DSS 4.0 scope when prompts, retrieved context, outputs, or tool calls store, process, or transmit account data, or when the system connects to the cardholder data environment. Compliance leaders should map each AI execution path to the relevant requirement areas (data protection, access control, logging and monitoring, scoping), enforce the controls that must act before data moves at runtime, and retain per-interaction evidence an assessor or examiner can test.

    Audit-ready evidence for AI interactions and tool calls

    • The identity of the user, application, or agent making the request
    • The policy version and rule that applied to the interaction or tool call
    • The action taken: allowed, redacted, blocked, or escalated, with the approver where relevant
    • The tool or MCP server targeted and the data classes detected, without storing the raw account data
    • A timestamp and a tamper-resistant record that can be exported in a format a QSA can sample
    • Retention and residency settings that match your documented policy

    Where LLM applications and AI agents enter PCI DSS scope

    PCI DSS applies to system components that store, process, or transmit account data, and to components that are connected to or could affect the security of the cardholder data environment. An AI system does not need to be built for payments to meet that description. A support assistant that accepts a pasted card number, a retrieval index built from case notes that contain account data, or an agent with a tool that queries a payment record can each bring the model, its orchestration layer, and its connected tools into the assessment conversation.

    Agents widen the question. NIST describes AI agents as systems that embed general-purpose models in software scaffolding so the model can manipulate tools and take actions beyond text output. Each tool is another path along which account data can travel, so scoping analysis should cover every tool an agent can invoke, not only the model endpoint.

    PCI DSS assessments conclude with an Attestation of Compliance based on the Report on Compliance, and both describe a defined scope. How AI components are represented in that scope is a decision to make with your QSA early, before the architecture is fixed. Avoid describing any AI deployment as certified or compliant in general terms. The attestation covers the assessed scope and nothing wider.

    Mapping AI execution paths to requirement areas

    The table links each path to the exposure it creates and the PCI DSS requirement areas most likely to be tested. Confirm exact requirement numbering and applicability against the current standard and with your QSA.

    • Prompts: Exposure: account numbers entered by users or passed by upstream systems. Areas: data protection, transmission security, scoping.
    • Retrieved context (RAG): Exposure: indexes and retrieved passages that contain account data. Areas: data protection, access control, scoping.
    • Model outputs: Exposure: responses that repeat, infer, or reveal account data to the wrong audience. Areas: data protection, access control.
    • Tool and MCP server calls: Exposure: agent actions against systems holding account data, under whatever identity the agent uses. Areas: access control, authentication, logging and monitoring.
    • Conversation memory: Exposure: account data persisted between turns or sessions. Areas: data protection, retention, scoping.

    Controls that must be enforced at runtime, not reviewed after the fact

    The distinction that matters to an assessor is whether a control acted before the data moved or only described it afterward. A reviewer who reads transcripts next week can detect that a card number reached a model, but cannot undo it. Controls that decide whether data is allowed to reach a model, a retrieval store, or a tool therefore need to operate in the execution path.

    Three control types fit this pattern. The first is data handling: detecting account data in prompts, context, and outputs, then redacting or blocking before the next hop. The second is access decisions for tool calls: the agent's identity, the tool requested, and the data involved are evaluated before the action completes. NIST's NCCoE work on agent identity and authorization makes the same point, that controls must be applied as agents access diverse data sets, tools, and applications. The third is escalation: where a request is permitted only with human approval, the approval must happen before the action, not as a retrospective sign-off.

    Runtime enforcement has limits. Redaction before a model call can reduce the account data a model sees, but it does not by itself settle whether the surrounding systems are in scope. Where account data legitimately must reach a component, that component is an architectural scoping decision, handled through segmentation and assessor agreement rather than a filter. Treat runtime enforcement as a way to narrow data flows and prove the narrowing, not as a substitute for scoping.

    Evaluating allow, redact, block, and escalate

    Each enforcement action suits a different situation, and a sound policy uses all of them. Allow applies when the identity, data, and destination are permitted, and the decision should still be recorded. Redact fits cases where the task can proceed without the account number, such as summarizing a dispute. Block fits cases with no legitimate need, such as a model call carrying full account numbers to an external endpoint outside the approved boundary. Escalate for human approval fits rare, consequential actions where a person with the right authority must decide before the tool call executes.

    The tradeoffs are practical. Aggressive redaction can degrade answers and drive users to unmanaged tools. Detection will miss some formats and produce some false positives, so policy owners need a way to tune rules and see the effect. Escalation adds latency and an approver workload that must be sized. Evaluate any approach on whether each action can be applied to prompts, retrieved context, outputs, and tool or MCP calls, and whether the decision happens before data reaches the model or tool.

    Deployment questions belong in the same evaluation. Check where the enforcement layer runs, whether it can be self-managed inside your cloud so account data does not leave your boundary, and whether retention and data residency are configurable. These settings affect scope and what evidence you can legitimately keep.

    Where runtime enforcement from Trussed fits

    Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, at the level of individual interactions and tool calls. It can allow, redact, block, rate limit, or escalate for human approval, and the enforcement itself produces continuous audit-ready evidence. It deploys self-managed in the customer cloud, with configurable retention and data residency.

    Be precise about the boundary. Trussed's shipped evidence mappings cover HIPAA, NAIC, NYDFS, the EU AI Act, and NIST. Mapping that evidence to specific PCI DSS requirements is work to do with your QSA, and it does not replace the assessment. Trussed's production experience is in regulated sectors, with insurance where it is deepest, so financial services teams should validate fit against their own cardholder data environment and assessor expectations.

    A practical next step is to inventory every AI execution path, record which could carry account data, document a risk assessment for those systems, define agent identities and authorization boundaries before granting tool access, and write data-handling rules as testable policy checks with captured evidence.

    Account data paths in an AI interaction

    Prompt

    Users or upstream systems may submit primary account numbers.

    Retrieved context

    Retrieval can pull account data into the model's input.

    Model output

    Responses can repeat or reveal account data.

    Tool call

    Agents act on systems that may hold or transmit account data.

    Evidence

    Each decision is recorded with identity, policy, and action.

    Test your AI execution paths against assessor questions

    Walk through how runtime enforcement and per-interaction evidence apply to your prompts, retrieved context, and agent tool calls.

    Request a Demo