Implementation Guide

    Routing n8n AI Agent Node Calls Only to BAA-Covered LLM Providers

    To keep n8n AI Agent calls inside BAA coverage, define coverage per provider, product and endpoint, then enforce an allowlist in the execution path of every model call. Configuration and periodic review cannot stop a workflow author from swapping a model node, changing a base URL or adding an HTTP request. Route all model traffic through one enforcement point, restrict n8n egress to it, and record each decision, including blocked calls.

    Evidence that shows the restriction held

    • Per-call decision record: time, workflow, node, requested destination, resolved provider, product and endpoint, and the decision
    • Policy version in force for each decision, with change history for the approved list
    • Blocked-call records, including the reason and who was notified
    • Redaction and escalation records, including approver and outcome
    • Negative test results showing unapproved providers, base URLs and direct HTTP requests fail
    • Documented retention settings for the records, and the BAA reference behind each approved entry

    Configuration-only controls compared with runtime enforcement

    Restricting which model nodes are available

    Reduces accidental use but does not stop base URL overrides, personal credentials or direct HTTP requests. Depends on author behavior.

    Periodic workflow and credential review

    Finds drift after the fact. A non-covered call can occur and complete between reviews.

    Network egress allowlist alone

    Blocks unapproved hosts across all paths. Cannot distinguish products or endpoints on an approved host, and cannot redact content.

    Enforcement point in the model call path

    Evaluates each request against provider, product and endpoint policy, and produces a decision record. Adds a hop and a component to operate.

    What BAA coverage means in practice

    A Business Associate Agreement is an agreement about specific services. Treating it as a property of a vendor name is the first mistake. A single provider can offer several products, several endpoints and several data-handling settings, and only some combinations may fall under an executed agreement. The approved list should therefore be written at three levels: the provider, the product within that provider, and the endpoint that n8n actually calls. Retention and training-use settings belong in the same record, because they are commitments the provider makes and the FTC has said that model-as-a-service companies that fail to abide by such privacy commitments may be liable.

    A BAA also does not replace your own risk work. The ONC HTI-1 final rule preamble states that under the HIPAA Security Rule, business associates have an independent obligation to identify and manage risks, regardless of whether a BAA exists. Separately, the HIPAA Security Rule amendments published as a proposal in January 2025 state that ePHI, including ePHI in AI training data, prediction models and algorithm data maintained by a regulated entity for covered functions, is protected by the HIPAA Rules. As of November 2026, treat that text as a proposal and not as a requirement in force. Either way, the practical conclusion is the same: the approved list is an input to a risk analysis, not a substitute for one.

    How an AI Agent node can reach a model

    The AI Agent node does not call a model by itself. It uses a chat model sub-node, and that sub-node uses a credential. Each of these is an opportunity for the destination to change without anyone reviewing the workflow as a whole.

    There are four paths to account for. First, the workflow author can add or swap the chat model sub-node for a different provider. Second, the credential can be personal or shared, so a key issued outside your procurement process can be attached to a node. Third, a node can be pointed at a custom base URL, which changes the destination while the node still appears to be a familiar provider. Fourth, a separate HTTP Request node can call a model API directly and bypass the agent's model configuration entirely.

    A node-level restriction addresses only the first path, and only when authors comply. The deployment model matters too. Where you operate the n8n runtime, you can constrain network egress yourself. Where another party operates it, egress controls are not yours to set, so the enforcement point has to be something the workflow must call in order to reach any model. Verify each path in a test environment for your own deployment rather than assuming it is closed.

    A routing and enforcement pattern

    The pattern has four parts, and they work only together.

    First, put a single enforcement point between n8n and every model provider. Configure AI Agent chat model nodes to use it as their base URL, with credentials issued for that purpose. Second, hold the real provider credentials only at the enforcement point. Workflow authors then have no key that works against a provider directly, so personal and shared credentials stop being a bypass. Third, restrict n8n egress to the enforcement point where you control the network, which closes the HTTP Request path as well. NIST guidance on agent environments recommends limiting network access through a strict allowlist where possible and limiting access to what is necessary. Applying that to LLM egress from n8n is a reasoned extension, so test that unapproved destinations fail. Fourth, have the enforcement point resolve each request to a provider, product and endpoint, compare it with the approved list, and apply the current policy version.

    The tradeoffs are real. The enforcement point is in the availability path, so it needs capacity planning and a defined failure mode. A fail-closed design protects PHI but interrupts workflows. The approved list needs an owner and a change process, because adding a provider is a compliance decision as much as a technical one. Where n8n is operated by a third party, you rely on the credential and base URL controls rather than network egress, so the verification effort is higher.

    Handling disallowed calls: allow, redact, block, escalate

    Every request needs a predictable outcome, and the workflow must be able to handle it. Define four.

    Allow applies when the destination is on the approved list and the content is permitted for that destination. Redact applies when the destination is approved but the request contains data that should not leave, so the content is modified before sending and the modification is recorded. Block applies when the destination is not approved. The caller should receive a consistent error so the workflow can branch on it instead of failing in an undefined way. Escalate applies to cases a rule cannot settle, where the request is held for human approval and the approver and decision are recorded.

    Decide in advance which workflows may continue after a block, which must stop, and who is notified. A block that silently triggers a retry against another model node defeats the purpose, so retries should pass through the same enforcement point and meet the same policy.

    Trussed enforces policy in the execution path across models, agents and tool calls with these outcomes, and the same design can be built with other components. What matters is that the decision is enforced at runtime, not reviewed after the fact.

    BAA-scoped routing for n8n AI Agent calls

    Label: Define

    Description: Approved provider, product and endpoint combinations, each tied to an executed BAA.

    Label: Route

    Description: All AI Agent model traffic passes through a single enforcement point.

    Label: Decide

    Description: Each request is allowed, redacted, blocked or escalated at runtime.

    Label: Prove

    Description: Decision records and policy versions show that the restriction held.

    Review your n8n AI Agent model paths

    Walk through the routing, enforcement and evidence design for your n8n deployment with a Trussed expert.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.