Logging Copilot Studio Generative Orchestration Decisions for FINRA Rule 4511
FINRA Rule 4511 does not name orchestration decisions, but it requires members to make and preserve books and records, and FINRA treats its rules as technology neutral. Whether a given agent interaction or tool call is a required record depends on its content and what it relates to, which is a firm interpretation. Capture the user request, selected tool, parameters, returned data, policy decision, human approval and timestamp as evidence at the point of enforcement, and store it under retention and format controls your counsel has approved.
- {"title":"1. Request interpretation","description":"The agent reads the user request and decides what it is asking for."}
- {"title":"2. Selection of topic, tool or connector","description":"The agent picks which capability to invoke, which determines what systems are touched."}
- {"title":"3. Parameter construction","description":"The agent fills in the inputs it sends to the tool, which may include client or account data."}
- {"title":"4. Knowledge source retrieval","description":"The agent pulls content from a knowledge source to ground its answer."}
- {"title":"5. Returned data and response","description":"The tool output and the final reply the user receives."}
- {"title":"6. Approval or escalation","description":"A human reviews and approves, or rejects, a proposed action."}
- Completeness: does it capture every field in the schema for every tool call?
- Enforcement linkage: is the policy outcome recorded with the action, not inferred later?
- Durability: can records be retained for the required period in a format your counsel accepts under SEA Rule 17a-4?
- Redaction: is sensitive data handled before it is stored or returned?
- Residency and control: where does the evidence live, and who controls the store?
- Retrieval: can you produce a specific interaction for an examiner quickly?
- Inventory agent interactions and classify which relate to the firm's business as such, using the content-based test.
- Map each in-scope record type to its governing rule and retention period, applying the six-year default only where none is specified.
- Adopt an evidence schema and confirm with counsel which fields each record type requires.
- Compare native transcripts and telemetry against the schema and document the gaps.
- Decide whether capture and enforcement sit in the execution path, and assess this before deployment.
- Update written supervisory procedures to cover agent activity, approvals and evidence review.
What Rule 4511 requires, and what it leaves to your interpretation
FINRA Rule 4511(a) requires members to make and preserve books and records as required under FINRA rules, the Exchange Act and the applicable Exchange Act rules. Rule 4511(b) sets a six-year minimum for FINRA books and records where no period is specified elsewhere. Rule 4511(c) requires records to be preserved in a format and media that comply with SEA Rule 17a-4. As of November 2026, the rule is published in the FINRA rulebook with no reported amendment.
The six-year period is a default, not a universal answer. Where a FINRA rule or an SEA rule specifies a period for a record type, that period applies. SEA Rules 17a-3 and 17a-4 set the minimum requirements for which records broker-dealers must make, how long they keep them and in what formats.
Two FINRA positions shape the analysis. First, Regulatory Notice 24-09 states that FINRA rules are technology neutral and apply to generative AI, whether a firm builds the tool or uses third-party technology, including embedded features in existing products. Second, FINRA guidance on communications holds that content, not the device or technology, determines whether something must be retained. Together these mean a Copilot Studio agent is in scope to the extent it is used in the firm's business, and that each record's status turns on what it contains.
Neither position says that an orchestration decision or tool call is itself a required record. That conclusion is a legal and compliance interpretation each firm must make and document. The rest of this guide assumes you are building evidence so that interpretation can be applied consistently and defended to an examiner.
Define the evidence schema for each decision point
Evidence differs from a raw log in that it is structured around a question an examiner would ask and can be produced on request without reconstruction. For each tool call, the schema should answer: who asked, what was decided, what was sent, what came back, what policy applied and who approved.
The fields below are a working schema, not a regulatory prescription. Treat them as a starting point and have counsel confirm which are required for each in-scope record type.
Retention should follow the record type rather than the platform. Classify each captured interaction by content, map it to the rule that governs it, and apply the specified period. Use the six-year default only where no period is specified. Approvals deserve attention: for communications, SEA Rule 17a-4(b)(4) covers copies and any approvals relating to the firm's business as such, so a human approval step in an agent flow plausibly needs its approval captured. Counsel should confirm that inference.
- User request: the verbatim input, user identity and channel.
- Selected tool and alternatives: which tool, connector or knowledge source was invoked.
- Parameters: the exact inputs sent, with sensitive values redacted per policy.
- Returned data: the tool output, or a reference to it, and the response given to the user.
- Policy decision: allow, redact, block or escalate, with the policy that produced it.
- Approval and timestamp: approver identity, outcome and time for each step.
What Trussed enforces and records, and what stays with the firm
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 produces audit-ready evidence. Policy is enforced at runtime, not reviewed after the fact. Deployment is self-managed in the customer cloud, with configurable retention and data residency, so the firm decides where evidence is held and for how long.
The boundaries are specific. Trussed does not provide end-to-end tracing of agentic workflows or transaction-level governance across tasks and agents. It operates per interaction and per tool call. It has not been shown to serve broker-dealers, and its production record is strongest in insurance, another regulated financial sector, alongside pharma, healthcare and education. Its evidence is currently mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST.
Several decisions remain the firm's. Counsel must decide which interactions are required records, which retention period applies to each, and whether the chosen storage meets SEA Rule 17a-4 format requirements. Written supervisory procedures under Rule 3110(b)(1) should also be updated to cover agent activity and evidence review. A control cannot make those determinations, but it can make the resulting policy consistent and provable.
From user request to examiner-ready evidence
Request
The user asks the agent to do something.
Orchestration decision
The agent selects a topic, tool, connector or knowledge source at runtime.
Policy outcome
Allow, redact, block or escalate for human approval, enforced before the call proceeds.
Evidence record
A durable, retention-controlled record tied to that single interaction or tool call.
Review your agent evidence against the schema
Walk through how policy enforcement in the execution path produces per-interaction and per-tool-call evidence for a regulated enterprise.
Explore Agent and Tool-Call EnforcementHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.