A state insurance examiner asks a carrier how a specific claim was adjudicated. The answer that used to be acceptable was "an adjuster reviewed it against the policy". The answer that used to be acceptable when the adjuster's judgement was assisted by AI was "the adjuster reviewed it against the policy, with some help from a tool". The answer the examiner accepts now is the seven-step execution trace of the agent that assisted the decision, showing what was retrieved, what was reasoned over, what policy the recommendation was validated against, and what the human accepted.
If your architecture cannot produce that trace, your architecture cannot pass a state insurance audit. If it can, the AI-assisted claim is easier to defend than the manually-adjudicated one, because the reasoning is explicit.
The trace is not a log file. It is a structured record of the agent's execution, in seven steps that every defensible agent architecture produces the same way.
The seven steps
01 · Input. What triggered the agent's execution, on what record, at what time, by what user. This is the primary key of the trace. Everything else in the record ties back to this event.
02 · Retrieval. The specific documents, records, and prior cases the agent read to assemble the context. Not "the policy library" but the six specific documents pulled, their versions, their access-controlled paths, and the retrieval scores that surfaced them.
03 · Reasoning. The intermediate steps the agent took. Not the raw model output but the structured sequence of sub-decisions: identified the coverage in question, extracted the relevant policy clause, matched it against the loss report, computed the applicable limit, checked the fraud signal, drafted the coverage analysis.
04 · Tool selection. Every external call the agent made. The specific API, the parameters passed, the response received, whether the call succeeded. If the agent looked up a claim history in a case management system, that lookup is in the trace with the request and the response.
05 · Policy validation. The runtime check against the firm's policy layer. Which policy version was applied. Which rules the recommendation was tested against. Which passed. Whether any raised a policy exception that had to be escalated.
06 · Human decision. The recommendation that was surfaced to the human, the human's response (accept, modify, override), and the reason the human recorded for their decision. If the human did not have to intervene because the recommendation cleared all gates, that too is recorded.
07 · Output. The final action taken. The record that was updated, the message that was sent, the decision that was written back to the workflow tool.
Seven steps. Every agent execution. Structured records that a compliance office can query directly, without a data scientist writing a bespoke report.
What "defensible" means at this level
The state examiner does not want a screenshot of the AI's chat interface. She does not want a text description of what the agent did. She wants to run a query against the audit store and see: for claim number 2026-CL-0042184, at 14:37 on August 4, the agent retrieved these six documents, reasoned through these four sub-decisions, passed policy validation version 7.3, surfaced this recommendation to Adjuster Chen, who accepted it with the note "consistent with prior treatment of similar losses at this policy limit".
That query is a JOIN across the execution store, the policy version registry, the user access log, and the workflow tool's own case record. It runs in seconds. It produces a defensible narrative. The examiner reads it, checks two or three other claims against the same pattern, and moves on.
If the AI is architected in the way most current tools are, that query cannot run. The retrieval was implicit inside a prompt template that got composed at request time and thrown away. The reasoning was inside the model's chain of thought and never surfaced. The policy validation happened on the output text via a regex. The human decision was captured as a checkbox with no rationale. There is no way to reconstruct the seven-step story, so there is no way to defend the decision.
Where the trace comes from
Intelligence Fabric is the orchestration layer that produces the seven-step trace as a first-class output. It does this not because tracing is a feature but because the architecture itself is built around structured execution. Every agent runs as a sequence of typed sub-operations, each of which writes its own record to the trace store. Retrieval writes a retrieval record with the documents and scores. Tool calls write a tool-call record with the request and response. Policy validation writes a validation record with the rules that fired.
The trace is a side effect of the architecture, not an instrumentation layer bolted on top. That is why it is complete. If your agent architecture produces the trace by piping the model's output through a logging library, the log will be incomplete because the model itself does not know what to log. If your agent architecture routes every sub-operation through a typed contract, the log is complete by construction.
Why unified retrieval matters
Step 02 in the trace, retrieval, is the step where most agent architectures leak defensibility.
An agent that assembles context by reading whatever documents happen to be in the current session cannot answer "why did you retrieve document X and not document Y". The retrieval was implicit and per-request. It happened inside the model's context window with no accountable component.
Fabric routes every retrieval through a single scoped retrieval layer. The layer knows what the user is allowed to see. It knows what document classification the request is happening at. It knows what the retention policy for the retrieved content is. It logs the query, the returned documents, the retrieval scores, the ranking model version, and the access-control decisions applied.
When the examiner asks "why did you retrieve the 2019 policy interpretation memo and not the 2021 update", the answer is "the 2019 memo scored 0.87 and the 2021 memo scored 0.71 against this query, both retrieved, both cited, the 2019 was the higher-scored match". That is a defensible answer. "The model chose it" is not.
What ships in a Fabric deployment
A structured execution trace store. Every agent invocation writes a full seven-step trace. The store is queryable by the compliance office directly, with SQL or the compliance office's existing BI tool.
Unified retrieval semantics. Every agent reads through the same scoped retrieval layer, which means access control, retention, and classification are enforced consistently and logged consistently.
Consistent observability and policy boundaries. AI Gateway sits underneath for model routing and cost, AI Guardrails sits alongside for policy enforcement, and Fabric is the coordination layer that makes both of them consistent across agents.
Deploy targets appropriate to the classification of the workload. SaaS on shared infrastructure for low-classification work. Dedicated cloud tenants for medium. Sovereign VPC deployment inside the customer's own accredited boundary for federal or classified work. The trace architecture is the same across all three; the deployment posture changes with the data.
The state insurance examiner walked into the boardroom with a question about a specific claim. She walked out with the seven-step trace and moved on to the next carrier. That is what Intelligence Fabric is for.
Tell us the highest-stakes agent workload your compliance office has questions about. We will run an architecture review on the specific traces you would need to produce, and show you what the Fabric trace looks like on your workload before you commit to anything.