A broker that puts AI between its advocates and its clients’ claims will be asked hard questions by its E&O carrier, its compliance team and its clients. These are the answers InsightAdvocate is built to give.
| The question | Where the answer lives |
|---|---|
| Can a model invent a deadline, limit or floor? | No. Engines compute every figure; an answer with an ungrounded number is discarded. See grounded figures. |
| Can an agent send something on its own? | No. Every action is capped at L2, ask first; sends are dry runs today. |
| Who decided, and on what? | The decision record: control hits with rule versions, prepared decisions, answers, sign-offs. |
| What did the model see and say? | A trace per agent step: model, tokens, cost, the facts and the output. Prompts are versioned, with local defaults. |
| Whose data can a user see? | What their role and scope allow; the copilot’s lookups use the same scope as the screens. Every row belongs to an organization. |
| Personal data in prompts? | Agents reason over the engines’ numbered facts, not whole documents, so a prompt carries what the step needs and little else. The model runs in your own Bedrock account. |
| Are the regulatory timelines law? | No. They are illustrative models, labelled as such on every screen. A practice sets its own with counsel. |
| Do tests call a model? | No. Bedrock is refused at the AWS client in the test suite; engines are tested on known answers. |
flowchart TB
subgraph Controls
G["Grounding check"]
L["L2 cap and tenant ceiling"]
R["Rules Studio, versioned"]
end
subgraph Evidence
T["Traces per step"]
D[("Decision record")]
A["Audit trail"]
end
G --> D
L --> D
R --> D
T --> A
D --> A
Every Analyze shows its own summary: which agents ran, whether each reasoned with the model or fell back to the engines, and what each prepared. A reviewer can start there and follow any line to its facts.
