The autonomy dial and prepared decisions

Every Insight workbench has the same autonomy dial: four levels, per agent and per action, with a tenant ceiling above them. In claims or underwriting a carrier may, with evidence, let some actions run at “act and tell”. In an acquisition there is nothing an agent should do on its own. InsightRollup ships with every action capped at L2 — ask first, whatever the dial says.

The four levels

Draft

Prepares a draft; a person does the work.

CeilingL2

Ask first

Prepares the decision and asks a person before it acts.

Act and tell

Locked for every acquisition action.

Act

Locked for every acquisition action.

Three caps, the lowest wins

flowchart LR D["Dial: the agent's level
for this action (and state)"] --> MIN{"Lowest of"} T["Tenant ceiling
MERGE_AGENTS_MAX_LEVEL (L2)"] --> MIN H["Hard cap per action
(L2 for every acquisition action)"] --> MIN MIN --> L["Effective level"] L --> P["Prepared decision
waits for a person"]

The effective level for an action is the lowest of three: the level set on the dial for that agent and action (optionally per state), the tenant ceiling from configuration, and the hard cap in code. The seven actions an acquisition agent can propose — ask the seller, propose retention agreements, confirm client matches, request appointments, adopt the plan, finalize the memo, the practice leader's sign-off — are all hard-capped at L2. What goes to a seller, a carrier or an investment committee is a person's call.

The autonomy page — Autonomy: the tenant ceiling at L2 ask first, agreement over answered decisions, and each agent's level per action, with L3 and L4 locked.

Prepared decisions

Everything an agent wants done arrives as a prepared decision: a title, the lines of what it proposes, the payload, why, the facts behind it, a confidence and whether it was reasoned by the model or stands on the engines. The person answers Accept, Reject or Ask. The answer is signed on the decision record as that person and counted as evidence.

stateDiagram-v2 [*] --> Open: agent prepares Open --> Superseded: a newer one for the same action Open --> Accepted: person accepts (carried out as them) Open --> Rejected: person rejects, with a note Open --> Open: person asks; the agent answers from the facts Accepted --> [*] Rejected --> [*] Superseded --> [*]

Agreement as evidence

The autonomy page shows agreement — the share of answered decisions people accepted — and overrides. In products where promotion is allowed, an agent moves up when people agree with at least 90% of 50 answers and down at once below 70% of 20. In InsightRollup the evidence is still recorded, so a practice leader can see which agents are trusted, but the hard cap holds.