AI agents

Agents that prepare the complaint, and people who decide it

Eight agents for triage, vulnerability, research, remediation, replies, supervision, root causes and QA. They never send, pay, classify, close or approve.

The eight agents

One step each, and a line none of them crosses

Six agents work on a case, in order, when someone presses “Run the agents”. Three work across the book: the triage agent over the inbox, the root-cause analyst and the QA reviewer.

AgentWhereWhat it preparesWhat it never does
Triage agentCase and inboxIntent, product, sentiment and the quoted dissatisfaction; treating a hidden complaint as a complaint; cases for inbox complaints, most urgent firstDecide whether a case is a complaint
Vulnerability spotterCaseEach indicator quoted; how to adapt the case, from the knowledge baseInvent an indicator
Case researcherCaseThe fees in question, related cases, prior complaints, a shared fee run; the root-cause tagsInvent a charge: every fee is the firm’s record
Remediation calculatorCaseThe refund plus interest the calculator worked out, and whether a lead must approve itApprove or pay redress
Reply drafterCaseThe acknowledgement, response or final response letter, with citations, heldSend to a customer
SupervisorCaseWhere the case stands against the controls and the clock; an outcome where the records support oneRecord the outcome or close the case
Root-cause analystOversightClusters by cause and product, and taking ownership of the systemic onesInvent a cause: every cluster counts real complaints
QA reviewerOversightA sample of closed cases checked against the checklist; a pass for the clean onesRecord a QA result
Autonomy

Autonomy is earned, and capped

Each agent’s actions start at L1 and are hard-capped at L2. An administrator can enable or disable an agent and set its level within the cap. The platform recommends a promotion only when people agreed with at least 90% of at least 50 answers in 30 days, and a demotion when agreement falls below 70%.

LevelNameWhat the agent may do
L1DraftPrepares a draft; a person does the work. Every InsightService agent starts here
L2Ask firstPrepares the decision and asks a person before it acts. The ceiling for every InsightService action
L3Act and tellActs, tells the person, and lets them pull it back. Not available to InsightService agents
L4ActActs on its own. Not available to InsightService agents
Grounding

The engines compute. The model explains.

Classification, deadlines, fee matching, interest and the controls are deterministic code. Where a language model is used, it explains and drafts over numbered facts, and an answer that carries a name or number not in those facts is thrown away.

Deterministic

Arithmetic is code

Deadlines, duplicate fees, refunds and interest are calculated, not generated, so the same data gives the same answer.

Grounded

Facts are numbered

Each agent reasons over numbered facts from the case and cites them; anything else is discarded.

Fallback

Model off is a supported mode

With SERVICE_AGENTS_LLM=0 every agent runs on its rules and says “reasoned by rules” on each step.

Prepared decisions

Accept, reject or ask

What an agent prepares lands on the case and on the handler’s home. Accepting carries it out as the person who accepted; rejecting records that a person disagreed; asking sends a question back. Each answer counts toward the agent’s agreement rate.

Copilot

A read-only colleague on every case

The copilot answers questions about the open case from read-only tools and never changes anything.

Watch the agents prepare a complaint

We run them on the demo firm’s cases and walk through every prepared decision.

Request a demo