Visualisation: a sound waveform built from slender graphite rods stands in a dark hall with tall windows, a thin cyan scan line running along it.

Promeing ConceptScenario set in Paris

Autonomous Customer Intelligence System

IndustryInsurance customer operations

Visualisation

Status
Promeing Concept
Scenario
Paris
Industry
Insurance customer operations
Capabilities
AI AgentsAI AutomationAI Infrastructure
Sheet
02 / 06

Voice and text agents that resolve routine insurance requests within written policy limits, cite the policy wording they rely on, and hand everything else to a person with the case already written up.

A Promeing Concept: a studio study of a system we would build, set in a realistic scenario. It describes no client, no deployment and no measured result.

Problem

Routine requests and first notices of loss arrive by phone, chat and email at all hours, and the answers are spread across a large library of policy wordings. Customers wait for questions that have clear answers, and when a case does need a person, they explain it all over again.

Context

The scenario: an insurer’s customer operations team in France, handling motor and home policies across several product generations, each with its own wording. Service staff are skilled and stretched; the knowledge is written down, but finding the right clause for the right policy version takes time. Any automation has to respect strict rules on what can be decided without a person, and every decision has to be explainable to a supervisor and an auditor.

Visualisation: close detail of polished steel rods of different heights forming a waveform, crossed by a cold cyan scan line and reflected in a dark floor.
VisualisationFig. 01

System

Voice and text agents handle the routine end to end: changes of address, policy documents, payment questions and the first notice of a claim. Before answering, an agent retrieves the wording of the customer’s own policy version and cites the clause it relies on. Explicit limits are written as rules, not prompts: anything outside them, such as an injury, a disputed cover or a vulnerable customer, goes to a person with a written summary, so the customer never repeats themselves. A supervisor console shows the decision trace behind every action. It would be deployed privately and evaluated on consented, recorded interactions before every release.

AI architecture

Channels feed one conversation layer. Agents act only through a policy engine that holds the limits; retrieval grounds every answer in the customer’s own policy wording, and every step lands in the trace a supervisor reads.

AI architecture of Autonomous Customer Intelligence System

Layers

  • Interface: Phone, chat and email, Supervisor console
  • Agents: Service agents, Policy limits engine
  • Retrieval: Policy wording retrieval
  • Models: Speech-to-text with diarisation, Private language model
  • Data: Policy and claims records
  • Infrastructure: Evaluation on consented recordings, Decision trace store

Connections

  • Phone, chat and email → Speech-to-text with diarisation
  • Speech-to-text with diarisation → Service agents
  • Private language model → Service agents
  • Policy wording retrieval → Service agents
  • Policy and claims records → Policy wording retrieval
  • Service agents → Policy limits engine
  • Policy limits engine → Supervisor console
  • Policy limits engine → Decision trace store
  • Decision trace store → Supervisor console
  • Evaluation on consented recordings → Private language model
ArchitectureWhat people use at the top, what it stands on at the bottom

Interface

The supervisor console: the live queue, the conversation as it happens, and the decision trace beside it, with the policy wording each step relied on.

Interface mockSchematic, with sample data

  1. A: Every channel in one queue, with the state of each conversation stated plainly.
  2. B: The decision trace: intent, the clause it relied on, the limit it checked and the action it took.
  3. C: A handoff carries the whole case to a person, written up, so the customer never starts again.

Numerals appear only in dates, IDs and timestamps. Indicators read — until there is something true to show.

Technology

  • Speech-to-text with diarisation, and natural text-to-speech
  • Telephony and chat integration through existing contact-centre platforms
  • Retrieval over versioned policy wordings, with clause-level citations
  • Policy limits as explicit, testable rules
  • Private language model deployment in a chosen EU region
  • Evaluation harness on consented, recorded interactions
  • Decision trace store with role-based access for supervisors and audit

Expected outcome

Expected

Faster answers to routine requests at any hour, handoffs that arrive written up, and human time spent on the cases that need judgement.

Qualitative and unmeasured. A pilot measures it against acceptance criteria agreed at the start.

What we would measure

  1. Resolution without handoff
  2. Time to first answer
  3. Handoff quality, rated by the people who receive them
  4. Customer effort
  5. Citation accuracy against the policy wording

Metric names only. Values exist once a pilot measures them.

What a pilot would need

  1. One request type with clear, written rules to start, such as changes of address
  2. The policy wordings for the products in scope, with their version history
  3. Consented, recorded interactions to build the evaluation set
  4. Read access to policy and claims records; write access only where the rules allow
  5. A supervisor team to set the limits and review traces daily
  6. Sign-off from compliance on what an agent may decide on its own

Promeing ConceptScenario set in Paris

Building something like this?

Bring the problem in its messiest form. We’ll tell you honestly whether AI is the right way to solve it.

Discuss a system like this All work