← Mir Quadri · AI Lab
● Live buildAgentic customer resolution

// Built from the customer problem backward

Utility Bill Clarity Agent

A customer-resolution prototype that explains unexpected utility bill increases, separates confirmed evidence from uncertainty, recommends practical next steps, and keeps consequential actions behind explicit human confirmation.

Launch live demo

What it proves

3
synthetic billing scenarios
ADC
server-side authentication
Human
confirmation before case creation

Architecture

  • Google Agent Studio
  • Vertex AI
  • Agent Runtime
  • Cloud Run
  • Synthetic data

Independent demonstration · Synthetic data · Not affiliated with any utility company

// Production architecture

One resolution agent. Clear boundaries.

This MVP deliberately uses a single customer-facing agent, not a multi-agent swarm. Its job is bounded: explain a synthetic bill change and distinguish facts from assumptions. Chat can recommend a review, but it cannot create a case.

Agents
1

Resolution agent

HTTP endpoints
2

Health + agent API

Demo scenarios
3

Synthetic records

Customer PII
0

Real accounts used

  1. 01

    Customer browser

    Chooses a synthetic scenario and asks a bill question.

  2. 02

    Cloud Run application

    Serves the demo UI and validates POST /api/agent requests.

  3. 03

    Vertex Agent Runtime

    Runs one Utility Bill Clarity resolution agent with ADC.

  4. 04

    Grounded response

    Returns evidence and opens a dedicated review panel when a case is requested.

Trust boundary

The browser never receives Google Cloud credentials. Cloud Run uses its service account through Application Default Credentials to call the Agent Runtime server-side.

Human approval gate

The dedicated review panel shows the selected customer, evidence, case category, and demo-only destination. Only its final Confirm action sends a structured approval request. The server validates that request, records the approval method and timestamp, then generates a simulated case ID. Cancel—or ordinary chat—creates nothing.

API map

  • GET/api/healthService health and runtime location check
  • POST/api/agentValidated chat or structured case-confirmation request
  • SDKAgent Runtime streamQueryServer-side call to the deployed reasoning engine

// How It Works

Architecture

This demo uses one customer-facing resolution agent, not a multi-agent swarm. Its job is bounded: explain a synthetic bill change, separate confirmed evidence from assumptions, and keep case creation behind an explicit human confirmation. Chat can recommend a review. It cannot create a case.

01

Customer browser

Chooses a synthetic scenario and asks a bill question. The browser never receives Google Cloud credentials.

02

Cloud Run application

Serves the demo UI and validates POST /api/agent requests. Cloud Run calls the Agent Runtime server-side with ADC.

03

One resolution agent

Vertex Agent Runtime runs a single Utility Bill Clarity agent. It returns grounded evidence and opens a dedicated review panel when a case is requested.

04

Human confirmation

Only the dedicated review panel's Confirm action sends a structured approval request. Cancel — or ordinary chat — creates nothing.

// Why This Pattern Works

Key Insights

  • ✓

    One agent, one job

    A bounded resolution agent is easier to test than a swarm. It explains the bill. It does not create a case in chat.

  • ✓

    Grounded answers

    The agent separates confirmed evidence from assumptions so the customer can see what is known versus inferred.

  • ✓

    Human approval gate

    Consequential actions stay behind explicit confirmation in a dedicated review panel. Ordinary chat creates nothing.

  • ✓

    No customer PII

    The demo uses synthetic billing scenarios. Real accounts are not used.

// Adapt This Pattern

Use This Anywhere

Bound the agent, ground the answer, and keep consequential actions behind a human. The same shape works anywhere you need to explain a record and escalate with approval:

  • Insurance claims: extract coverage details, verify eligibility, calculate payout
  • Contract analysis: parse terms, flag risks, summarize obligations
  • Invoice processing: extract line items, validate against POs, route for approval
  • Medical records: extract diagnoses, medications, lab values, build timeline

The key insight: bound the agent, ground the answer, and keep consequential actions behind a human.

🎯

Build Your Own System

Need multi-agent orchestration for your problem? I'll architect the system and integrate your data.

Discuss Your Use Case
🔓

Early Access to Playbooks

I'm building production playbooks for agent orchestration. Sponsors get first access before public release.

Become a Sponsor
💡

Fund Specific Research

Shape research priorities. Your sponsorship funds specific agent patterns and production strategies.

Sponsor Research
Back to all demos