Agent Architecture

BindPilot's agents are autonomous AI systems that handle repetitive insurance operations without human intervention. Think of them as autopilots for your agency.

Autopilot vs. Copilot

Copilot: AI assists you (you make the final decision)

• Example: Form auto-fill, email draft suggestions

Autopilot: AI acts on your behalf (you review the result)

• Example: Auto-shop renewals, auto-respond to COI requests

BindPilot is autopilot-first. Our agents run tasks end-to-end, then notify you for review/approval.

Available Agents

BindPilot ships 12 agents wired into the API, in two families.

Harness agents — invoked directly at /api/agents/{agent}:

| Agent | Role | Status | |-------|------|--------| | Onboarding Agent | Client intake, document processing, risk profile construction | Built | | Self-Service Agent | Client-facing policy questions, certificate requests | Built | | Claims Agent | First-notice-of-loss intake and claims triage | Built | | Compliance Agent | Licensing, surplus lines, and filing-obligation checks | Built | | Knowledge Base Agent | Retrieval over carrier documentation and guidelines | Built | | Security & PII Agent | PII detection and redaction, data classification | Built |

Orchestrator agents — registered at /agents and run on demand:

| Agent | Role | Status | |-------|------|--------| | Quoting Agent | Appetite filtering, parallel carrier submission, ranking | Built | | Renewal Agent | Renewal pipeline, re-shop, proposal preparation | Built | | Servicing Agent | COI requests, endorsement changes, account maintenance | Built | | Analytics Agent | Book analytics, retention, carrier performance | Built | | Notification Agent | Outbound notifications via the transactional outbox | Built | | Proposal Agent | Client-facing proposals from quote comparisons | Built |

Status is "Built", not "Live in production." BindPilot is pre-launch. These agents are implemented and reachable through the API, but none has yet executed against a production customer book — there are no agencies, policies or quote requests in production today. Each orchestrator agent carries a cron expression, but they are triggered through the API rather than by a scheduler.

Retention and New Business agents appear elsewhere in these docs. They are not implemented — there is no code for either one. They are on the roadmap, undated.

How Agents Work

1. Trigger

An agent is triggered by a condition or event:

• Renewal Agent: Expiration date reaches X days (e.g., 60)
• Client Service Agent: Incoming email with keyword "COI" or "certificate of insurance"
• Quoting Agent: Manual trigger (runs when you click "Shop All Carriers")

2. Action

The agent executes a workflow:

• Query carrier systems
• Evaluate options
• Generate documents (proposals, responses)
• Send communications (emails, notifications)

3. Result & Review

The agent notifies you of the outcome:

• Notify Only: You review the agent's work, approve or override
• Auto-Send: Agent sends the output (proposal, response); you get a notification
• Manual Approval: Agent flags for your approval before sending

4. Learning

As you override or approve agent actions, BindPilot records the signal so the agents can adapt to:

• What you typically approve
• Which carriers you prefer for certain risks
• Your tone/templates for client communications

The feedback loop that consumes these signals runs hourly. It is built, but it has processed no production signals yet — accuracy improvement is the design intent, not a measured result.

Configuring Agents

Each agent has tunable settings:

1. Go to Settings > Agents > [Agent Name]
2. Adjust:
   • Enable/Disable — Turn agent on/off
   • Trigger Timing — When to run (e.g., "60 days before renewal")
   • Action Level — Notify only, draft for approval, or auto-send
   • Thresholds — Criteria for agent decisions (e.g., "only select quotes under $X")
   • Templates — Email templates, proposal designs agent uses

See Customize Agents for detailed configuration.

Agent Safety & Overrides

Safety features:

• Approval Gates: Critical actions (sending proposals, rebinding) require your approval
• Audit Trail: Every agent action is logged with reasoning
• Override: You can always override agent decisions
• Disable: Pause any agent anytime

If something goes wrong:

• Agent actions are logged and reversible where the underlying action allows it — a draft can be discarded, a proposal cancelled before send, a scheduled task stopped. Anything already delivered to a third party cannot be recalled, which is why send-level actions default to requiring your approval
• Team members can intervene in real-time
• Manual process takes over

Agent Performance & Monitoring

BindPilot tracks agent performance:

1. Go to Analytics > Agent Health
2. View:
   • Uptime — Is the agent running?
   • Success Rate — % of tasks completed successfully
   • Approval Rate — % of agent suggestions you approve
   • Avg Task Time — How long agent takes per task
   • Errors — Any failures or issues

Use this to fine-tune agent settings.

---

Next: Learn about each agent — Quoting Agent, Renewal Agent, Client Service Agent, and more.