Honeycomb Insurance
Meetings: 25 | Period: 2025-07-16 → 2026-04-28
Account Summary
Honeycomb Insurance is a US-based InsurTech focused on landlord insurance. They use OneAI voice agents to qualify inbound leads by phone — screening prospects before routing them to live reps or binding flows. The engagement is deep and long-running (25 weekly syncs), with a dedicated technical stakeholder (rafael) who built a custom supervisory AI layer on top of OneAI’s platform. The relationship has weathered a serious operational incident in late January 2026 (spike in false disqualifications causing churn) and recovered. Current focus is script quality, conversational naturalness, and a second agent use case for producer onboarding.
Use Case
Primary: Inbound Lead Qualification (Landlord Insurance)
- AI voice agent calls inbound leads, runs through a qualification script (~3 core questions).
- 12% of callers need insurance validation before routing.
- Qualified leads are routed to live reps; disqualified are dropped.
- A 4-way triage classification system determines disposition.
Secondary (in progress): Producer / Agency Onboarding
- ~60 new insurance agencies onboarded per month.
- Goal: AI-driven onboarding calls to verify and update agent contact information.
- rafael has provided a use case overview and script; integration is underway.
Relationship Health
🟡 Recovering — Active, engaged, but history of instability
- Strong technical engagement from rafael (built 1,000-line supervisory LLM layer).
- Jon is hands-on with script quality and pushes for fast iteration.
- January 2026 disqualification spike caused real churn and rafael paused lead ingestion — that trust dent is real.
- Weekly cadence maintained throughout, including through the incident. Relationship is stable but needs consistent delivery on script quality and classification accuracy to stay green.
Key Contacts
| Name | Role | Notes | |
|---|---|---|---|
| Jon Repka | jon@honeycombinsurance.com | Primary stakeholder | Drives script tone, pacing, and copy. Sends Word doc drafts. Pushes for tight, natural-sounding scripts. |
| rafael | rafael@honeycombinsurance.com | Technical lead | Built 1,000-line supervisory AI + LLM audit layer. Owns data review, lead ingestion, and edge case flagging. Paused ingestion during janj crisis. |
| janj | jan.j@honeycombinsurance.com | Operations / monitoring | Monitors live calls for false positives. Sends flagged examples to the team. |
What’s Working
- Script iteration process: Jon sends Word doc drafts with tracked changes; Daniella applies on platform. Clear loop.
- Supervisory AI layer: rafael’s custom audit layer (LLM + 1,000-line code) catches misqualified deals before they cause downstream issues. Built on top of OneAI.
- Model retraining pipeline: Amit retrained the AI model in early Feb 2026, correcting 10 prior errors. Materially improved call flow.
- Business results: 12% increase in quote rate and 14% increase in bind rate were confirmed in November 2025 — concrete ROI numbers they’ve seen.
- Collaboration depth: rafael’s supervisory layer and janj’s manual monitoring show strong internal investment. They’re not passive customers.
- Holiday management: Holiday pauses handled proactively (Daniella sent confirmed schedule in Nov 2025).
Open Challenges
- Script naturalness: Robotic tone and slow cadence remain complaints as of March 2026. Jon explicitly flagged this. Script tightening is ongoing.
- Redundant questions: Multiple rounds of deduplication (March 2026 sync: consolidating effective date questions). Script has accumulated cruft over time.
- Caller confusion: Prompts are unclear enough that Daniella and Michael launched an initiative (April 2026) to rewrite intros and remove redundant greeting phrases.
- Direct contact transfers: Planned experiment (April 2026) not yet executed. No confirmed outcome on record.
- Classification accuracy: 4-way triage was at 70% accuracy as of January 2026 — Amit updated the prompt, but no confirmed accuracy improvement on record since.
- January 2026 incident aftermath: The spike in false disqualifications caused real churn. rafael paused lead ingestion. Full confidence may not be restored. Ongoing monitoring by janj is a signal that trust in the system isn’t fully automatic.
- Second agent (producer onboarding): rafael provided use case overview. Integration progress not confirmed as complete.
Open Action Items
From the most recent meetings (as of 2026-04-28):
| Owner | Action | Source Meeting |
|---|---|---|
| Daniella | Update call script intros with clearer questions; remove redundant greeting phrases | 2026-04-28 |
| Jon | Send updated script with highlighted changes in ‘draft two’ Word doc | 2026-03-24 |
| Daniella | Confirm Jon’s script changes and apply platform updates | 2026-03-24 |
| janj | Continue monitoring live calls; send new false positive examples to team | 2026-02-17 |
| Daniella | Follow up with John on future meeting participation | 2026-02-17 |
Recent Meeting Highlights
2026-04-28 — Script Clarity Initiative Daniella and Michael leading effort to reduce caller confusion. Explicit questions being added; redundant greeting phrases being removed. Direct contact transfer experiment planned but not yet run.
2026-03-24 — Voice Naturalness Push Jon flagged robotic tone and slow cadence. Agreed to tighten the script for better flow. Jon to send ‘draft two’ Word doc; Daniella to apply.
2026-03-17 — Question Deduplication Removed redundant effective date questions to reduce caller confusion. Amit to send example calls; rafael to review.
2026-03-10 — Script Optimization to 3 Core Questions Compressed qualification to three main questions. Identified that 12% of callers need insurance validation for routing.
2026-02-17 — False Positive Review 3 false positives confirmed, 1 error fixed. System declared stable. janj monitoring ongoing.
2026-01-27 — Crisis: Disqualification Spike Significant increase in disqualifications disrupted operations and caused churn. rafael paused lead ingestion. Amit committed to full review and benchmark tests. This was the low point of the relationship.
2025-11-25 — Strong Business Results 12% increase in quote rate, 14% increase in bind rate confirmed. Plans to automate producer channel onboarding. rafael to share detailed performance data.
Prep Tips for Next Meeting
- Script update status: Did Jon send the ‘draft two’ Word doc (due after 2026-03-24)? Come prepared with confirmation or a flag if it’s still pending.
- Direct transfer experiment: The April 2026 sync planned a direct contact transfer experiment. Know whether it ran, what the outcome was, and what the next step is.
- Naturalness / tone: Jon has been persistent on this. If you’re applying script changes, bring specific examples of what changed and why it should sound better.
- False positive rate: janj is still monitoring. Ask for the latest numbers. Classification accuracy was 70% in January 2026 — push to get a current benchmark.
- Producer onboarding (second agent): Check integration status. rafael provided the overview months ago. If it’s stalled, name it.
- Bring rafael into any technical discussion: He built the supervisory layer himself. He will catch anything vague or inconsistent. Treat him as a technical peer, not just a stakeholder.
- Reference the business results: The 12%/14% improvement (Nov 2025) is their internal win. Reference it when discussing continued investment or new use cases.