November 2025 - Enrichment Cascade, Interest Signals, and Conversation Depth¶
Context¶
With basic visitor profiling in place, November built the buyer-intelligence layer: a multi-source enrichment cascade, LLM detection of buying signals, structured dialog-state extraction, and a conversation-depth experiment. These are the month's retained R&D, with the detector-inspection surface as supporting infrastructure; only the client-facing dashboard product and browser-compatibility maintenance are excluded, listed concisely at the end. November is also the first month with enough production data to measure the foundation's effect, so its business-impact numbers are genuine production measurements.
Maps to 2026 R&D projects: Compounding Context Engine (enrichment, interest signals, dialog state); Adaptive Multi-Tenant Conversation Orchestration (signal-driven demo arbitration); Closed-Loop Agent Evaluation and Optimization (measured Q4 impact).
Multi-Source Enrichment Cascade¶
Project and lock¶
Single-source IP enrichment identified only a minority of visitors and added latency on every miss. The lock is whether several enrichment sources of differing confidence can be combined into one cascade that raises coverage and completeness while keeping latency bounded, without polluting the profile cache with low-confidence (e.g. shared-IP) data.
This month's work¶
The core of the work was the cascade logic and its confidence-aware caching policy: sources are tried in order of confidence and cost (cached profile, known-account lookup, client-side reveal, session-based identification, server-side IP enrichment), and whether a result is written back to the IP→account cache depends on the source's confidence — high-confidence reveals are cached, IP-only results are not, because a shared IP would poison the mapping. Integrating each individual enrichment vendor's API was ordinary integration.
Results, proof, and next step¶
The cascade replaced the fragile single-source path: it sustains the higher identification level reached in October at materially lower average latency through caching, and raises data completeness. The research core was the cascade ordering, the confidence-based store/skip decision, and the caching policy; the per-vendor API integrations were ordinary integration.
Interest-Signal Detection and Demo Arbitration¶
Project and lock¶
Demo proposals were mistimed — too early felt pushy, too late missed the visitor, too often annoyed. The lock is whether buying intent can be detected from conversation content reliably enough, and scored, to time a demo proposal automatically and per-client.
This month's work¶
An LLM-based detector reads the conversation for signal types (pricing inquiry, demo request, technical deep-dive, competitor mention), each with a confidence and a configurable per-industry weight, and a weighted-threshold rule decides whether to propose a demo. The per-client configurability is what lets one detector serve different industries without code changes — a different signal means a different thing for an accounting tool than for an HR tool.
Results, proof, and next step¶
Detection was evaluated against 200 manually labeled conversations, reaching above 85% precision/recall on the high-value signals (demo request, pricing inquiry) and lower on the fuzzier ones (technical deep-dive, competitor mention). Signal-based timing replaced manual/random proposal timing; the conversion reference of 2.89% (the single Q4 baseline) rose to 3.15% (+9%) into December as the detector matured. The conversation-and-signal analysis views built this month are the surface on which the detector's output is inspected against real conversations and its thresholds tuned; the client-facing dashboard product, authentication, and access control around them are the only ordinary parts. The research was the detection and the configurable scoring/arbitration, measured against the 200-conversation labeled evaluation set.
Dialog-State Extraction¶
Project and lock¶
Beyond intent in the moment, conversations carry durable markers — a demo was requested, an objection was raised, a next step agreed. The uncertainty is whether these can be extracted as structured, reliable state from free conversation so downstream workflows can trigger on them.
This month's work¶
A structured-output extraction step pulls markers (demo requested/booked, email captured, objections, next steps, conversation stage) from the conversation and persists them with the conversation, enabling state-based queries and downstream CRM triggers. The webhook/CRM integration itself was ordinary integration; the research was the reliable structured extraction of dialog state.
Results, proof, and next step¶
Extraction accuracy was high on the crisp markers (email captured, demo requested) and lower on the interpretive ones (objections, next steps), measured against test conversations with known markers. The research was the structured dialog-state extraction; the CRM webhook plumbing was ordinary integration.
Conversation-Depth Experiment (suggested answers and follow-ups)¶
Project and lock¶
Conversations often ended after one or two turns because visitors did not know what to ask next. The hypothesis: offering click-to-send suggested answers and proactive follow-up questions, generated from conversation context, increases conversation depth.
This month's work¶
LLM-generated suggested answers and follow-up questions were produced from the conversation history and visitor context, sent with a single click to remove friction, and A/B compared against the no-suggestion experience.
Results, proof, and next step¶
Engaged conversations (two or more turns) rose from 29.24% to 39.27% (+34%) over the period — a measured production result attributable to this change. A second, separate presentation-style experiment was also launched in November — a controlled comparison of two answer-layout styles for their effect on engagement — running on live traffic from early November; its result matured the following month and is reported in December's prior-month results review.
Measured Business Impact (Q4 production)¶
These are genuine production measurements, the first the platform could make, and they validate the retained work above:
| Metric | October 2025 | November 2025 | Change |
|---|---|---|---|
| Initial engagement | 1.24% | 1.73% | +40% |
| Engaged conversations (2+ turns) | 29.24% | 39.27% | +34% |
| Visitor identification | 35% | 72% | +106% |
Attribution: the engagement gain to page-aware questioning (October) plus richer enrichment; the conversation-depth gain to suggested answers and follow-ups; the identification gain to the enrichment cascade fed by inference-and-merge.
Not retained as R&D¶
The client-facing dashboard product (auth, multi-tenant access control, UI) and the Safari connection/compatibility fix — product delivery and maintenance with no nexus to a named experiment.
Next Work (December)¶
- Intent router for multi-agent orchestration (the routing investigation begins).
- Per-client agent configuration.
- Parallel graph execution for latency.