Relationship Intelligence

The Relationship Intelligence panel on a contact or deal detail page is the rep-facing surface of KarmaFlow's Contact Intelligence Pipeline. Instead of showing semantically-similar snippets, it produces a current read on the relationship: temperature, intent, what matters, what to avoid, and the best next move.

This article covers what you see on the card, how the system builds it, what to do when something looks wrong, and the privacy and data-use principles behind the feature.

What the Card Shows

The card is split into three layers that you can scan from top to bottom.

Headline layer (always visible)

• Headline — a one-line summary of the relationship state.
• Current read — two or three sentences interpreting what's going on right now.
• Temperature — one of hot / warm / neutral / cool / cold / risky / unknown, color-coded.
• Intent level — high / medium_high / medium / low / none / unknown.
• Confidence — high / medium / low. Tells you how much weight to put on the read.
• Recommended next move — a specific, plain-English next action.
• Suggested angle — how to frame the conversation when you make that move.

Signals layer (progressive disclosure)

Click any heading to expand:

• Direct signals — things this contact has actually said or done that matter, drawn from their own calls / chats / emails / SMS / notes.
• Pain points & needs — explicit or strongly-inferred business needs.
• Objections & friction — concerns, hesitations, blockers, complaints.
• Risks & cautions — data-quality issues, identity uncertainty, compliance flags, low-trust signals, wrong-number indicators.
• Avoid — what not to do or say next.
• Broader account signals — company-level context, clearly labeled as account context, not contact facts.
• Similar deal patterns — patterns from past WON / LOST deals that look like this one, framed as guidance, not facts about this contact.
• Worth remembering — durable preferences and facts the system thinks should carry into future interactions.

Provenance layer (bottom)

• Evidence note — a short statement about the strength of the evidence behind this read (gaps, staleness, contradictions, coverage).
• Generated timestamp — when the snapshot was last produced, and whether it's a cached or fresh read.

How the Card Stays Fresh

Snapshots are persisted per (contact, deal) pair and follow a stale-while-revalidate policy:

• Less than 15 minutes old? Served immediately.
• Older than 15 minutes? Served immediately and a fresh one is computed in the background — the next page load picks it up.
• Cold start (never generated)? Computed on first read; you'll see a brief "Reading the relationship…" indicator.

A Refresh icon in the panel header forces a fresh regeneration on demand, bypassing the cache. Use it when you've just added a note or completed a meaningful interaction and want the card to reflect it immediately.

Background refreshes also fire automatically when meaningful new evidence arrives — call summary saved, note added, chat session summarized, inbound or outbound email, inbound or outbound SMS, task completed, deal stage changed. You don't need to do anything to trigger them.

What Goes In: Evidence Sources

The system mixes deterministic and semantic retrieval:

• Deterministic pulls (always run, no LLM): contact and company profile, current deal context, open tasks, recent notes and activities, durable contact memory.
• Pre-extracted signals (built up over time): a structured store of pain points, objections, next steps, intents, and risks extracted from each conversation as it happens.
• Semantic search fallback: when signal coverage is thin, KarmaFlow runs a vector query over the long tail of activity to make sure nothing important is missed.
• Similar deals: closed-won and closed-lost neighbors are pulled in as pattern guidance, never as facts about this contact.

Each evidence item is tagged with a scope — contact, company, deal, or similar_deal — so the model knows what to treat as direct fact versus broader context.

What You Can Do With It

• Brief yourself before a call. Open the contact, read the headline + current read + recommended next move, drill into the relevant signals.
• Hand off cleanly. When you reassign a deal, the new owner gets the same up-to-date intelligence the previous owner had.
• Quickly spot risky records. The risky temperature plus the Risks & cautions section flag wrong-number records, unsubscribe intents, compliance concerns, and identity uncertainty before you reach out.
• Trust your AI agents. When a TaskAgent runs against a contact, it reads from the same intelligence snapshot your card shows — so your reps and your AI agents work from one source of truth.

When the Card Looks Wrong

The model will occasionally surface something stale, generic, or off. Here's how to triage:

1. Did you just add evidence? Click Refresh — the in-memory cache might still be serving an older read.
2. Is the confidence low? Look at the evidence note — it usually explains the gap (e.g. "thin direct activity in last 30d, drawing mostly from company context"). Add a note or log a call summary; the next refresh will lean on the new evidence.
3. Is the temperature stuck at unknown? This means there's almost no usable evidence in the CRM yet. The card will populate as activity accumulates. There's no harm in clicking Refresh after adding a meaningful note — you'll get a more substantive read.
4. Is the read confidently wrong? Use the Refresh button first; if it stays wrong, flag it to your admin. Admins can inspect the full evidence bundle behind any snapshot via the Intelligence — Admin page.

The system is designed to fail safe: when evidence is thin or the model can't be reached, the card returns a placeholder ("Not enough evidence to form a read yet") rather than guessing.

Privacy and Data Use

• Tenant isolation. Every read is grounded in your own tenant's CRM. The system never reads from or writes to other tenants.
• No external enrichment. Relationship Intelligence does not call public web databases, social profiles, or third-party enrichment APIs. The only external touchpoint is the reasoning model that produces the judgment.
• Cascade on delete. When you delete a contact, deal, or company, KarmaFlow eagerly purges the associated signals and snapshots — including for soft-deletes. If a contact is later restored, the card regenerates from current evidence on next read.
• Risk flags surface, never hide. Wrong-number records, unsubscribe intents, hostility signals, identity uncertainty, and compliance concerns appear explicitly in the Risks & cautions section so they cannot be missed.
• Conversation summaries, not raw transcripts. When a call or chat ends, KarmaFlow extracts durable structured signals (pain point, objection, next step, etc.) from the summary of the conversation, not the raw transcript. The extracted signals are concise paraphrases or short evidence excerpts, designed to capture the actionable read without retaining lengthy verbatim content.
• History is bounded. The card keeps only the last three snapshots in its drift-debugging history. Older judgments are not retained.

Where to Go Next

• AI Search & Find Similar — semantic search, find-similar panels, and the broader CRM AI surface.
• Deals — pipeline, stage progression, and deal-health scoring that pairs with the intelligence card on deal pages.
• Contacts — managing contact records, presence, and the spreadsheet view.
• Tasks — log-complete with outcome + note to give the next snapshot a richer read.