SocialHub.AI

Capture · the signal model

Events, Behavior & Intent — from what happened to why, now.

Flash reads a member in four layers: raw records become one standardized stream of events, the member's own actions are distilled into behaviors, and a recent window of behaviors is read as intent— the “why now”, always stated as a hypothesis to confirm, never as a fact the platform pretends to know.

The signal modelfour layers
Intent意图

why, now — as a hypothesis

Behaviors行为

what the member did

Events事件

everything that happened

Raw data原始数据

orders · points · visits · coupons

Each layer is built from the one below · read identically by the profile panel, intent tags and the AI agent.

The problem

Behavior is loud, but the “why now” is buried.

A member is browsing, going quiet, or hunting for a deal right now — but the raw records live in a dozen different tables, and most teams only see the pattern weeks later in a report, or bolt on an AI that guesses at raw events with no shared, governed read.

Scattered, mismatched records

Orders, points, visits and messages each live in their own table with their own shape — no single stream to reason over.

Signal buried in noise

System churn — points accruing, offers issued — drowns out the handful of actions the member actually took.

Everyone infers differently

The operator, the segment builder and an AI agent each read intent their own way, so they disagree about the same member.

The model

Four layers, one direction of travel.

Each layer is built from the one below it — raw data standardized into events, events filtered into behaviors, behaviors read as intent. Every step is explainable back to what the member actually did.

01

Raw data

原始数据

The records your systems already keep — orders, points, page visits, coupons, messages — scattered across a dozen tables, each with its own shape.

02

Events

事件

Everything that happens to a member, from every source, standardized into one timeline and one vocabulary — both what the member does and what the system does to their account.

03

Behaviors

行为

The customer-active slice of that timeline — what the member actually did — filtered clean of the system noise, so the signal isn't buried.

04

Intent

意图

The inferred why-now, read from a recent window of behaviors and stated as a hypothesis your team or your AI can confirm before acting.

Layer 2 · Events

Everything that happens — in one shape.

An event is anything that touches a member, from any source, translated into one standard vocabulary — so a purchase, an email click, an expiring point balance and a store visit all read the same way.

Crucially, the events stream is two-directional: it records both what the member does and what the system or brand does to their account — points earned or expired, an offer issued, a device passing a store. That full history is the honest record everything above is built on.

One standardized member stream

  • MemberPurchased in storemember did
  • MemberAbandoned a checkoutmember did
  • MemberRedeemed a couponmember did
  • SystemPoints earnedhappened to them
  • SystemPoints expiredhappened to them
  • BrandCoupon issuedhappened to them
  • SystemEntered a store areahappened to them
  • MemberReferred a friendmember did

Layer 3 · Behaviors

What the member actually did — the signal, minus the noise.

Behaviors are a clean filter of the events stream: only the things the member themselvesdid. Points accruing or an offer being issued is context — it isn't behavior, and it doesn't get to speak for the member.

Kept — the member acted

  • Browsed the portal or a web page
  • Abandoned a checkout
  • Purchased / submitted a receipt
  • Redeemed points or a coupon
  • Checked in at a store
  • Referred a friend or shared content

Kept as context, not behavior

  • ·Points earned or expired
  • ·An offer issued to them
  • ·A device passing a store geofence
  • ·An admin points adjustment

Layer 4 · Intent

Seven intent signals, read from the behavior window.

Intent reads a recent window of behaviors and surfaces the hypotheses it supports — dynamic, refreshed as the member acts, and always explainable back to the exact behaviors behind it.

High purchase intent

Repeated browsing plus an abandoned checkout with no purchase after it — wants to buy but stalled, often on price.

Engaged

Clicked in an email, or opened a push / read an in-app message — reachable and paying attention right now.

Churn risk

Opening emails but not clicking, with activity thinning toward the end of the window — interest may be fading.

Coupon-sensitive

Recurring coupon or points redemptions — responds to incentives; size the offer rather than discount deeper.

Brand advocate

Invited friends or contributed content — likes the brand enough to promote it; recognition may beat a discount.

Newly acquired

Registered or enrolled with no purchase yet in this window — an onboarding nudge, not a win-back offer.

Disengaging

Hold

An explicit opt-out — unsubscribe, consent withdrawal, or an SMS STOP. Treat as a hold, never a target.

Guardrail: Disengaging is a suppression signal — it tells the platform to hold back and give a member space, never to target them with more outbound.

One core, three surfaces

One reading of intent — shared by people, tags and the agent.

Behaviors and intent are computed one way, in one place, and surfaced three ways — so a human, your audiences and the autonomous agent all read identical signals, never three drifting versions of the truth.

For the operator

A “Behaviors & Intent” panel on the member profile

Open any member and see the intent read alongside their behavior stream — the why behind the signal, in plain language, so a person can confirm it before acting.

For audiences

Segmentable intent tags

The same signals materialize as tags you can build segments and campaigns on — intent flows straight into the governed audience layer next to lifecycle and value.

For the AI agent

The same vocabulary the agent reasons with

The autonomous SoClaw win-back agent reads the exact same signals from the exact same behavior stream — so the operator, the tags and the AI never work from different intent.

Honest by design

A hypothesis to confirm — not a verdict to obey.

The higher you climb the layers, the more inference is involved — so intent is weighted accordingly. An agent trusts a certified metric more than a rule-based tag, and a tag more than an intent hypothesis.

Stated as a hypothesis

Every signal reads as “the behavior suggests…”, with the exact events attached — so a person can agree or overrule it.

Absence is not evidence

A signal only ever fires on behaviors that are present. Intent is never inferred from what a member didn't do, and a short window is never treated as a lifetime.

Confirmed before it acts

Intent informs a decision; it doesn't silently pull the trigger. Money-moving actions still pass through human approval and guardrails.

One semantic layer, three projections

State, inference, aggregate — unified by governance.

Intent is the inference projection of the AI-facing semantic layer. Alongside Tags and Metrics, it's one of three governed reads of the same customer — the layer every AI agent, dashboard and API sees the member through, so they never disagree.

A member's intent read alongside their behavior stream — a hypothesis a person can confirm.

Turn everything a member does into an honest read of why — now.

400M+
50+
12+