SocialHub.AI
Try CWMBook a demo
Resources · Videos

See the Platform in Motion

Product tours, explainers, and stories from SocialHub.AI — explore the Consumer World Model, AI colleagues, and the complete marketing loop.

Visit our YouTube channel ↗

Visit our channel for the latest releases.

Latest video

The Consumer World Model — From Our Founder & CEO

Bill Huang, founder and CEO of Socialhub.AI, on why we built the Consumer World Model. A language model has never met your consumers — it writes a beautiful message but doesn't know who's about to buy or about to leave. Old analytics flattens each customer into a few averages. So we built a model that learns from real behavior in time order — every visit, every order — trained on 50 million members and 890 million behavior records, and improving recursively: results feed back into the model every day, so this month's model beats month one. Live today with 50+ global brands across 12+ countries, covering over 400 million consumers, built with SAP as a partner. A photo tells you where a consumer has been; a world model tells you where they're going.

Open & share →

Explore the video library

Product tours, explainers, and stories. Latest releases first.

28 videos

The Consumer World Model — From Our Founder & CEO

Bill Huang, founder and CEO of Socialhub.AI, on why we built the Consumer World Model. A language model has never met your consumers — it writes a beautiful message but doesn't know who's about to buy or about to leave. Old analytics flattens each customer into a few averages. So we built a model that learns from real behavior in time order — every visit, every order — trained on 50 million members and 890 million behavior records, and improving recursively: results feed back into the model every day, so this month's model beats month one. Live today with 50+ global brands across 12+ countries, covering over 400 million consumers, built with SAP as a partner. A photo tells you where a consumer has been; a world model tells you where they're going.

Open & share →

The Full Loop — One Marketer, One Complete Campaign

Follow one marketing manager through a complete cross-border campaign. It starts with a high-value audience and the Customer World Model's read on who these people really are, then an ad campaign built in the operator with a 40/60 Google/Meta split. New members arrive, and one typed business goal becomes a full campaign — audience, schedule, email and send, live as components on the canvas, with a branded EDM and hero illustration already in place. Before a dollar goes out, every A/B variant is played against member digital twins: simulate first, then spend. Behavior flows back into the model, orders drive a per-customer Next Best Action, and members crossing the churn line enter the win-back queue automatically — one plan per person, with 23% incremental lift measured against a holdout, not guessed. Then she reviews the whole campaign just by asking. Your stack stays your stack — what changes is the decision behind it: from segment operations to customer-level decisions. English narration, real product capture.

Open & share →

SocialHub.AI for SAP — A complete cross-border campaign, from audience to review

Follow a marketing manager through a complete cross-border campaign: begin with the high-value audience synced from Emarsys and the Customer World Model analysis, then create a Google/Meta advertising campaign with a 40/60 split in Joule with one request. New members arrive, and one business goal generates a complete campaign — audience, schedule, email and delivery appear directly on the canvas, with branded EDM and illustrations prepared by AI. Digital twins simulate before spending, and the campaign returns to Emarsys for normal execution. Behavior flows back into the Customer World Model, orders drive a per-customer Next Best Action, and members crossing the churn-risk threshold automatically enter the win-back queue — one plan per person, with 23% real incremental lift measured against a control group. Finally, review the campaign directly in Joule. SAP remains SAP and Emarsys remains Emarsys; the decision behind them changes from segment operations to customer-level decisions. Real product footage, Chinese narration.

Open & share →

The Win-Back Shift — Understand and win back each member

Each year, 20–30% of members approach churn — six to nine million people need to be won back, for different reasons: poor service, price sensitivity, or simply moving house. Sending everyone the same coupon does not work. When the Consumer World Model identifies churn risk, it automatically queues members for win-back. Multiple AI employees follow configured Skills and standard operating procedures within their own guardrails, diagnose the reasons, and create one plan per person: autonomously provide care and win-back outreach, hand the case to a person when information is insufficient, or take no action when the member is likely to return anyway. Real product footage follows all three branches; a holdout group measures actual incremental spending. Chinese narration.

Open & share →

AI Retail Sell-Through — From member assets to real in-store sales

As the market adjusts, inventory is stuck in the channel instead of becoming bottles opened by consumers. Headquarters has 30 million members, but stores are still waiting for walk-ins. Members choose a local store and order nearby; stores see CWM prompts for their assigned members, trigger outreach in one click, plan in-store VIP tasting events, and measure actual sell-through with bottle-opening QR scans. Consumer intelligence at headquarters becomes a sales capability every store associate can use. Today's opportunities also include business customers: a VOC data integration reveals that a corporate buyer recently booked a partner conference, so AI identifies the event's baijiu demand in advance and proactively makes contact. Actual store-manager workbench footage, Chinese narration.

Open & share →

Supplements: The Right Move for Every Customer

For a DTC supplements brand where every needless discount comes off a thin margin, the profitable Next Best Action engine reads five signals per member — purchase imminence, churn risk, price sensitivity, value tier, intent — then scores five moves (Wait / Reminder / Offer / Bundle / Service) by expected contribution, not by which is easiest to close. Five customers with the same "ready to buy" signal get five different right answers — Sarah left alone, Maria a free reminder, James a basket-building bundle, Linda a nudge, David a service credit — so no discount is ever wasted. Every option is scored right on the member's own page: a planning estimate over real unit economics, never inflated into proven lift.

Open & share →

Socialhub.AI for SAP Emarsys

Maya runs marketing at a fast-growing retail brand on SAP Emarsys — with a hard line from her CEO: cut the agency, still lift repurchase. This is Socialhub.AI for SAP Emarsys, the AI email layer that sits on top of Emarsys. Maya describes an email in one sentence and gets a finished, on-brand, personalized campaign — greeted by name, with the member's own points and coupon. Product Grid and Coupon blocks pull products, prices and offers straight from the Shopify catalog and the SocialHub CDP — member, order, and recommendation data hydrate every block. One click pushes the finished campaign back into Emarsys as a draft, with its own campaign ID: Emarsys stays the engine. Built-in tracking and a lightweight SDK give true attribution across opens, clicks, visits and purchases, and a purchase-intent score for every member hands Maya the next campaign to run. All of it runs through SAP Joule, informed by CRM and ERP data. Predict. Personalize. Create. Act — inside the engine she already runs.

Open & share →

Business Goals: Set the Target, Break It Down, Track It Home

Yunshang, an apparel chain, sets one big goal — sixty million in annual GMV. But a number on the wall answers to no one. This is Business Goals: set a target on any governed KPI, then break it down two ways on one tree — by time (year → quarter → month, with Q4 carrying the 11.11 and year-end peak) and by organization (the whole company → store regions → each store). Every goal tracks the same metric your dashboard uses; the actual figure rolls up automatically from real orders — store into region, month into quarter, quarter into year — with red / yellow / green and a pace projection showing at a glance which store is falling behind, and an alert the moment a goal turns off track. Includes a walkthrough of the live product — the goals board and the hierarchy tree rolling up in the real app. Available in English and 中文.

Open & share →

Shopping Mall: Cross-Tenant Loyalty in Action

A shopping mall stops being a landlord and starts managing consumers alongside its merchants — through one cross-tenant loyalty network. Follow May, newly arrived in Hong Kong, and her friend Sarah, start to finish: scan to join, earn points from any store's receipt, redeem a reward, invite a friend so both earn, ask the member assistant what's on, grab a coupon, and even generate a personal key to connect the membership into Claude Code. Then see the engine behind it — the merchant redemption terminal (scan or type the code, even on a weak signal), the self-serve store dashboard, the cross-merchant member network and behavior model, and win-back. And on the operator's side — stored-value float with hard ceilings, month-end settlement with POS three-key reconciliation, the AI coupon matrix, and compliance as the foundation. From collecting rent to lifetime value — turning footfall into loyalty. Available in English and 粵語 (Cantonese).

Open & share →

Enterprise Skill Library: Stop Improvising. Run Your Playbook.

11:47 on a Tuesday night, Maya abandons an $86 cart — and the question is what your AI agent does next. Improvise a coupon and teach her that walking away pays? Flash's answer is the Skills Library: your best play written down as a managed asset — trigger, required Consumer-World-Model state, tool allowlist, exact steps, and hard rules a machine-checkable verifier enforces. A tour of the live Skills console: the governed lifecycle (draft → verified → published, second-person approval, version lineage, deprecation the runtime honors instantly), the 31 built-in skills across six lifecycle stages, then Maya's save end to end — Cart Abandonment Recovery's three-step ladder, where the incentive may only appear at step three, every touch re-reads the timeline, and eval hooks judge the skill on governed metrics, not vibes. Judgment where judgment belongs; rules where rules belong.

Open & share →

The Socialhub.AI Architecture, End to End

The whole platform on one map — three layers. On top, the applications that meet your customer across the full lifecycle: acquire, activate, grow, retain, loyalty. In the middle, two brains working together: a big brain (the LLM) that reasons, writes and plans, and a small brain (the Consumer World Model) that learns each customer's patterns from your own first-party data and predicts the next move — connected by a neural link, where the small brain grounds the big brain and the big brain turns predictions into words and decisions. At the bottom, a governed first-party data foundation that fuels both. Part one walks every layer and what it does; part two follows a real win-back — Mia, quiet for 87 days — as the data flows hop by hop through the diagram: real-time events, a churn score, auto-enrollment, an agent's play, a one-to-one send, and honest holdout attribution — then loops back to retrain. Don't guess — know. Available in English and 中文.

Open & share →

Smart Discounting: Different Discounts for Different Customers

A blanket 20% coupon works — sales go up — but most of it lands on people who were going to buy anyway. Flash asks a better question: who actually needs a discount? The behavior model scores every member's purchase imminence live (80,119 already close to buying get zero discount), preset tags grade discount sensitivity, and the ladder builds itself: no coupon for high intent, 10% for the coupon-sensitive fence-sitters, 20% only where lifetime value justifies it. Digital twins pretest the offer structures before budget burns; branch nodes route each member by tags in one canvas; and for the judgment calls, hire an AI employee on the canvas — brief it in plain language and it picks each member's offer and writes the matching message, inside hard budget ceilings, with a holdout in every tier so the campaign only takes credit for sales it caused.

Open & share →

Ads on Real Behavior: Targeting, Delivery & the Closed Loop

A $50,000 ad budget usually means guessing twice — who to target, and what happened after the click. This is the same money with nothing guessed. Captured behavior trains the brand's private world model: 200,000+ members carry a live prediction, 80,119 flagged close to a purchase. Predictions become audiences in Segments — imminence bands, behavior-embedding lookalikes — then push to Google and Meta as hashed, consent-filtered Customer Match lists; do-not-sell never syncs. Digital twins pretest the offer before budget burns. And the click comes home: the landing page carries Flash's tracking SDK, so page views, UTM source and referrer stream back as first-party behavior — anonymous at first, stitched to the member at login — measured against holdouts, retraining the model.

Open & share →

Your Customers Are Talking: NPS, Chat & Behavioral Signals

After the Mother's Day campaign closes, one question remains: how did it actually land? Your customers are already answering — in three languages. What they say: the portal's NPS survey rolls every score into one dashboard — promoters, detractors, the full distribution, and a written reason attached to each response. What they ask: a member-facing concierge answers from the member's real account behind six fail-closed gates, and every question lands as a de-identified feedback signal. What they do: every visit, click, check-in and redemption on one member timeline, where the behavior model reads the pattern, predicts the next move, and turns listening into a suggested next step with a confidence score. Three languages, one member record.

Open & share →

Preset Metrics & Tags: Answers on Day One

Maya is five days into running CRM at a 197,000-member brand when the board asks: how many members are we losing — and who, exactly? Flash answers on day one, because the definitions ship with the platform. 129 preset metrics across nine business categories on one governed semantic layer — certified badges, tunable calibers, version-controlled definitions, and the same number on every surface from dashboards to agents. And 29 preset tags — seven exclusive profiles plus live, sticky and manual lifecycles the engine keeps re-evaluating — that have already profiled all 197,550 members: 104,945 at risk. Three tag chips build the win-back audience, and a certified uplift metric only takes credit for returns it actually caused.

Open & share →

Who's Watching the Money? Campaign Cost Governance

One Mother's Day campaign at a 20-store retailer, and the three gates its $200,000 coupon budget walks through. Tiered authorization gives HQ, regions and stores their own campaign spaces with their own approval chains; a two-step sign-off reviews the spend and locks the budget into a hard ceiling the system itself enforces at issuance time. Then the campaign goes live: the in-flight dashboard catches one store burning through its quota in under three days and another paused at its limit, a one-click top-up moves cards from the HQ reserve, Mother's Day demand triggers an audited break-glass raise with a mandatory 24-hour review — and when it's over, finance exports the full cost ledger: every card, every store, every signature.

Open & share →

Build on Flash — Open by Design

Flash isn't a black box — it's a platform you build on, with three ways in on one governed core. An MCP server with 70+ tools lets Claude or your own agent operate Flash directly, every call scope-checked, guarded and logged. A full REST API with webhooks and native connectors — Shopify, Amazon, TikTok Shop, POS/ERP — plus zero-copy queries into Snowflake, Databricks and Salesforce Data Cloud. Embed any page into your product behind SSO and SCIM. Extend the model itself, shown live in the console: describe an object in a sentence and AI drafts it, define metrics as SQL over governed views. Even the AI employees take your skills and playbooks. Every door runs the same spine: scoped keys, fail-closed guards, one certified metric layer, full audit trail.

Open & share →

Privacy & Consent, by Construction

Privacy as machinery, not a page in the footer. AI drafts your policy into a reviewed, versioned skeleton — it never invents a legal fact — and every member's clickwrap binds to the exact published version, with consent models for 94 countries built in. Saying no is one click: unsubscribe honored everywhere, a preference center, and Global Privacy Control obeyed automatically. Member rights run on rails — DSAR export, erasure down to the model embeddings on day zero, retention that actually deletes, names masked before any text reaches a language model, and every look at personal data logged.

Open & share →

Store & Location Marketing: Reach Members Near Your Stores

Location marketing done properly: a lightweight SDK inside your own app, behind your brand and your consent screen — Flash never tracks anyone from the outside. Every store becomes a geofence with its own radius in meters, and each member's phone quietly watches the twenty nearest fences, always up to date. Crossing one lands as a first-party behavior — entered or lingered — deduplicated, rate-limited, never a raw coordinate; campaigns and journeys trigger on near-a-store like any other behavior, with holdout-proven lift.

Open & share →

Loyalty, in Your Members' Pocket

A tour of the loyalty module, led by the portal your members hold in their hand: build every screen as blocks with AI and publish without developers, then walk the member side — tiers and perks, a transparent points ledger, the rewards mall, coupon wallet, referrals, social and UGC, ambassador program, stamp cards, wallet passes, marketplace-order self-reporting, the Loyalty Concierge, and a member-generated read-only MCP key for their personal AI.

Open & share →

Zero-Copy & Enterprise App Integrations

Integration without copies. Zero-copy queries Snowflake, Databricks, Salesforce Data Cloud and Delta Sharing where the data lives — live queries, no ETL, no permanent copy — through a governed five-step loop: pick the data, spend-check first, query the warehouse, hold only a temporary snapshot, record the cost in the audit log. Enterprise connectors — Shopify verified points, Amazon & TikTok Shop, POS/ERP/OMS via one governed API, fraud-checked receipts — land everything on one verified customer record behind the semantic layer.

Open & share →

Decide: Strategy, Analytics & Insights

Three ways Flash reads your business, all on one governed semantic layer so they can never disagree. Analytics: live dashboards with certified metrics, the behavior radar, and AI suggestions — the Engage funnel flags its own bottleneck. Insights, shown live: a question typed in plain language — how's business this month? — answered from certified metrics with a suggested next step, then saved to the personal analytics feed in one click, exportable as Markdown or HTML. Strategy: MarPro, the AI research analyst — a real Deep run shown end to end: issue tree, parallel research, 53 sourced facts, seven chapters, G-Eval quality gates, and a boardroom-grade report with 11 cited sources. English narration with burned-in subtitles.

Open & share →

One Sentence. One Campaign.

Tell the AI your goal in plain words — win back lapsed members with a 15% coupon, email plus push, warm tone — and watch Flash draft the real campaign: type, audience resolved against live data, offer, channels, as an editable draft that never sends without you. The email arrives written and rendered in your brand voice from the Brand Center, the audience estimates live, digital twins rehearse the send, coupon governance guards the budget, and Track measures it all on one calendar.

Open & share →

Capture: The AI-Ready CDP

Ask a traditional CDP what GMV was last month and you get three answers; point an AI at raw tables and it's confidently wrong. Capture is the AI-ready CDP: every touchpoint captured live at 200M+ member scale, then — the hard part — a governed semantic layer on top: a business ontology, governed tags, and certified metrics defined once on one engine. Agents ask the layer and get certified answers, every consumer reads the same number, and governance and consent are built in, not bolted on.

Open & share →

What Is a Consumer World Model?

What a Consumer World Model is: a multi-expert behavioral model learned from 55 million members and 890 million behaviors, then trained privately for each brand — the Behavior Model learns the past, the Digital Twin rehearses the future. Inside, a fleet of small domain experts is combined by a learned gate that beats a single pooled model on every arena. The video maps its 23 capabilities across eight families — what the fleet wins, what stays with classical ML, and what splits by domain — then walks two cases start to finish: winning back a lapsing VIP, and converting an about-to-buy audience. And one honest rule: if it can't beat the baseline, it never goes live.

Open & share →

Meet the So Family: AI Colleagues on Governed Data

Three AI colleagues, one governed data layer. SoCode pulls certified numbers into your IDE and turns them into working software — governed data in, dashboards and decks out. SoTag makes Slack a shared workspace for humans and agents: a real captured session where a −61% GMV alert gets diagnosed live, honestly. SoMe gives every member a personal AI connection — mint a key from the portal, check points, redeem rewards, hands-free.

Open & share →

AI Employees: The Win-Back Shift

A tour of the live Win-Back console: the guardrails you set once (holdout, complaint breaker, money limits, kill switch), AI employees with budgets and goals, and their task queue. Then a real agent session in five moves — boot, analyze a gone-quiet $46k regular, decide on a $20 surprise gift, pressure-test it on 188 digital twins, act — and when delivery snags, it hands that item to a human in the "Needs you" queue, every step on the record.

Open & share →

Want a Guided Walkthrough?

Book a live demo and see the Consumer World Model on your own use cases.

Book a Demo