Levanta "Divi"
Agents that act on their own, modeled on monday.com — and a chatbot that helps in the moment, modeled on Upwork's Uma. Built for creators and sellers.
Dormancy and activation:
Accounts still live, that never got started or went quiet.
What each number argues for: 3 in 4 claimed creators go quiet for 90 days, mostly after activating — the case for a re-engagement agent and chatbot alike. The 28% who converted once and vanished are the sharpest version of that pattern. And the creator-alert feature only works if sellers have something new to alert on — 1 in 4 don't, which argues for a seller-side agent that flags a stale catalog, and a digest, not a raw trigger, since large catalogs debut dozens of new products a quarter.
Source: Levanta Team, Product, and daily reporting tables (Hex). Creator figures exclude 71,160 unclaimed Creator Connections placeholder accounts, never signed into. "Stale catalog" uses a product's first real click or sale, not its catalog-sync date. Archived/cancelled accounts are excluded at sync, so dormancy measures live-but-inactive accounts, not formal cancellation.
A chatbot answers. An agent acts.
monday.com draws this exact line between two of its own products — sidekick is a conversational AI Chatbot, built for on-demand questions, while AI Agents run processes end-to-end, monitor continuously, and act on triggers, around the clock. Levanta needs both patterns, not one or the other.
| AI Chatbot | AI Agent | |
|---|---|---|
| Trigger | User types a question | Account event, schedule, or a standing goal |
| Output | An answer, drafted in the moment | The task is already done — link generated, message drafted, rate requested |
| Context | Whatever the user brings to that one conversation | Live account state, page, lifecycle stage, history |
| Initiative | Waits to be asked | Notices and acts first, reports back |
| Scope | One question at a time | Runs a goal continuously, even while you're offline |
Source: monday.com, "AI Agents on monday.com" support documentation.
AI Agents
Reference: monday.com — continuous, trigger-based execution on structured data, with full governance.
Structured records, scoped agents, full governance
monday's agents act on typed board data under defined permissions — not open-ended chat. Levanta's data is the same shape: accounts, partnerships, and commissions as structured records.
| monday.com | Levanta "Divi" | |
|---|---|---|
| Acts on | Boards, items — a typed, permissioned schema |
Seller: Partnerships, deals, commission records
Creator: Storefront links, partnership status, payouts
|
| Agent shape | Purpose-built Expert Agents, not one do-it-all bot |
Seller: A Commission Agent, a Discovery Agent
Creator: A Link Agent, a Re-engagement Agent
|
| Trust layer | Activity tab logs every action + reasoning | An approval gate logs every send, publish, or rate change — for either side |
| Runs on | Triggers, schedules — "even while you're asleep" |
Seller: A weekly creator partnership review, run unprompted
Creator: A payout flagged before you'd notice
|
Source: monday.com, "AI Agents on monday.com" support documentation.
Five actions the agent takes for you
Grounded in what Levanta already tracks per seller account — the agent learns your pattern on each, then acts on it, continuously.
Five actions the agent takes for you
Built around the core lifecycle actions on Levanta — link generation, partnerships, campaigns, samples, and payouts.
It doesn't just do the task. It tells you what to do next.
Executing a task is half the model. The other half is the unprompted suggestion — the agent noticing a pattern continuously and proposing the fix before you'd think to look.
Delivered as a short weekly digest — two or three suggestions, not a dashboard you have to go dig through.
Two sides of one marketplace, helped in the moment
Uma ships distinct features for clients and freelancers, each triggered by a specific action — not a continuous background process. Levanta's seller and creator sides map directly onto it.
| Uma (Upwork) | Levanta "Divi" | |
|---|---|---|
| Client / Seller | Draft a job post in seconds | Draft a deal or campaign brief in seconds |
| Instant interviews score every candidate | Score a creator application on request | |
| Freelancer / Creator | Stand out — instant interview, no cover letter | Lower-friction partnership acceptance |
| AI companion — brainstorm, review, analyze, write | Draft a caption, review a pitch, explain a payout | |
| Triggered by | A specific action in the moment | A specific question the user brings to it |
Source: upwork.com/uma.
Help the moment you ask for it
Not continuous, not unprompted — the chatbot answers exactly what's asked, pulling the same live data the agents already use.
Illustrative — inspired by Uma's on-demand assist pattern, in Levanta's own UI.
One workflow, two models
The agent works in the background. The chatbot shows up the moment there's a question. Together, they close the loop.
Agent monitors
Continuously, in the background
Flags & drafts a fix
A conversion drop, a rate opportunity
Chatbot explains
User asks "why?" — gets the reasoning
Agent executes
Approved — the fix goes live
Example: a creator's conversions drop 20% — the agent notices overnight and drafts a fix. The creator opens the app, asks the chatbot "why is this happening?", and gets a plain-language answer pulling the same data. They approve, and the agent executes the fix — the loop closes without the creator ever having to dig through a dashboard.
Built on what's already shipped
The Levanta MCP already exposes this same data to Claude, ChatGPT, and other AI tools externally. Divi's chatbot and agents would call the same layer — not a new one.
What's still needed — this isn't a finished agent on its own:
Write-capable tools — draft, propose, send — behind the same approval gate already in this deck, plus a trigger/scheduler layer, so agents act on events and schedules instead of only answering when called.
Agencies like Pattern are a different shape of problem
Not one portfolio — a portfolio of portfolios. Levanta's schema already models this relationship, and agency-managed sellers already look different.
What a different shape of problem means for the agent:
Cross-account review — one weekly digest across every managed client, not 40 separate logins
Playbook enforcement — the agency's standard rate/response rules, applied consistently per client
Fast client onboarding — bootstrap a new brand against the agency's existing playbook
Client reporting, reused — the same approval-gate Activity Log doubles as the monthly client report
Source: Levanta Team table, agencyId field (Hex). Sample is small (126 agency-managed sellers) and concentrated in one agency (73 of 136); the gap likely reflects some selection effect — sellers who hire an agency may already be more committed. Agency names/attributes sit in a table not yet available to this query.
Agents that act. A chatbot that answers.
Not one or the other — the background work of an agent, and the in-the-moment help of a chatbot, in the same assistant.