Skip to content
Back to results
Live product

PalmBiz

WhatsApp multi-agent business assistant — live product

PalmBiz is a Hebrew-first AI business assistant that works over WhatsApp and Telegram. An orchestrator (Sonnet) routes every message to one of seven domain agents — documents, customers, inventory, expenses, reports, suppliers and CRM — that act directly on five ERP systems (iCount, EasyCount, Morning, Priority, Invoice4U) and Arbox CRM. Voice notes and receipt photos become structured actions, and it is multi-tenant from day one.

Industry
SMB operations — Israeli businesses
Year
2026
What we ran
Product & architecture, full multi-agent build (TypeScript), ERP/CRM integrations, ongoing operations
Seconds

From a WhatsApp message to a filed invoice, expense or report — in seconds

5 ERPs

Works with the accounting system the business already runs — no migration

24/7

The whole back office — invoices, customers, inventory, reports — day or night

The problem

Business owners live in WhatsApp, but their books live in ERP systems no one wants to learn. Bookkeeping is done by hand, after hours, and error-prone — and every business runs a different ERP vendor.

How we solved it

We built a team of agents: an orchestrator routes each Hebrew message to the right specialist; agents call a unified adapter layer across every ERP; fast-paths bypass the model on routine actions; smart model selection uses a cheap model for reads and a stronger one for writes; and a confirmation guard gates every money-moving action before it runs.

What it does

01

An orchestrator (Sonnet) routes every Hebrew message to one of seven domain agents — documents, customers, inventory, expenses, reports, suppliers and CRM.

02

One adapter layer speaks to five ERP systems — iCount, EasyCount, Morning, Priority, Invoice4U — plus Arbox CRM, behind a single tool interface.

03

Voice notes (Whisper) and receipt photos (vision OCR) turn into structured actions — issue an invoice, log an expense — with no typing.

04

Fast-paths handle confirmations and common patterns without the model; a confirmation guard gates every money-moving action.

05

Smart model selection (cheap for reads, strong for writes) and Redis conversation memory keep it fast and low-cost; multi-tenant from day one, with each business's credentials encrypted (AES-256) and signed webhooks.

Under the hood

  • Multi-agent orchestration
  • Claude (Sonnet + Haiku)
  • WhatsApp / Telegram
  • ERP + CRM adapters
  • TypeScript · Redis · PostgreSQL

What would your expert team take on?

Tell us the work that eats your week. We'll show you which AI expert takes it off your plate — working inside your existing systems — and what it saves you.