Smart Box
Every reply is read and sorted — reply needed, meeting, possible opportunity, needs review — with the reason it was flagged and a draft answer ready.
An AI sales workspace for email and WhatsApp. It reads every reply, drafts the answer in the rep’s voice for them to approve, and moves the pipeline forward from what actually gets sent.
Outreach tools made sending easy, but the work after a prospect replies — reading it, deciding what to do, following up, updating the CRM — still happens by hand, and that is where deals go quiet. Ferouchi joins outreach, replies and the pipeline into one loop for sales teams, and a person approves every AI reply before a prospect sees it.
Every reply is read and sorted — reply needed, meeting, possible opportunity, needs review — with the reason it was flagged and a draft answer ready.
Replies follow the rep’s own tone and the team’s approved knowledge, with warnings when a draft quotes an unapproved price, makes a banned claim or still has a placeholder.
Email sequences and Meta-approved WhatsApp campaigns run from one workspace, and each contact keeps a single conversation history.
Sending a scheduling link or a proposal moves the deal to the next stage, with the source recorded. Ferouchi never moves a deal backward or closes one by itself.
Mailbox warm-up, safe daily limits spread across mailboxes, a preflight check before each campaign and an automatic pause when bounces rise.
Role-based access, private or shared mailboxes, an audit log of every AI action, and AI routed between a private model and an online provider.
Prompts, output schemas and safety rules live in a private Python service running open models. The API sends it a thread and gets back validated JSON, tagged with the model and prompt version that produced it.
Unsubscribes, bounces and out-of-office replies are caught by rules and never reach a model. Deal facts only stand when a quote from the thread backs them, and a deal is only marked won on the prospect’s own words.
AI work goes to the private model by default, with an automatic fallback when it is busy or failing, and the router backs off a failing model for a while instead of retrying it on every request.
The team’s approved knowledge is embedded and searched with MongoDB Atlas Vector Search, so drafts use real pricing and positioning. A policy check then flags invented prices, banned claims and unfilled placeholders.
Sends, sequences, warm-up and inbound mail run on BullMQ workers in their own process. Scheduled events go through an outbox with unique ids, so a retry never sends twice and a pending send can still be cancelled.
Gmail and Microsoft 365 connect over OAuth through the Google and Microsoft Graph APIs, and WhatsApp runs on Meta’s Cloud API with templates, consent and opt-out tracking.







