CloudSell Copilot

Overview

CloudSell Copilot is an AI assistant embedded directly inside CloudSellPOS — the cloud ERP/POS platform covered in a separate case study — that lets any user ask financial and operational questions in plain Arabic or English and get an instant, data-backed answer instead of hunting through reports and menus.

In Action

A live conversation answering a sales comparison question with a chart, and the same assistant embedded side-by-side with the rest of the workspace.

The Challenge

A full accounting and POS system accumulates dozens of specialized reports and screens — sales, account statements, inventory, journal entries, rep performance — necessary breadth that becomes a daily burden. Reaching one simple figure meant knowing exactly which report held it, opening it, and setting the right filters, and new users faced a real learning curve just locating where information lived. Any AI layer added on top had to answer with real numbers, not generated guesses, and could not be allowed to see or reveal data the asking user wasn't already permitted to see.

Engineering Solution & Architecture

Copilot is a conversational layer sitting on top of the existing platform — not a replacement system — that turns a natural-language question into a governed, live data query:

• Semantic intent understanding: recognizes a question's intent by semantic similarity rather than keyword matching, so different phrasings and Arabic dialects of the same question are understood as one underlying request.
• Tool calling over guessed text: every question type maps to a dedicated action that runs a real query against the system's own data; the model selects the correct action via tool calling rather than pattern-matching free text.
• Permission-aware execution: every query runs under the identity of the actual requesting user, filtered by the exact same module- and record-level permission rules enforced everywhere else in the platform — the assistant holds no elevated access of its own.
• Live, tenant-scoped data grounding: financial answers are never generated text — each one is pulled live from the tenant's own database at the moment of the question, with a direct link back to the full report for anyone who needs to drill down.
• Documentation-grounded help layer: non-financial "how do I...?" questions are answered from a knowledge base built from the product's own documentation, not generic public knowledge that might not match its actual behavior.
• Disambiguation & short-term memory: when a name or item could match more than one record, the assistant presents the matching options instead of guessing, and retains short-term conversation context so follow-up questions don't need to restate everything.

Key Results

Delivered an embedded financial assistant that answers everyday accounting questions in natural Arabic or English with live, permission-filtered data — with zero elevated access, full usage-cost visibility for administrators, and a modular architecture where every new financial skill ships as an independent unit.