Conversational AI & Chatbots

A Multi-Agent Voice & Chat Shopping Assistant Built on a Coordinated Squad Architecture

How Pfactorial Technologies built Jessica, a retail assistant that handles product search, order management, returns and appointment booking through a coordinated group of assistants behind one seamless session.

August 21, 2026

OUTLINE

ENGAGEMENT SNAPSHOT

Pfactorial_Case_Study_Retail_Assistant_Jessica image 1

Figure 1 - Key figures from this engagement, at a glance.

EXECUTIVE SUMMARY

Our client wanted a retail assistant that could handle the full span of customer support - product questions, order tracking, returns, appointment booking - across voice and chat, seven days a week, without routing customers between disconnected systems.

Retail support genuinely spans multiple domains - inventory, order management, scheduling - each with its own data source and logic. Handing all of that to one monolithic assistant strains a single behavior definition, but customers shouldn't feel like they've been transferred between systems either.

Pfactorial built Jessica using a VAPI Squad - a coordinated group of assistants behind one Squad ID - so retail-specific domains are handled by specialized logic while the customer experiences one continuous conversation from first contact through a resolved return or booking.

Why this engagement is representative This engagement shows Pfactorial's approach to conversational AI that spans multiple business domains - using a coordinated multi-assistant architecture where it genuinely helps, without letting that coordination become visible or disruptive to the customer experience.

THE CHALLENGE

Serving the full span of retail support through one assistant experience required solving for domain breadth without sacrificing conversational continuity.

1. Retail support spans too many systems for one flat assistant

Inventory lookups, order management, returns processing and appointment booking each have distinct data sources and logic - treating them as one undifferentiated capability set strains a single assistant definition.

2. Real-time data has to actually be real-time

Stock levels and order status shown to a customer have to reflect the current system state - stale inventory or order data directly costs sales or creates support escalations.

3. A return needs to feel resolved, not deferred

Customers expect an RMA number and a refund timeline in the moment, not a promise that someone will follow up later.

4. Multi-assistant coordination has to stay invisible

Using specialized assistants behind the scenes should never surface to the customer as a handoff, delay, or loss of context mid-conversation.

The real brief Not “build a retail chatbot” but “coordinate specialized retail logic behind one continuous customer conversation, so the architecture never becomes the customer's problem.”

THE SOLUTION

Pfactorial built Jessica around a VAPI Squad architecture with a central tool-dispatch layer, so specialized retail logic runs behind the scenes while the customer experiences one uninterrupted session.

Pfactorial_Case_Study_Retail_Assistant_Jessica image 2

Figure 2 - From contact to a resolved retail action, in one session.

Architectural principles

Squad architecture for coordinated specialization - A group of coordinated assistants referenced by one Squad ID handles the session, rather than a single assistant carrying every retail domain's logic at once.
Central tool dispatch - Every tool call - inventory, order, booking - routes through one /vapi-tools endpoint that dispatches to the correct handler, keeping backend logic centralized and consistent regardless of which part of the squad triggered it.
Alternatives over dead ends - When an inventory check returns unavailable, the response includes a reason and a list of alternative products rather than a bare decline, keeping the conversation productive.
One session, multiple resolved actions - A customer can check an order, initiate a return, and place a replacement order within the same conversation, without restarting or losing context between actions.

CAPABILITIES DELIVERED

Each capability replaces a step a customer would otherwise navigate across separate retail systems.

CAPABILITY	WHAT IT DOES
Product & inventory search	Natural-language catalogue search with real-time stock levels and pricing.
Order status & tracking	Lookup by order number or linked phone number, with full order details.
Returns & exchanges	RMA generation, refund timeline guidance, and same-session replacement ordering.
In-store appointment booking	Consultation and fitting appointments with real-time availability checking.
Promotional code validation	Coupon and discount code checking against order totals.
Full session history	Paginated past sessions with transcripts and recording playback.

Pfactorial_Case_Study_Retail_Assistant_Jessica image 3

Figure 3 - A squad of coordinated assistants, indistinguishable from a single-assistant session.

Design note Routing every tool call through one central endpoint - rather than letting each part of the squad talk to backend systems independently - kept the retail logic consistent and made adding new tool calls (like the promotional code check) a matter of extending one dispatcher, not touching every assistant.

ENGINEERING FOR SCALE AND RELIABILITY

Five decisions keep the squad's coordination invisible and the session resilient to real-world edge cases.

Structured, alternative-aware error responses

Backend responses for unavailable inventory or unmatched orders include a specific reason and actionable alternatives, which Jessica relays naturally instead of a flat decline.

Configuration-guarded initialization

If the VAPI public key or Retail Squad ID environment variable is missing, the system shows a clear configuration-error status rather than attempting and failing a confusing connection.

Safe audio resource cleanup

Audio cleanup is wrapped in try-catch so Web Audio API issues during session teardown never crash the interface, and resources are released immediately to prevent memory accumulation across sessions.

Non-blocking history persistence

If a transcript fails to save after a session ends, the failure is logged for investigation but never affects the customer, since the session has already completed successfully from their perspective.

Lightweight, optional local data model

Session history uses SQLite with optional retail-specific tables for inventory and orders, letting the system operate standalone or integrate with the retailer's existing systems without requiring either.

DELIVERY APPROACH

The engagement built the squad architecture and tool dispatch layer first, then layered retail-specific handlers on top.

1. Squad architecture design - coordinating specialized retail logic behind one Squad ID rather than a single monolithic assistant.

2. Central tool dispatch backend - the /vapi-tools endpoint routing inventory, order, and booking calls to their handlers.

3. Retail system integrations - inventory, order management, and appointment booking logic, with optional local fallback tables.

4. Error & edge-case handling - alternative-aware responses for unavailable items and unmatched orders.

5. Session frontend - the React interface for voice and chat, with live transcription and session history.

6. Performance tuning - paginated history, lazy-loaded recordings, and memory-conscious audio resource management.

RESULTS AND IMPACT

Pfactorial_Case_Study_Retail_Assistant_Jessica image 4

Figure 4 - Key outcomes from this engagement.

Customers get end-to-end retail support - from a product question through a completed return and replacement order - in one continuous conversation across voice or chat.

The squad architecture lets the client extend or adjust individual retail domains (inventory, orders, booking) without redesigning the entire assistant experience.

What it enabled commercially

The client gained 7-day, always-available retail support that resolves the full range of customer needs in a single session, rather than routing customers through disconnected tools or deferring resolution to a follow-up.

WHY PFACTORIAL

This engagement reflects Pfactorial's conversational AI product engineering service line, applied to a domain where coordinating multiple specialized capabilities behind one seamless session is the actual engineering challenge.

Pfactorial_Case_Study_Retail_Assistant_Jessica image 5

Figure 5 - Service lines this engagement draws on.

Engagement enquiries Pfactorial Technologies works with retailers and e-commerce operators that need conversational support spanning multiple business domains. If you're evaluating a multi-domain AI shopping assistant, we're happy to give you an honest read on scope and risk before anyone commits to anything. · pfactorial.ai

APPENDIX A - TECHNOLOGY STACK

The technology stack underpinning the system, grouped by the layer it serves.

Pfactorial_Case_Study_Retail_Assistant_Jessica image 6

© 2026 Pfactorial Technologies. Client identity and product-specific implementation detail are withheld or generalized; no client data, credentials, source code, or infrastructure detail is included in this document.