Skip to content
Gimpact

Case studies

AI Support Assistant on WhatsApp

A grounded assistant answering order, delivery and returns questions in Arabic and English, with a designed handover to a human.

Sample project

Illustrative project — regional retail group

Related services

AI SolutionsAutomation

Technology

Claude API · Python · pgvector · PostgreSQL · WhatsApp Business API · AWS

Illustrative example. Not a real client engagement — results are described qualitatively because no measured client figures have been cleared for publication.

01Challenge

Challenge

A three-person support team answering the same order-status, delivery and returns questions hundreds of times a week, in Arabic and English, across WhatsApp and Instagram — with nothing outside working hours and no record a manager could search.

02Approach

Approach

  1. 01

    Started from 600 real historical conversations to find what customers actually ask, rather than what the FAQ page assumes.

  2. 02

    Grounded every answer in the order record or the published policy, with the assistant explicitly unable to state a price or date it has not retrieved.

  3. 03

    Built an evaluation set including Gulf-dialect and code-switched messages, and gated launch on accuracy against it.

  4. 04

    Made escalation a first-class path: the human receives the transcript, the retrieved sources and the assistant's uncertainty.

03Solution

Solution

A retrieval-grounded assistant connected to the order system and the returns policy, able to answer in Arabic, English or a mix of both, and to escalate anything outside its scope with the full conversation context attached.

04Features

Features

  • 01

    Grounded answers with citations

    Every response is retrieved from the order record or the published policy, and carries a source the support team can check.

  • 02

    Arabic, English and code-switching

    Handles Gulf-dialect phrasing and messages that switch language mid-sentence, which is how customers actually write.

  • 03

    Designed escalation

    Out-of-scope questions route to a human with the transcript, the sources consulted and the assistant's confidence — not a dead end.

  • 04

    Hard guardrails

    The assistant cannot state a price, date or policy it has not retrieved. Refusing is a valid answer; inventing one is not.

  • 05

    Evaluation harness

    A fixed set of real historical questions, including adversarial ones, run before every prompt or model change.

  • 06

    Full transcript logging

    Every conversation is stored and reviewable, so quality is something the business inspects rather than hopes for.

05Technology architecture

Technology architecture

How the pieces fit together.

  1. Frontend

    • Next.js admin console
    • TypeScript
  2. Backend

    • Python
    • FastAPI
    • Node.js webhooks
  3. AI

    • Claude API
    • Retrieval-augmented generation
    • Evaluation harness
    • Guardrail policies
  4. Database

    • PostgreSQL
    • pgvector
  5. Infrastructure

    • AWS
    • Docker
    • Queue workers
  6. Integrations

    • WhatsApp Business API
    • Instagram Messaging
    • Order management system

06Results

Results

  • Removed repetitive manual replies from the support team's day.
  • Extended coverage beyond working hours without adding headcount.
  • Created a searchable record of every customer conversation.
  • Made assistant quality measurable against a fixed evaluation set.

Outcomes are described qualitatively. No measured figures are claimed for illustrative work — real metrics appear here only when a client has approved publishing them.

More walkthroughs

  • 01Sample project

    Real Estate Sales Platform

    One inventory record, automatic lead routing and a commission ledger drawn from the same data as the sales pipeline.

    Industry

    Real Estate

    Related services

    Custom Software · Web Development · Automation

    Technology

    Next.js · Node.js · PostgreSQL · Redis · WhatsApp Business API

    Read the walkthrough

  • 02Sample project

    Field Operations SaaS

    A multi-tenant scheduling and job-management product with an offline-first technician app, so work is captured where it happens.

    Industry

    Construction & Development

    Related services

    SaaS Platforms · Mobile Apps · UI/UX Design

    Technology

    Next.js · React Native · PostgreSQL · Prisma · Stripe

    Read the walkthrough

  • 03Sample project

    Clinic Booking & Reminder Automation

    Online booking against real practitioner and room availability, with automated multilingual reminders and no-show reporting by branch.

    Industry

    Healthcare

    Related services

    Web Development · Automation · Custom Software

    Technology

    Next.js · Node.js · PostgreSQL · Redis · WhatsApp Business API

    Read the walkthrough

Next step

Have a similar challenge?

Most of what you have just read can be adapted. Tell us your version of it and we will come back with an approach, a timeline and an indicative range.

We reply to project enquiries within one business day.