Case studyDecathlon

Autonomous B2B Onboarding Agent

A three-week manual onboarding process, reduced to a conversation of a few minutes

Onboarding time

3 weeks → minutes

End-to-end partner integration

Admin burden

Eliminated

No more manual re-keying

Interface

Google Chat

No new tool for business teams

Status

In production

Architecture validated internally

Overview

Decathlon's wholesale teams onboarded every new B2B partner by hand: collecting documents, re-keying data into SAP, checking it, chasing what was missing. I designed and shipped an end-to-end GenAI agent covering the whole path — conversational intake in Google Chat, multimodal document extraction, verification, and automated writing to SAP and BigQuery under human approval.

The Challenge

Integrating a new B2B partner took around three weeks of back-and-forth between the wholesale team and the partner. Every day of delay was a day of deferred revenue, and the process consumed senior team time on data entry rather than on the commercial relationship.

Key pain points

  • Partner data scattered across PDFs, images and email threads
  • Manual re-keying into SAP, with the error rate that implies
  • Back-and-forth on missing or inconsistent fields stretching the cycle to three weeks
  • No traceability on what had been checked, and by whom, before a partner went live
  • Business teams unwilling to adopt yet another standalone internal tool

The Solution

An orchestrated agent, reachable where the teams already work, that collects, extracts, verifies and writes — keeping a human in the loop on every write to the core systems.

1

Conversational entry point in Google Chat

Built on Google ADK and Gemini, the assistant drives the intake dialogue and answers questions in real time through a RAG module indexed on internal onboarding documentation. No new interface for the teams to learn.

2

Multimodal extraction pipeline

PDFs and images — contracts, certificates, identification documents — are parsed into structured fields, removing the re-keying step entirely.

3

Interactive verification & validation workflow

The agent cross-checks collected data, flags gaps and inconsistencies, and asks for exactly what is missing instead of returning a generic error.

4

Automated writing to SAP & BigQuery, human-in-the-loop

Nothing reaches the core systems without explicit human approval. LangGraph handles agent state, Airflow the scheduled steps, and Model Armor filters incoming prompts against injection and data leakage.

Results & Impact

3 weeks → minutes

Onboarding cycle

The purely administrative parts of the process — collection, entry, cross-checking — now run inside a single conversation.

Administrative burden eliminated

Team time reallocated

The wholesale team stopped re-keying documents and went back to partner relationships.

Production-grade guardrails

Human-in-the-loop + Model Armor

Every write to SAP and BigQuery is human-approved, and prompts are filtered before reaching the model — the conditions that made the architecture reviewable and approvable.

Validated for production

By Decathlon's AI architecture team

The system passed internal architectural validation, including data security and GDPR requirements.

Technologies used

PythonVertex AIGoogle ADKGemini 3Model ArmorGoogle Chat APICloud RunLangChainLangGraphRAGBigQueryFirestoreAirflowDocker

He built an end-to-end AI agent solution that I was able to safely validate for production. Being able to rely on an expert like Yanis is invaluable: his autonomy and technical rigor greatly facilitate the process of architectural validation.

Médéric H

AI Architect at Decathlon

Ready to move your AI project forward?

Let's talk about what a similar engagement would look like in your context.

Get in touch