Fnac Darty: complaints cut fourfold and NPS doubled with AI-powered Voice of Customer

Overview

Fnac Darty runs its marketplace on a simple but demanding rule: it never touches the boxes, yet it owns the experience. The brand promise rests on more than 3,000 active third-party sellers, each handling its own logistics, after-sales and customer contact. When one of them botches a delivery or a repair, customers blame Fnac Darty first, not the seller.

The challenge

Covid turned into a full-scale stress test. Order volumes surged, carriers got overwhelmed, complaints spiked, and the central CX team was always a step behind. It revealed a structural gap: the marketplace was run on a handful of quantitative KPIs, with no reliable way to hear the customer's actual voice at scale.

An unreadable, volatile NPS

Swings by seller, category or market were nearly impossible to explain. Every Monday, two analysts hand-coded a few hundred verbatims, with all the bias that invites, and turning the result into an action plan was slow.

600,000 conversations a month, unread

The back-and-forth among customers, sellers and the call center never got analyzed. Phantom sellers, cut-short legal warranties, off-platform refunds and counterfeits stayed invisible, because none of them set off a complaint or a call.

No CX governance beyond the CX team

Account and commercial teams are measured on growing seller revenue, not on policing quality. With no shared, aggregated view of the qualitative signal, an insight had no path to becoming an action the right team would own.

Goals

• Move marketplace CX from reactive to proactive, protecting the brand promise before the customer is hurt.
• Make NPS readable and explainable at seller and category level, quickly and without manual bias.
• Detect seller risk (fraud, legal non-compliance, unresponsive phantom sellers) at scale, including cases that never surface as a complaint.
• Turn Voice of Customer into an operating system shared across CX, account and commercial teams.
• Shift people from reading verbatims by hand to acting on them, coaching sellers and driving change.

The initiative

The program rolled out in stages over roughly three years, starting small and proving value before scaling.

• 2022 — Proof of concept. A raw batch of marketplace NPS verbatims moved the team from quantitative KPIs to structured qualitative analysis, and paid off within three months.
• Phase 1 — Analysis platform (live in ~2 months). The platform cross-referenced three Voice-of-Customer sources — NPS verbatims, customer reviews of sellers, and the 600,000 monthly conversations — against order metadata (carrier, delivery time, seller history) for one unified read.
• 2024 — Scaling to production. Multi-source analysis went to full production volume across the marketplace.
• Phase 2 — Agentic workflow (live September 2025). Built in about six months, a workflow of six AI agents now detects, scores and recommends actions on seller risk, day after day.

At the center of Phase 2 is the seller-risk agent the team named "Charlie". It scores every seller as low, medium or high risk, drills down to the individual order, and gives concrete recommendations instead of vague flags — for example: on the orders shipped with a consistently late carrier, add one day to the delivery estimate, because that is what is driving the delivery-related verbatims. The workflow drives a weekly cadence: analysis on Monday, a seller review every Wednesday.

Buy-in and collaboration

Bilel Raouli, Head of Marketplace Performance & CX, sponsored the initiative and helped build it. CX defined the risk framework, account and commercial teams became consumers of the same data, and legal ran a full data-sovereignty and security review during procurement.

Change management

Two decisions made adoption stick. First, one interface for every marketplace team, so CX, account and commercial all read the same reality. Second, an explicit human-in-the-loop design: the AI proposes, a person decides. Every consequential call and every legal term is signed off by a person at the Wednesday review. The team went from roughly five days a month reading verbatims and filling spreadsheets to under one, and put that time into coaching sellers.

The role of technology

Glanceable runs on a proprietary AI built over three and a half years on its own compute, not a wrapper around an off-the-shelf model. That narrows hallucination risk and lets the system name themes it was never given up front. Reliability comes from cross-referencing several data sources against order metadata, and from keeping people on the decisions that count. As the team puts it: AI builds the case; a person makes the call.

Results

• Marketplace NPS doubled (x2), from a level already among the best in e-commerce, with variation now explained week over week.
• Customer complaints cut fourfold (÷4) since go-live.
• ~600,000 conversations analyzed and cross-referenced every month, versus a manual sample of a few hundred.
• Day + 1 risk detection on the previous day's exchanges, before any complaint — including phantom sellers and fraud.
• NPS variation analysis in three clicks, by seller and category.
• Manual analysis effort down from ~5 days to under 1 day per month, reallocated to action and seller coaching.
• Corrective action within 48 hours on at-risk sellers and detractor cases.

Key insights

• Quantitative alone misses things. A top seller with an excellent NPS turned out to be answering French customers in English and German, against the rules — no KPI would have caught it, and the fix went out within 48 hours.
• Voice of Customer is a growth lever, not just a control: the same signals that flag failing sellers help the best ones improve.
• One shared view aligns the company and ends the siloed meetings where every team brings its own version of reality.
• Keep a person accountable and you keep the ROI accountable.

Practical advice

• Build a reliable, multi-source Voice of Customer first — it is the highest-leverage step.
• Move your people from spreadsheets to action.
• Keep it simple, end to end.
• Keep a person on consequential calls.
• On a marketplace you have two customers — the end customer and the seller. Read the feedback well and you protect one while growing the other.