Three languages, one small team.
Customers call from France, the UK, the US and Italy. The support desk is a handful of people at head office, strongest in French, thinner in English and Italian.
Customer case study · Le Marquier
A €70,000 support line,
answered for €14,000.
Same callers. Same questions. Eighty percent less money.
Le Marquier makes planchas people save up for. Their customers call in French, English and Italian, and they call most in the weeks nobody on the team is free. Donna picks up.
Premium consumer brand · France · Voice support in three languages
Le Marquier support lineInbound call · French detected
SATURDAY · 20:40Out of office hours
ILLUSTRATIVE CALL · NOT A CUSTOMER RECORDING
01 Product and care questions
02 Warranty intake
03 Dealer lookup
04 Order and delivery status
01 — The context
Premium products.
A phone line that could not keep up.
Le Marquier builds planchas and outdoor cooking equipment in France. The product is expensive, so the service has to match it. In grilling season it stopped matching.
The same six questions.
Product care and seasoning, warranty terms, spare parts, the nearest dealer, order status, delivery dates. Expert answers to questions that never change.
€70,000 a year, and climbing.
Peak-season hold times ran past fifteen minutes. The annual support bill sat at €70,000 and grew with the catalogue, not with the revenue.
02 — On the call
What Donna actually does
between hello and resolved.
Answers in the caller’s language.
French, English or Italian, picked up from the first sentence. No menu, no “press 2”, no apology for the accent.
One line, three languagesKnows the catalogue.
Trained on Le Marquier’s product data, care guides and support documentation, so a seasoning question gets the same answer at 21:00 as at 11:00.
Answers from approved materialTakes the claim while the caller is still on the line.
Model, purchase date, description of the fault. The claim is opened during the call instead of after three emails.
Claim raised in one conversationFinds the nearest dealer.
By postal code, from the current dealer list. A buyer who wanted to see the plancha in person gets an address rather than a promise to call back.
A next step, not a noteReads order status from the system of record.
Connected to the ERP, so delivery questions are answered from the live record instead of a guess.
Status without a human lookupHands over when it should.
Complaints, safety questions and anything outside the approved answers go to a person, with the conversation attached. Escalation runs at 2% of calls.
2% reach a humanOn the same line
Lead qualification by voice.
Qualified leads, straight into Zendesk.
Not every caller is a customer with a question. Some are buyers still deciding which plancha to buy. The same agent tells the two apart and handles the buyers differently.
Works out who is calling.
An existing customer with a support question stays in the support flow. A new prospect is recognised as a lead in the first moments of the call.
Support and sales separated on the callAsks what sales would ask.
The agent qualifies the lead during the conversation, in the caller’s language, so the answers are captured while the buyer is still interested.
Qualified during the callCreates the lead in Zendesk.
A qualified lead is written into Zendesk with what the caller said, so the team picks it up from the tool they already work in instead of from a voicemail.
Qualified leads land in Zendesk03 — The rollout
Live in eight weeks.
Not live on day one.
Nothing was switched over on faith. The agent took a fifth of the call volume first, and the rest only after the transcripts said it should.
- W1–2
Listen to what already happens.
Six months of call recordings reviewed. The top 50 call types accounted for 90% of the volume, and that list became the scope.
- W3–4
Train the agent and pick the voice.
Conversation flows built on the product catalogue, warranty terms and care guides, then tuned until the tone matched the brand rather than a call centre.
- W5–6
Connect the systems.
ERP, CRM and knowledge base wired in so order status and warranty entries are real. The internal team tested it before any customer did.
- W7–8
Soft launch at 20%.
One call in five went to Donna. Real conversations drove the corrections. Full rollout followed once the escalations stopped being about things she should have known.
04 — The results
The bill did the arguing.
80%Lower support cost
€56,000Saved per year
98%Resolved by the agent
0 minHold time
| Measure | Before | After |
|---|---|---|
| Annual support cost | €70,000 | €14,000 |
| Annual saving | — | €56,000 |
| Average hold time | 15 minutes | 0 |
| Calls resolved without a person | — | 98% |
| Calls escalated to the team | All of them | 2% |
| Monthly conversation minutes handled | — | 2,500 minutes |
| Time to deploy | — | 8 weeks |
Figures as published in the Le Marquier deployment write-up on rajsuyash.com. They are Voxdonna’s own published figures, not an independently audited result. Volume is counted in conversation minutes, not calls.
05 — In their own words
The CEO of Le Marquier,
on camera.
06 — How it was built
The engineering,
written up separately.
The short version: one knowledge base in French rather than three, language detection on the first utterance, an ERP connection for anything with a number in it, and an escalation rule that fires before the caller has to ask for a person. The long version — the model choices, the latency work and the two things that had to be rebuilt — is in the founder’s write-up, building the Le Marquier French voice agent.
If your own callers are asking the same six questions in more than one language, the pattern travels. The multilingual support playbook for specialty brands walks through it in order.
Your turn
Bring us your busiest month.
Tell us what your support line costs today and which languages it has to answer in. We will scope one workflow, price it, and say so if the maths does not work.
Book a discovery callA short qualification form first. A call if there is a fit.
Prefer email? hello@voxdonna.com