Convert and deflect around the clock, so your team can handle the rest.
I build and run AI agents for online retailers. They turn product questions into orders and clear the order and delivery questions filling your queue, day or night. They reply in your voice and hand to a human the moment a decision needs a person.
35,000+*
conversations handled in two months
85%*
average customer satisfaction
40%*
requests resolved without escalating to the team
*Measured in production, on live customer conversations.
Your store
AI agent, replying now
An example exchange. Lookups shown as they happen.
Trusted by
What it does
It reads your live order, delivery and catalogue data to answer the questions that actually fill a queue, and it goes live on your own catalogue in under a week.
Opens the actual order
Status, carrier, delivery window and what was bought, pulled live as the customer asks.
Answers from your catalogue
Dimensions, materials, compatibility, will it fit, even from a photo of what they want.
Knows what it must not answer
Refunds, money, commitments. It says what it knows, never invents a reason, and stops there.
Two models, one send gate
One model writes the reply, a second checks it is safe. Only grounded answers send on their own.
Hands over without losing the thread
The human gets the transcript, the order and what is established. No repeating.
Rides a surge without a deploy
Arm a promo or a shipping delay as temporary context; it expires on its own.
Track and fine-tune within your admin dashboard
Included with every deployment. Your team steers it directly.
- Every conversation, with the exact order, tracking or product data it pulled to answer
- Corrections stick: a miss this week becomes a rule next week
- Email replies your team approves before they send, once you extend it there
- Volumes, satisfaction, what it handled alone and what it escalated, in one place
- Arm a promo or a shipping delay as temporary context, expiring on its own
- One toggle routes everything back to your existing setup, any time of day
Security
Connecting an agent to your support stack means giving it access to real customer data. Here is what it can reach, what happens to that data, and what stays under your control.
// enforced in the code path itself
Connecting your stack
Least-privilege access
Read-only scopes to the data it needs: orders, delivery, catalogue. The key you issue can look things up and nothing else.
Encrypted, and under your control
Credentials are held as encrypted secrets and customer data is encrypted in transit. The access you grant stays yours to change or withdraw at any time.
Your data, and how long it stays
Used only to run and improve your agent, isolated per client. You set the retention window, and everything is deleted on request.
Where it runs
On infrastructure I build and operate, encrypted in transit, with a separate environment per brand.
GDPR
Lawful basis and roles
You remain the data controller. I act as your processor, under a DPA signed before anything connects.
Data minimisation
The agent reads the fields an answer needs, scoped to the customer asking, and keeps no more than that.
Subject requests
Access, correction and erasure requests are answered from the logs, per conversation or per customer.
One person accountable
One named person builds it, runs it and answers for it, so there is no chain of vendors to trace when a privacy question comes up.
Integrations
The agent sits inside your conversation channels. Answering a customer means calling your back-end and third-party APIs and reading what they return, mid-conversation, not at the end of it.
Back-end
Conversation channels
Third-party APIs
Ratings from real customers
Every conversation ends with a thumbs up or down and an optional comment. These are real ones, left by the customers the agent served.
“The best AI Assistant that I've encountered. Questions are understood and the answer is directly related to the question.”
“The AI ADVISOR is impressive. I want the same one for my company.”
“Best ever if this is AI!! I have customizations from the day before, so we'll get it sorted out via phone or email. thanks!”
“Excellent service, I am stunned by the quality of the exchange, bravo!”
“One of the best customer service chats I have dealt with. Thank you so much for all your valued help.”
The satisfaction score behind them
Chat satisfaction over one week: 388 rated good against 60 rated bad. Across the last two months the agent has held 85% over 35,000+ conversations.
Who builds it
Plenty of what is sold today is a chat window on top of a help centre, set up in a week by people who never worked a support shift. It demos well and comes apart on the questions that generate real tickets. This was built the other way round.
Fabien Monnier
Gem CX
A decade in support, first
Ten years in CX, several of them managing a support team through tens of thousands of conversations a year. The hard parts of this system, escalation quality and knowing when to stay quiet, come from that.
I built the tools before the system
Years building custom applications for support teams, one problem at a time. This agent is those pieces brought into one system, written against your own APIs, not a template dropped on a help centre.
I run it every day
It sits on infrastructure I maintain: health checks, automatic retries, a backstop that catches handovers the agent promised but never fired, and a satisfaction survey on every conversation. Logs are monitored daily.
Where built-in AI stops
The AI built into a helpdesk takes you from no self-service to a reasonable one quickly, and that is worth something. Going from reasonable to excellent is the hard part: months of tuning against limits the platform sets for you. This agent has answered the same web chat, the same customers and the same questions as a leading built-in AI, and retailers have moved across more than once.
Hard cases already solved
Reading a prescription from a photo. Finding a product from a picture of a living room. Resolving a carrier's real tracking number buried in a secondary field. Blocking the agent from inventing why an order is on hold. Every one of these was built for a real case. Yours gets the ones that match your own.
You work with me directly
One person builds it, runs it and answers when something breaks. Independent, based in France, working with teams across Europe, the US and Canada.
Where to start
The fastest way to judge an AI support agent is to ask it the questions your team gets every day. So the first step is a working agent on your own catalogue.
A demo agent on your catalogue, in 72 hours
I build it from what is already public: your product catalogue, your published policies, your help centre. No access to your systems, no credentials, nothing to sign. You open a link and ask it whatever you want.
- Built on your own products and your own published policies
- Public data only, no connection to your systems
- Free, and yours to poke holes in
If it earns it, we connect it for real
It starts with an audit of your actual support: a sample of your tickets, and how your helpdesk is wired, custom fields, triggers, groups, macros and channels. The agent is shaped around what really fills your queue, then connected to your order and delivery data and your tone.
- An audit of your real tickets and helpdesk setup, included
- Connected to your commerce platform and your helpdesk
- Rules on what it must never answer, agreed with your team
- Live on a slice of your chat traffic, then email when it earns it
Nothing to configure on your side. You point me at your helpdesk, we agree the rules together, and your team reviews the first conversations.
Book 15 minutes
Tell me your platform, your volume and the questions that eat your team's day. If a demo makes sense, I will have one ready for you.