Lea
Leafine’s first agent for e-commerce brands.
Meet Lea →An engineering lab · Paris
We build agents that can be held to a job.
Leafine builds AI agents for e-commerce brands — Shopify stores and D2C labels that run customer support, campaigns and ads with a small team. Our first agent, Lea, works today for French merchants. Every action is scoped, approved and recorded.
- Agents shipped
- One
- Built for
- Shopify & D2C brands
- Based in
- Paris · data in the EU
Why we exist
Behind every order there is a relationship. Software should protect it, not quietly spend it.
An agent working inside a store has access to money, customers and reputation. One bad reply, one campaign to the wrong list, and what breaks is trust a brand spent years earning. So we build agents the way a person is onboarded: a job first, then permissions, then a record of what they did. The model works inside those limits. The limits are the product — not a safety feature added later.
Agents
What we’ve shipped.
Agent 02
Agent 02
Verification requiredA designed placeholder only. A real agent must be named or this card removed before launch.
Register
Every capability, and its gate.
We publish what our agents may do, what waits for approval, and what they are not allowed to do at all.
| Capability | Scope | Gate |
|---|---|---|
| Reading | ||
| Read catalogue, orders and stock | shopify | open |
| Read customer conversations | instagram · whatsapp · gmail | open |
| Read campaign and ad performance | meta · email | open |
| Preparing | ||
| Draft a reply to a customer | all channels | open |
| Build a customer segment | shopify | open |
| Draft an email, post or ad creative | email · instagram · meta | open |
| Acting | ||
| Send a reply to a customer | all channels | |
| Publish to the merchant’s account | ||
| Send a campaign to a list | email · whatsapp | |
| Create or raise ad spend | meta ads | |
| Issue a discount code | shopify | |
| Not available to any agent | ||
| Refund, cancel or edit a paid order | shopify | withheld |
| Contact a customer who opted out | all channels | withheld |
| Change pricing or publish a product | shopify | withheld |
| Use merchant data to train a model | all | withheld |
A merchant can loosen a gate, one capability at a time, or leave them all closed. Moving a withheld line takes a release, a changelog entry and notice to every merchant before it ships.
Trust
We answer for the data.
This is where merchants check the commitments without taking an agent’s word for it.
DPA verification required: region, deletion window and sub-processor notice are draft claims.
- Hosting
- European Union, single region
- Model training
- No merchant data, no exceptions
- Retention
- Deleted within 30 days of account closure
- Access scope
- Per-capability, per-merchant, logged
- Sub-processors
- Published list, notice before change
- Documents
- DPA · security measures · privacy policy
Doing due diligence before connecting a store? We’ll send the full security pack the same day. No call required.
Request the packMethod
How we build an agent.
Permission before autonomy
Anything that spends money or reaches a customer stops and waits. The merchant opens it up capability by capability, or never.
capability registry · per-action approval gates · paused-state guards on every spend path
Every run leaves a record
What the agent read, what it decided, where it stopped, who approved it. Readable months later without us in the room.
append-only run ledger · durable suspend and resume · full input and output capture per step
Grounded, or silent
Answers come from the merchant’s catalogue, orders and policies. When the answer isn’t there, the agent escalates instead of improvising.
scoped retrieval per merchant · no generated answers on order state · escalation as a first-class outcome
French first, not translated
The merchant-facing product is written in French by people who speak it. The conventions of French retail are not a locale file.
French-native copy and prompts · EU hosting · GDPR terms drafted before launch
Notes
Engineering notes.
What we ran into, and what we changed. Written for people who build this kind of thing.
Who we are
A small team in Paris.
We’d rather build one agent properly than five convincingly. We work close to the merchants who use our work. If the register above is the kind of thing worth arguing with, we should talk.
Build with Leafine
One careful agent is worth a dozen convincing demos.
If you run an e-commerce team, build agent systems, or care about verifiable AI work, we would like to hear how you see the problem.