Your business is yours.
We just keep it honest.

i.What we hold, and why

To run your business we process your orders, customers, parcels, ad spend, WhatsApp conversations and courier settlements — synced from the platforms you connect (Shopify, Meta, couriers, WhatsApp Business). We hold them for one reason: to compute the truth about your business and act only where you've allowed it. We do not sell data, rent it, or train shared models on it.

ii.Isolation — your brand is a walled room, not a row in a table

• Per-client database isolation. Each tenant's data lives behind its own isolation rails. The build itself is checked: a deploy that could leak across tenants fails to ship — the wall is enforced by the pipeline, not by promises.
• Every AI answer is scoped before it speaks. The system resolves your tenant, then answers — and echoes your brand's name into the answer, so an operator running several brands physically cannot be told the wrong brand's number without seeing the wrong name attached.
• Our own brand plays by the same rule. Elyscents is a tenant behind the same walls. We eat what we serve.

iii.What the AI can and cannot do

• It cannot write. Not by policy — by privilege. Every AI query against your business data runs inside a read-only database transaction. The credential the AI uses does not have the power to change your data, so a bad answer can never become a bad write.
• PII redaction is on by default for client tokens. Where a token doesn't need names and phones to answer a question, it doesn't get them.
• Customer content is fenced as untrusted. Anything a customer writes — a WhatsApp message, a review — is wrapped in an ⟦UNTRUSTED⟧ fence before the model ever sees it, so a message can't instruct the AI to do something you didn't.
• Sends are earned, never assumed. Agents that can message customers graduate market-by-market through shadow mode and test phones, under a shared frequency cap — and the safety net (SENTRY) has no send path in the code at all today.

iv.The audit log includes refusals

Every AI call your business answers — and every one it refuses — lands in an audit log you can read yourself: ok, error, refused, rate-limited, with the parameters trimmed but reconstructible. A log of only successful calls answers "what did it do". Yours also answers "what did it try" — and the second question is the one that matters.

v.Credentials and access

• Tokens are stored only as SHA-256 hashes. If our database were read, your tokens would not be in it.
• Human access is role-scoped. A rider, a packer, an agent and an owner see different numbers — the morning brief pushes each person only what their role is allowed to see.
• One revocation story. API tokens can be revoked from your settings, and revocation actually ends access — there aren't two token systems that disagree.

vi.Your customers' data

Your customers remain your customers. Their messages are used to serve them — confirmations, recoveries, support — and, where you've enabled it, to mine phrasing for creative. Our mining pipeline re-verifies every customer quote as a literal substring of the real message and deletes what it cannot prove; the rejected count is logged, not hidden.

vii.What we don't do

• No selling or renting of data — yours or your customers'.
• No blended benchmarks that leak one client's economics into another's screen.
• No dark patterns around deletion: ask, and your tenant's data is exported or destroyed.

viii.Talk to a human

Privacy questions, export or deletion requests, or a data-processing agreement for your legal team: [email protected]. A person answers — the same people who run Elyscents on this system every day.