Skip to content

The AI agent

An AI that answers from your store, or not at all.

Most support bots are trained to always say something. Suvenna is built the opposite way: it answers from your policies and the customer's real order, and when neither covers the question, it hands off to your team. It never invents a tracking number, a discount, or a policy.

Request early accessSee the full setup

How grounding works

Two sources of truth. Nothing else.

Every answer traces back to something real: a policy you wrote, or an order in your Shopify store. That constraint is the whole design.

  1. 01

    Your policies are the knowledge base

    You paste your shipping, return, and refund policies, and your saved macros. That is the AI's entire world knowledge. It answers from what you wrote, in your terms, and nothing gets pulled from the open internet or from what a model thinks a typical store does.

  2. 02

    The real order, fetched live

    When a customer asks about an order, Suvenna fetches it from Shopify at that moment: status, items, fulfillment, refundable amount. The lookup is bound to the verified customer's email, so the answer is about their order, with current data, not a cached guess.

  3. 03

    Neither covers it? It abstains

    If the question falls outside your policies and the order data, the AI does not improvise. It tells the customer a teammate will follow up and escalates to your inbox with full context: the question, the order, and why it stopped.

The hard part

Abstaining is a feature

Any bot can produce a confident paragraph. The engineering is in knowing when not to.

The improvising bot

A typical AI helpdesk is optimized to respond. Ask it something its knowledge doesn't cover and it fills the gap with something plausible: a return window your store doesn't offer, a discount code that doesn't exist, a delivery date it made up. The customer acts on it, and now your team is cleaning up a promise the bot invented. The better it sounds, the worse the cleanup.

Suvenna

Suvenna treats "I don't know" as a valid, first-class answer. When your policies and the order data don't cover the question, it says a teammate will follow up, escalates with everything it gathered, and stops. Your team answers the hard one; the AI keeps clearing the ones it can prove.
  • Never invents a tracking number, discount, or policy
  • Escalations arrive with the question, the order, and the reason it abstained
  • Every abstention is visible in the inbox, so you can close the policy gap it found

Identity binding

Your order, and only your order

Order lookups are tied to the verified customer, not to whatever order number someone types.

An order number is not identity. Anyone can guess one, or read one off a screenshot. So Suvenna never looks up an order by number alone. Every lookup is bound to a verified email: on the email channel, the mailbox the customer actually wrote from; in the storefront chat widget, a shopper identity that Shopify itself verifies with an HMAC signature before we trust it.

Ask about order #1042 and you get an answer only if #1042 belongs to your verified email. Anyone else asking gets nothing: no status, no address, no items. There is no order snooping, by design, and the same binding governs every action the AI can take on an order.

Earning autonomy

Approve first. Autopilot when it's earned.

Trust is built per store, on your inbox, with your grading.

A journey, not a switch

Every store starts in approve-mode: the AI drafts every reply, and your team approves, edits, or rejects it before anything is sent. Each reply carries per-reply QA, a thumbs rating and notes, so grading the AI is part of working the inbox, not a separate chore. Reports show the QA scores and the AI's share of the work.

  • Approve-mode: AI drafts, your team sends
  • Per-reply QA: thumbs and notes on every AI reply
  • Flip on Autopilot per store when the scores earn it
  • Money actions stay confirm-gated either way, and live actions ship off until you enable them
The Suvenna inbox showing an AI-drafted reply awaiting approval, with per-reply QA controls and the customer's order in the sidebar

Questions

How the AI works, in short

What does the AI actually know?
Two things: the policies and macros you give it, and the verified customer's own Shopify order data, fetched live. It does not answer from the open internet or from general model knowledge. If your policies don't cover a question, it doesn't guess.
What happens when it doesn't know the answer?
It tells the customer a teammate will follow up and escalates the conversation to your inbox with everything it gathered: the question, the order, and why it abstained. Abstaining is a feature, not a failure. A bot that guesses about refunds is a liability.
Can a customer look up someone else's order?
No. Order lookups are bound to the verified customer's email. On email, that's the mailbox they wrote from. In the chat widget, the shopper's identity is HMAC-verified by Shopify. Pasting someone else's order number returns nothing.
Do I have to trust it on day one?
No. Every store starts in approve-mode: the AI drafts, your team approves. You grade replies with per-reply QA and watch the scores in reports. Flip on Autopilot when it has earned it. Live money actions ship off until you enable them, and every one is confirm-gated.

Get priority access

Suvenna onboards in limited waves. Join the list and you're first in line for the next one, with founding-member pricing locked in.

Request early accessSee pricing