The questions merchants actually ask.

A customer-facing AI that sells and supports on your storefront 24/7, plus a back office of specialist agents that work on your data — every consequential action confirm-gated and logged. Front of house, back office, and a memory that compounds.

It's a Shopify app. The dashboard lives inside your Shopify admin (or as a standalone app in its own tab), and the storefront chat ships as a theme app embed — you toggle it on in the theme editor, no code edits, no theme surgery.

No. Discount negotiation runs inside limits you set, and consequential actions — refunds, order changes, emails — are confirm-gated: the AI proposes, you approve. Everything lands in an audit trail.

Yes, deliberately. Independence is a dial you set per area of the store — fifteen of them, from discounts to triggers to FAQ — and every one of them ships at “confirm”, so a store that never opens the panel is never surprised. When you do raise one, it runs inside caps you set: what a single action may be worth, a monthly ceiling, how many actions a day, a discount limit, and quiet hours in your timezone. Refunds, customer emails, return decisions, support tickets and the security-sensitive settings can never run unattended at any level.

Yes — Excel, CSV, PDF or a photo of the page. Lumnya reads the lines, matches the SKUs against your own catalog rather than trusting what it read, and shows you every cost change side by side with what's there today before a single one is applied. A line whose SKU isn't in your store is named, not quietly skipped, and a price it can't read unambiguously — is “1,234” twelve hundred or one-point-two? — it refuses to guess and asks you. Your yes applies exactly the figures you were shown, and the whole batch undoes in one click. The same works for a returns policy: it drafts the FAQ answers and the return rules to match. One rule holds throughout: text inside a file is information, never an instruction — a document that says “refund this order” gets flagged to you, not obeyed.

Usually in one click. Every write lands in an action journal with what changed and how it was authorized, and reversible actions store their own inverse — undo replays the same guarded path the original change used, rather than editing your data behind the product's back. Actions that genuinely can't be reversed say so instead of showing a button that would lie. And you can ask Lumnya to simulate first: it runs the real decision path against your real settings and names the areas it could act in on its own versus the ones that would still come to you — changing nothing.

Urgent events email you immediately; everything else arrives as one daily digest rather than a stream of pings. Web push is opt-in per device and carries only a title, a preview and a link — never order values or customer data. Both lanes hold through quiet hours in your store's timezone, and a notification that has gone stale is dropped rather than sent late — it still waits for you in the app. Every message is labeled as coming from AI, per the EU AI Act.

Your store. Alerts, newsletter confirmations, cart reminders, return updates and the one-off note you send from chat all go out under your store's own name, and a reply reaches the admin email in your Lumnya Settings rather than us. Leave that field empty and no email promises a reply it can't deliver: it points the customer at your storefront instead. Mail Lumnya sends you is signed by Lumnya, so you can always tell which of the two you're reading.

Yes — that's the point of a manager. Every hour Lumnya checks whether orders have stopped coming in, whether the storefront widget is failing your customers, and whether people asked for a human and got nobody, and it messages you when one of those breaks. It compares against your own normal for that hour of the day, so a quiet Tuesday is never mistaken for a broken checkout, and one problem is one message rather than one an hour. It won't guess a cause it can't measure, and it changes nothing on its own.

It says so and escalates to you, instead of improvising. You'll see the conversation flagged in Live Chats and can jump in as a human at any point.

It waits for them to finish. “Do you have” / “boards for” / “beginners” is one question, so the assistant holds while they are still writing — the chat says it is waiting for them, not that it is thinking — and then answers all of it in one reply. If they start typing after the answer has already begun, the reply being written picks their words up mid-flight; anything it couldn't fold in is answered immediately afterwards as one follow-up, never one reply per line. Their message appears in the chat the moment they send it, either way. Up to five messages held at once; past that the text simply stays in the box rather than pretending to send. In Live Chats you read it the way it was meant: the pieces the assistant answered together are shown as the one question they were, with every word in the order the shopper spoke it.

The assistant answers in your customer's language, with the widget UI currently shipping in English, Greek and German — and more on the roadmap.

No. It's one app-embed toggle in your theme editor. You pick the accent color, the bot's name and the language; the widget takes care of the rest on desktop and mobile.

Right now, nothing. Lumnya is in private beta: the whole product, free, with no monthly conversation limit and no card — a paid subscription cannot even be started while the beta runs. Announced launch tiers start at $59/mo, and beta merchants keep a founding rate — see the pricing page for the full picture.

Cancel anytime from the dashboard — your store keeps running exactly as before. If you uninstall the app, Shopify's data-removal webhooks kick in and your shop's data is redacted on schedule.

No. Pricing is per connected store. Your whole team can use the dashboard at no extra cost.

Your store's data lives in an isolated, per-store space — strict tenant isolation is an architectural invariant, not a policy promise. It's processed to run the product, never sold. The plain-language version is on the security page; the formal one in the privacy policy.

No. A conversation is bound to the browser that started it: the id alone opens nothing, and reading or continuing a chat needs a key that stays in that shopper's browser and never appears in a URL. You, in your admin, can of course see every conversation in your own store.

Yes. GDPR requests — export and erasure — are honored end to end, including chat transcripts and uploaded attachments. Uninstalling triggers redaction automatically.

Only if the shopper asks for it, and only after they confirm. Asking sends them a one-time confirmation link and records nothing but a pending request; their click is what subscribes them, and it lands in your Shopify customers as a confirmed opt-in. We keep the consent evidence — when they asked, from where — and every email carries a one-click unsubscribe that also stops your cart reminders. That includes the confirmation email itself: anyone can type a stranger's address into a chat, so it offers a one-click “don't email this address again” that also makes the confirmation link inside it stop working. If you'd rather it never came up, one switch in Settings turns it off and the assistant says so plainly.

No. Conversations are processed to answer your customers, not to train foundation models. Model providers receive only what a given reply needs, under data-processing agreements.

We publish our promise in the app: a reply within 24 hours, usually within one. Tickets, screenshots and replies all live in the dashboard's support center.