Two doors, one idea. Inbox is where your customers' conversations land. Assistant is where you configure the AI that answers them.

The assistant is on for every plan. It is not a paid feature and never has been. The Free line is volume — 150 replies a month — not capability. If you only read one thing, read What happens when a customer messages you: it explains almost every "why didn't my bot reply?" you will ever have.

The inbox

The problem — WhatsApp, Telegram, web chat, voice calls and email, in five different apps, and you're the only one holding them together. Every one of them has its own notification badge, and none of them knows what the others said.

Where — Inbox

What you'll see

One list, newest last-message first, with threads that need you floated to the top — unread follow-ups and anything the assistant handed over. Each row carries a channel-tinted avatar so you can tell a WhatsApp chat from a voice call without reading a word: WhatsApp green, Telegram sky, Website violet, Instagram pink, Messenger blue, Voice call purple, Email amber.

The row itself gives you the customer's name, their country flag, a device chip where the channel reveals one, and a one-line preview prefixed with "You:" or "Bot:" so you know who spoke last. On the right: a relative timestamp (now, 5m, 3h, Yesterday, Wed) and a badge.

The badges

Set it up

1. Open Inbox. Threads needing you float to the top; the rest are most-recent-first.
2. Filter by channel — the pills only show channels you actually use, each with its own count.
3. Tap a row to read and reply. Opening a thread marks it read and clears its count.

Free and Pro

Details worth knowing

• The channel pills count everything, not just what's filtered. Switching to WhatsApp doesn't change the other pills' numbers — otherwise the filter would hide the very picture you opened it to see.
• Silent threads are hidden. Web chat and voice open a conversation the moment someone lands, so sessions where nobody actually said anything would otherwise fill your inbox with ghosts.
• Except when they need you. A thread that opened, said nothing, and escalated still shows — labelled "Opened a chat but didn't send a message", because "Needs you / No messages yet" reads like a bug.
• The list shows your 100 most recent threads.
• The Needs you count rides on the Inbox door from every page, so you don't have to keep checking.

Works with — Reply, take over, hand back, Insights

Reply, take over, hand back

The problem — The AI is handling it, you can see it's going wrong, and you want the wheel. Or the opposite: you answered one question and don't want to own the thread all night.

Where

What you'll see

The customer's name up top with a live status badge — Bot is handling · Needs you · You're handling · Closed — and, where they exist, View customer → and View linked order → shortcuts.

If the assistant sent any pay links in this thread, a Payment links panel shows each one with its click count and whether it's ✓ paid or unpaid. That's the assistant's actual scoreboard, sitting in the conversation where it earned it.

Under that: the conversation. Under that: the reply box.

Set it up

1. Type in the reply box and Send reply — this automatically takes the thread over from the bot. The box says so, under the button.
2. Attach a photo or a PDF with Attach a file. With a file staged, the box becomes an optional caption and the button becomes Send attachment.
3. Use Take over (while the bot has it), Hand back to bot (while you have it), or Close.

Free and Pro

Details worth knowing

• Replies cap at 4000 characters; attachments are images or PDF.
• "Hand back to bot" isn't offered on email threads. There's nothing to hand back to — see Email threads.
• Taking over is permanent until you undo it. A thread you claimed never auto-resumes; only ones the assistant handed over by itself do. See The 30-minute hand-back.
• Timestamps are in your workspace's timezone, not your device's — you might be reading this from another country, but your business doesn't move.

Common problems

Works with — Approve what the assistant proposes

Who am I talking to?

The problem — Replying to a name with no idea whether they're a first-timer or your best customer. The answer changes what you say, and it's sitting in your order history where you can't reach it mid-conversation.

Where

What you'll see

Their avatar and handle — tappable, straight through to Telegram or Instagram — and their phone as a WhatsApp deep link or a tel: dial. Then their country with the flag, and, honestly labelled, where that guess came from: from phone number, from app language, or approx. from network. A location Selza is unsure about says so rather than presenting a guess as a fact.

Below: First seen, the message count, Has ordered / No orders yet, and their 👍/👎 if they rated the chat. Then a collapsible Customer history & AI snapshot:

• Lifetime value, order count, average order, "Last order 3d ago", and "Pays by PayNow"
• A churn-risk pill, their top products with quantities, your tags, and your own notes
• Full customer profile →
• A one-line AI snapshot with Generate / Refresh

Free and ProRefreshing the AI snapshot uses your monthly AI allowance — it's a tool for you, not a customer reply.

Details worth knowing

• The history section starts collapsed unless there's something in it, so the reply box sits higher on a phone. Data is never hidden — only emptiness is.
• Only web chat can reveal a device. Every other channel says "Device not shared" rather than implying we just didn't capture it.
• "Opened chat from…" tells you the page they were on when they started typing.
• Instagram and Messenger avatars expire. A stale one is hidden rather than shown broken.

Works with — The inbox, Insights

Approve what the assistant proposes

The problem — You want the AI to be useful without ever letting it change an order behind your back. Those two wants look incompatible until you separate deciding from doing.

Where — Inbox → (conversation) — the request appears inline, above the messages

What you'll see

A highlighted card titled Bot suggests · refund (or reschedule, cancel…) with a plain-English summary of what it wants, and two buttons: Approve & apply and Reject. Once decided, it drops to a quiet line under the thread — ✓ Applied or ✕ Rejected — so the record survives the decision.

Set it up — Tap Approve & apply or Reject. That's the whole contract: the assistant never changes anything itself. It asks; you decide with one tap.

Free and Pro

What it can ask for — Reschedule · Update contact · Cancel · Add a note · Take bookings · Refund. You choose which of these it's allowed to propose (see Changes it can ask you to approve).

Details worth knowing

• The assistant is never allowed to say it's done. Its reply may only say it's been passed to you — and that's one of the things the double-check explicitly looks for. A customer told "your refund is on its way" by a bot that can't issue refunds is a complaint you'd have to clean up.
• Approving re-validates everything. The proposal is a request, not a stored instruction; the real action re-checks the order from scratch when you tap.
• A proposal can only be decided once. "That suggestion was already decided."

Works with — Refunds

Channels

Where

What you'll see — A card headed "Where your AI answers" with a live "N of 5 live" count, then one row per channel — Web chat · WhatsApp · Telegram · Instagram & Facebook · Voice AI — each with a badge and a one-line status you can act on. Every row jumps to that channel's card further down the page.

Your 24/7 AI assistant link

The problem — You want to be answerable right now, before connecting anything to anything. Every other channel on this page asks you to go and set something up somewhere else first.

Where

What you'll see — Your link in a copy box, and next to it a Selza-branded QR code, generated by Selza itself rather than fetched from some third-party QR service that could go away or start watching your customers. Under it, a line telling you what the link actually does for your shop right now — with WhatsApp connected it reads "On phones, WhatsApp users go straight there; everyone else chats with your AI on the page."

Set it up

1. Copy the link, or download the branded QR code.
2. Put it in your bio, on your receipts, on a card by the till.

Free and ProAlways live — zero setup.

Details worth knowing

• On a phone, WhatsApp wins. If you have WhatsApp connected, visitors with the app are sent straight there — a chat they keep. Everyone else chats in the browser, no app, no install.
• The link follows your shop handle, and renaming your shop doesn't break the old one. Anything you already printed keeps working. That's the whole reason it's safe to print.
• Every order page carries a "Chat about this order" button that opens this same chat, already knowing which order they mean — so nobody has to type an order number at you.

Works with — Web chat

WhatsApp — the shared number

The problem — Being on WhatsApp normally means a Meta account, a Business API, and a week you don't have. Most shops never get past step one, and their customers message them on WhatsApp anyway.

Where

What you'll see — On, with "Selza's shared number · +65 8281 0737 · no setup needed". It is already working. There is no connect button because there is nothing to connect.

Set it up — Nothing. Your shop is already live on Selza's shared number. Your support link opens a WhatsApp chat routed to you.

Free and Pro

How one number serves many shops

WhatsApp keys a chat by our number, so every shop on it shares one thread on your customer's phone. Two things make that work rather than break:

• The shop picker. A returning customer who saved the number and sends a bare "hello" — no link, no code, and the old thread long closed — gets a tappable list of only the shops they've already messaged, newest first, each labelled "Chatted 3 days ago". They can summon it any time by sending shops.
• Quoted replies route perfectly. If a customer uses WhatsApp's reply gesture on one of your messages, it lands with the right shop with certainty rather than a best guess.

Details worth knowing

• The picker is not a directory. It is built from that phone's own history — so it's a handful of rows however many shops share the number, and it can never reveal a shop the customer hasn't already talked to.
• It caps at 10 rows — WhatsApp's own limit. Past that, the last row becomes "Another shop" pointing them at the link you shared, rather than silently dropping the shop they wanted.
• Shops that can't take the chat aren't listed — support switched off, or since moved to their own number. A tap that leads nowhere is worse than a row that isn't there.
• A customer sending STOP unsubscribes from every shop on the number, and is told exactly that. They see one business; one STOP has to be enough.

Your own WhatsApp number

Customer photos

The problem — "Is this the right size?" with a photo attached — and a bot that can't see it. The customer has done the helpful thing and gets a reply that ignores it.

Where — Automatic on WhatsApp

Set it up — Nothing. The assistant reads photos customers send and answers about them.

Free and Pro

Details worth knowing

• Photos only. Video, documents and voice notes are acknowledged and handed straight to you — "Thanks — I've shared that with the team, and someone will follow up with you shortly." Then the chat is yours.
• Very large photos are treated the same way. There's a ceiling on what the vision model will read, and a photo over it goes to you rather than getting a guess. WhatsApp compresses photos well under that ceiling, so this is rare in practice.
• An honest hand-off beats a confident guess. Everything the assistant can't read escalates rather than being answered blind — the same rule that governs everything else it doesn't know.

Works with — When the assistant hands over

Telegram

The problem — Be on Telegram instantly, or under your own brand. Most tools make you choose "set up a bot" before you can find out whether your customers even use it.

Where

What you'll see — Either "Selza's shared bot · @… · or add your own", or, once you've connected one, "Your own bot · @yourbot".

Set it up

• Shared bot — nothing to do. It's already on.
• Your own bot — get a token from @BotFather, paste the whole line in, press Connect. The token is checked against Telegram before it's saved, and the webhook is set up for you.

Free and Pro

Details worth knowing

• Telegram is the only channel that can send broadcasts. WhatsApp requires Meta-approved templates and a 24-hour window; Telegram has neither restriction.
• Your own bot means your own brand in the customer's chat list — worth it if Telegram is where your customers actually live.
• The shared bot has no 24-hour window, so a follow-up can always reach a Telegram customer.

Instagram & Messenger DMs

The row shows Soon, not Off, on purpose: Off would invite you to connect something that then silently does nothing. Nothing is missing on your side and there is nothing you can do to speed it up. When it clears, the card becomes a connect button and this page changes with it.

Email threads

The problem — Every email your shop sends — receipts, status updates, quotes — and every reply, scattered across your personal inbox with no idea which order any of it belongs to.

Where

What you'll see — One thread per customer email address, holding both the automatic mail your shop sent and anything either of you typed. Replying sends a shop-branded email that chains the subject (Re: …) and carries the order's number and a View order link, so your customer sees a conversation rather than a pile of unrelated notifications.

Set it up — Nothing. Their replies land back in the same thread. The reply box says "Sends as a branded email from your shop. The customer can just hit reply."

Free and Pro

Details worth knowing

• This is why there's no "Hand back to bot" on an email thread. There is nothing to hand back to. The thread is marked as yours for life: never given to a bot that wouldn't answer, and never put in the waiting-for-you state that the 30-minute hand-back would collect.
• There's no Channels card for email and nothing to connect. It works because your shop sends mail, which it already does.
• One thread per customer, not per order. Email is how people talk to a shop, not to an order.

Works with — The 30-minute hand-back

Voice — "Call us"

The problem — Some customers would rather talk. Especially the ones who won't type, and the ones whose question takes ninety seconds out loud and nine messages in writing.

Where — Inbox → Channels → Voice AI (status only). Try it at Assistant → Test.

What you'll see — A card reading Voice AI ("Call us"), and, when it's on, "Live on your storefront" with your monthly minute allowance stated in the same sentence. There is no connect step and no phone number to configure.

Set it up — Nothing to provision. With voice on, a Call us button appears on your live storefront automatically.

Free20 minutes/month, 2 minutes per call. Pro — 200 minutes/month, 14 minutes per call. Free voice is live today, not a promise.

Details worth knowing

• Same brain as every other channel. It answers from your catalog and policies, can look up an order, and can send a pay link.
• Near a call's limit it winds down and moves the caller to text chat rather than cutting them off mid-sentence.
• Minutes round up. A 20-second call costs a minute — so 20 short calls is a Free month. The monthly pool is shared across all your callers and resets each month.
• The transcript lands in your Inbox as a Voice call conversation.
• It runs in the browser. There's no number to dial — no carrier sells a voice-capable Singapore number, so the storefront button is the line.
• Not available to property agencies. The voice assistant is a separate stack that never reaches the property grounding or its price and availability checks — a line it could never answer safely is worse than no line.

Works with — Test your assistant

The customer assistant

Where

The page is one long form, top to bottom: the master switch, the basics, what it knows, then advanced sections that start collapsed but open themselves if they already hold your settings — so nothing you configured is ever hidden behind a fold you didn't know to open. Changes need Save changes; until you press it you'll see Unsaved changes next to the button.

Turn the assistant on

The problem — The master switch.

Where

What you'll see — One switch at the top, described plainly: "When off, customers who open your link are told support is unavailable." Not silence — they're told.

Under it, a readiness banner reading your saved tests: "✓ All 4 saved tests passing", or "2/4 tests passing · 2 need attention — your assistant is live, so it's worth a look." With no tests saved it points you at Test it before you go live →.

Set it up — Toggle Turn the assistant on → Save changes.

Free and ProThe AI assistant is not a paid feature — see the two plans.

Worth knowing — Turning it off doesn't shut your inbox. Messages still arrive, still thread, and still wait for you — they just never get an AI reply. Off means you answer, not nobody answers.

Works with — Test your assistant, Your AI usage

Persona — name, first hello, tone

The problem — It sounds like a robot, and your shop doesn't.

Where

What you'll see — Three fields, all optional, and a Use an example button on the ones where staring at an empty box is the actual obstacle. The example is built from your shop name and your bot name, so it's a usable first draft rather than filler.

Set it up

1. Name — optional; blank stays unnamed. Up to 60 characters.
2. First hello — optional; blank uses a friendly default. Use an example fills one in for you. Up to 500 characters.
3. How it sounds — pick a tone: Friendly, Professional, Playful or Concise.

Free and Pro

Details worth knowing

• An unnamed assistant is a real choice, not an unfinished one. Plenty of shops would rather it read as "the shop" than as a character with a name.
• Tone is a chip, not a paragraph. Four options you can judge instantly beats a free-text box where "be professional but warm but not stiff" quietly does nothing.

Works with — How it talks, See how it replies

How it talks

The problem — Fine-tuning the voice without touching what the bot is allowed to do.

Where

Set it up

1. Use emoji — on or off.
2. Always reply in — e.g. "Bahasa Indonesia". Blank mirrors whatever the customer writes in.
3. Wording notes — style guidance, up to 500 characters. "address customers as 'kak', keep replies short, never promise discounts."
4. Sign-off — a closing line, up to 120 characters.

Free and Pro

Details worth knowing

• Wording notes are style only. They can never change what the assistant is allowed to do — you can't talk it into new powers here, and neither can a customer. Everything it may actually do lives in What it can do, behind switches.
• Blank "reply in" is usually right. Mirroring the customer's own language handles the customer who writes to you in Malay and the one who writes in English, which a fixed language can't.

Works with — Persona

What it knows

The problem — A bot that guesses is worse than no bot. One invented delivery time costs you more than a hundred correct answers saved you.

Where

Set it up — Type your policies, FAQ, anything customers ask. Up to 5000 characters.

Free and Pro

The grounding contract, stated in the product — "The assistant answers only from this, your published products, and the customer's order. If something isn't here, it asks you instead of guessing."

Details worth knowing

• Three sources, in order. Store information is preferred over Procedures, which are preferred over this free-text FAQ. Structured beats prose because prose can contradict itself.
• This box is the catch-all, for the things that don't fit a field or a playbook.
• Insights can fill it for you. A top question on Insights has an Add to FAQ button that drops it straight into this editor, already written.

Works with — Store information, Procedures, Improve

Store information

The problem — The same five questions, forever: delivery, returns, warranty, hours, sizing.

Where

What you'll see — A field per question, and a coverage meter showing what's still empty — so "what haven't I told it?" is answerable at a glance instead of by re-reading your own FAQ.

Set it up — Fill what applies, leave the rest blank.

Free and Pro

Details worth knowing

• These are preferred over your free-text FAQ, so the answer to "do you deliver to Jurong?" stays the same every time it's asked — which a paragraph in an FAQ box cannot promise.
• Blank is fine. Nothing here is required, and an empty field just means that question falls through to your FAQ or to you.
• The double-check reads these too. They're the facts a sensitive reply gets measured against, so filling them in makes the double-check sharper as well.

Works with — What it knows

Procedures

The problem — "How do I handle a return?" needs a repeatable playbook, not a paragraph. Written as prose it gets half-followed; written as steps it gets followed.

Where

What you'll see — Your playbooks, with the section already open if you have any. Each has a name, a "when it applies" trigger, and steps.

Set it up

1. Start from a ready-made template, write your own, or paste an existing SOP and let AI shape it.
2. Give it a name, a "when it applies" trigger, and plain steps.
3. Turn it on.

Free and ProWriting them by hand is free of any allowance; AI-shaping a pasted SOP uses your monthly AI allowance — that one's a tool for you.

Details worth knowing

• A procedure guides the assistant's wording, not its powers. It can't let the bot act on its own — it still only proposes changes for you to approve. A procedure that says "issue the refund" gets you a proposed refund, waiting for your tap.
• Procedures outrank your FAQ, so a playbook and a paragraph that disagree resolve in the playbook's favour.
• The preview tells you when one fired — "Following 2 procedures".

Works with — What it knows, See how it replies

Sell & send payment links

The problem — The assistant can answer questions but not close the sale sitting right in front of it. A customer who says "I'll take two" at 1am shouldn't have to wait for you to wake up.

Where

Set it up

1. Toggle Sell & send payment links on (on by default).
2. Set Max order value it can sell on its own, in your shop currency. Empty means no limit.
3. Above that limit, it tells the customer you'll finish the order personally and hands the chat to you — which is usually where a big order converts better anyway.

Free and Pro

Details worth knowing

• It only picks the item and quantity. The price always comes from your catalog. It cannot invent a price, and it is never allowed to compute a total — the figures in the reply are composed by Selza from your own catalogue, not written by the model.
• The link is real and single-use, and behaves like any other pay link at checkout.
• The cap is a judgement call, not a safety rail. The assistant can't misprice a S$2,000 order any more than a S$20 one — the cap is about whether you'd rather be the one to handle it.

Works with — Follow up on unpaid links, Order links

Suggest add-ons & deals

The problem — Basket size. A shop assistant would say "that goes well with the small one" and yours can't.

Where

Set it up — Toggle on. It recommends one or two matching items and mentions an active discount code.

Free and Pro

Details worth knowing

• It never invents items, prices or totals. Everything it names is something you actually sell.
• One or two suggestions, not a catalogue. An upsell that lists eight things is a menu, and a menu is what the customer was avoiding by asking a human.
• Only codes you've marked shareable ever reach a customer this way. A discount the assistant granted itself is deliberately excluded — see Offer a one-time discount.

Works with — Discount codes

Offer a one-time discount

The problem — A customer is hesitating on price at 1am. Your options were: the bot stonewalls them, or the bot freelances a number. Neither is good — one loses the sale, the other loses the margin and sets a precedent you never agreed to.

Where

Set it up

1. Toggle it on. Off by default — with it off, price negotiation comes to you.
2. The toggle alone ships a working default: 10%, no minimum, valid 48 hours.
3. Customize the discount to change the rate (1–50%), a minimum spend, and how long it lasts (1 to 720 hours).

Free and Pro

The guardrails, in the order they apply

• The rate is yours, always. The assistant asks for a discount with a reason; it never names a percent. The rate, the minimum, the expiry and the single use all come from your settings and are applied by Selza — the offer sentence your customer reads is composed by Selza, not written by the model.
• Once per conversation, then once per customer. Identity is tracked beyond the thread, so closing a chat and opening a new one can't farm a second code.
• A daily ceiling per shop backstops the whole thing, so a runaway loop stops at a bounded cost.
• The granting reply is always double-checked before sending — regardless of your review setting.
• The code is single-use, expiring, and re-validated at checkout like any code you typed yourself.
• Every grant is recorded with the assistant's reason, so you can read back why it thought this sale needed saving.
• The code never enters the assistant's own knowledge, so it can't be repeated to anyone else.

Worth knowing — A held reply returns the allowance. If the double-check blocks the granting reply, the code is archived and the customer's once-only chance is given back — they were never shown an offer, so it would be wrong to have spent it.

Follow up on unpaid links

The problem — They asked to buy, got the link, got distracted. One nudge recovers the sale; two nudges lose the customer.

Where

Set it up — On by default. Turn it off here.

Free and Pro

What your customer gets — One message on the same chat: "Hi! Just checking in — your [item] is still waiting whenever you're ready. Here's your link to finish up: …"

Details worth knowing

• Exactly one reminder per link. Ever. It's stamped as sent before it sends, so even a retry can't double-send.
• At least 3 hours after the link, and never more than 24. Sooner is nagging someone who's still deciding; later is a stranger reminding you about something you'd forgotten.
• Never if the thread moved on. If the customer messaged again after the link, the bot has already answered whatever came next — a nudge would be talking past them.
• Never if you took the thread over, never to someone who opted out of marketing, and on WhatsApp only inside its 24-hour window.
• It's a fixed template, not AI. Deterministic, free, and impossible to talk into saying something else. It costs you no allowance.
• This is for links the assistant sent in a chat — the ones with a conversation to follow up in.

Works with — Sell & send payment links

Let customers check an order

The problem — "Where's my order?" at 2am, forever. It's the single most common message any shop gets, and the answer is already on a screen the customer can't see.

Where

Set it up — Toggle on.

Free and Pro

Details worth knowing

• It only shares status after they confirm the email or phone on that order. An order number is not a password — it's short, sequential, and printed on things. Nobody looks up anybody else's order.
• A chat opened from an order page already knows the order, so there's usually nothing to confirm — the link they clicked is the proof.
• It reads the same source your order page does. The assistant cannot quote your customer a different status than you're looking at.

Works with — Order page

Greet browsers proactively

When it's live it will offer a hand when a visitor lingers on a product or looks about to leave with items in their cart — like a shop assistant walking over. It leads with help and never says it was watching, which is the difference between a greeting and a jump-scare.

Where

Free and Pro

Changes it can ask you to approve

The problem — Deciding how much rope the assistant gets. The honest answer differs per shop, so it has to be your call rather than ours.

Where

Set it up — Tick what it may propose: Reschedule · Update contact · Cancel · Add a note · Take bookings · Refund. (The first four are on by default; Refund isn't.)

Free and Pro

Details worth knowing

• Every one of these is propose-only. Ticking Refund doesn't let the assistant refund anyone — it lets it ask. The tap is still yours. See Approve what the assistant proposes.
• Refund is off by default because it's the only one that moves money, and a default that moves money is a default nobody chose.
• Unticking one doesn't stonewall the customer — the assistant hands the chat to you instead.
• The chips are the allow-list. A change the model invents can be asked for and still never execute.

When to involve you

The problem — There are things a bot should never handle alone, and only you know what they are. Your list isn't ours: one shop's routine question is another shop's lawsuit.

Where

Set it up

1. What to tell customers when it hands off — e.g. "We usually reply within 2 hours". Up to 200 characters.
2. Always hand to me — e.g. "complaints, wholesale or bulk orders, partnership requests".
3. Never discuss — e.g. "medical or legal advice, anything about competitors". It declines politely and hands off.

Each has a Use an example button.

Free and Pro

Details worth knowing

• "Always hand to me" and "Never discuss" are different tools. The first says someone should answer this, and it's you. The second says nobody here should answer this at all — it's declined politely, then handed over.
• The response-time note is the one that earns its keep. "We usually reply within 2 hours" turns a hand-off from an unanswered message into an answered one.
• "Never discuss" is enforced twice — the assistant is told to avoid those topics, and the double-check independently checks the drafted reply against the same list before it sends.

Works with — When the assistant hands over

Double-check sensitive replies

The problem — One wrong answer about a refund costs more than every right answer saved you.

Where

What it does — Before a sensitive reply sends, a second AI reads the draft against your store info, procedures, FAQ, catalogue and the linked order, and votes: send it, or don't. If it votes no, the draft is thrown away and the chat comes to you.

Set it up

1. On by default — nothing to do.
2. Choose Refunds, cancellations & payments (the default) or Every reply.

Free and ProThis is not a paid feature and doesn't use your allowance — caution should never be something you can run out of.

What it checks — that every concrete claim is grounded in your facts · that it never says a refund, cancellation or change has been done when it's only been proposed · that it stays off your "never discuss" list · that it promises nothing you didn't offer.

Details worth knowing

• It never rewrites the reply — only send-or-don't. A reviewer that paraphrased could introduce brand-new errors of its own, and then nothing would be checking it.
• Money turns are always checked, whatever mode you picked: a pay link, a proposed change, or a discount code. Even when the chat is already being handed over — your customer still reads that reply.
• A blocked reply's hand-off never auto-resumes. It was held because something was wrong; giving it back to the bot after 30 quiet minutes would just re-run the same mistake.
• It adds a moment's delay on those replies. That's the trade.
• Switching it off also removes the extra hold on payment replies.

When the assistant hands over

The problem — Knowing why a chat suddenly needs you.

It hands over when — you switched it off · it hit your plan's monthly reply cap · your shop went over its hourly burst limit · the reply failed its double-check · the customer's WhatsApp 24-hour window closed · it hit an error · the customer sent media it can't read · the sale was over your max-order value · a topic matched your "always hand to me" or "never discuss" list · or it decided the matter needs a human.

Your customer always gets a friendly holding line. Never a paywall.

Free and ProAlways on; not configurable.

Details worth knowing

• The volume guards run before the model, so a chat that's handed over for volume costs you nothing.
• One chat is capped at 20 bot replies an hour. Over that it sends one notice and goes quiet on that thread until the hour rolls — it won't argue with someone sending fifty messages.
• Your whole shop is capped at 120 bot replies an hour. That's a burst guard, not a plan limit; it protects against every chat firing at once, and it's sized so a genuinely busy day doesn't trip it.
• Nothing here loses a message. Every one of these paths stores the customer's message and puts it in your inbox first. The worst case is that you answer it.

Works with — The 30-minute hand-back, Your AI usage

The 30-minute hand-back

The problem — A chat escalated to you at 2am and you're asleep. The customer shouldn't wait until morning for a question the bot could have answered in a second.

Where — Automatic. A note in the thread tells you it happened.

Set it up — Nothing. If nobody replies within 30 minutes, the assistant steps back in on the customer's next message and leaves a note in the thread saying so.

Free and Pro

Details worth knowing

• It's reactive, never a push. The resume rides in on the customer's next message. A quiet thread stays quiet — nobody is messaged out of the blue at 3am, and on WhatsApp the 24-hour window would block it anyway.
• The clock starts at the hand-off note. For a chat you never answered, that note is the last thing anyone said — so the 30 minutes measures exactly how long your customer has been waiting.
• A thread you took over never comes back. This only collects chats the assistant handed over and nobody picked up. Press Take over and it's yours until you press Hand back to bot.
• If you reply first, the resume quietly does nothing. The bot can never talk over a human who just engaged.

Email is never handed back either — email is human-only, and was never the bot's to begin with.

Works with — Reply, take over, hand back, Email threads

Test, improve, insights

See how it replies

The problem — You've just changed the tone, added an FAQ line, and tightened a policy. Finding out whether that helped by watching a real customer is an expensive way to learn.

Where — Assistant → Assistant → the See how it replies panel, beside the form

What you'll see — A sticky panel that follows you down the page, with four one-tap questions — "Where's my order?" · "Can I get a refund?" · "Do you deliver to me?" · "This is taking forever 😤" — or a box to type your own.

The answer comes back as a chat bubble, plus what it would have done:

• 💳 Would send a pay link: 2× Chocolate Cake
• 📝 Would ask you to approve: refund
• Would look up order #1032 · Would suggest: … · Would mention offer …
• Would hand off to you — (and why)
• Following 2 procedures · Knows: Delivery, Returns · the voice it's using

Free and ProPreviews use your AI allowance and are hourly rate-limited — they're a tool for you, not a customer conversation.

Details worth knowing

• It runs against your UNSAVED edits. That's the entire point: you see the change before you commit it. The footer says "Previews use your current edits — Save to make them live."
• Nothing is sent and nothing is saved.
• Same engine as the real thing. The preview isn't a simulation of your assistant — it's your assistant, with the send switched off.

Works with — Test your assistant

Test your assistant

The problem — Finding out how your assistant answers by watching it answer a real customer badly.

Where

What you'll see — A fuller version of the inline preview, with the thing the panel can't give you: what it can see — the actual grounding it's working from, laid out. Plus the option to test against one of your real orders, so you can ask "where's my order?" as a customer who has one.

Set it up

1. Type a message.
2. See the reply, what it would do, and what it can see.
3. Optionally test against one of your real orders.

Free and ProNothing is sent to a customer and nothing is saved.

Details worth knowing

• You can talk to it by voice here before customers can — the same voice assistant your storefront button opens.
• "What it can see" is the debugging tool. When a reply is wrong, it's almost always because the grounding is missing something, not because the model is being stupid. This shows you which.
• Voice testing isn't offered to property agencies, matching the storefront.

Works with — See how it replies, Saved tests & regression suite

Saved tests & regression suite

The problem — You fix one thing and quietly break another. Six FAQ edits later, the answer that used to work doesn't, and nothing told you.

Where

Set it up — Save a scenario with the outcome you expect, then run the suite. An AI judge grades each run: did it address what the customer wanted, or correctly hand off?

Free and ProRuns use your allowance.

Details worth knowing

• "Correctly handed off" counts as a pass. A test suite that only accepted answers would train your assistant to answer things it shouldn't.
• Your results ride back to the Assistant page as the readiness banner, so a broken suite finds you rather than waiting to be checked.
• It's tucked behind a fold on purpose. Casual testing — send a message, read the reply — is the common case and stays the default view.

Works with — Improve, Turn the assistant on

Improve

The problem — Your assistant's gaps are buried in a hundred real conversations you'll never re-read.

Where

Set it up

1. Press Scan recent chats.
2. You get a short, ranked list of gaps — a missing answer, an empty store-info field, an action it isn't allowed to take — each linking straight to its fix.
3. Triage: Start · Mark done · Dismiss.

Free and ProScanning uses your allowance.

Details worth knowing

• It never piles up duplicates, never re-raises something you dismissed, and only re-opens a "done" gap if it actually recurs. A to-do list that regrows overnight is one you stop opening.
• Reading your chats changes nothing in them. The scan is read-only and proposal-only; no customer is messaged and no conversation is touched.
• Turn a bad chat into a test. Pick a conversation that went wrong and save it as a scenario. It becomes a permanent regression test, run by the same engine that serves production — so the thing that embarrassed you once is now the thing that can't happen again.

Works with — Saved tests & regression suite, Insights

Insights

The problem — Is the bot actually helping, and what do people keep asking?

Where

What you get

• Bot deflection — chats handled without you, as a percentage and a count.
• Rated helpful — real 👍/👎 from customers. Until you have ratings this slot shows Escalated to you instead, because an empty card teaches nothing.
• Actions approved — of how many proposed, and how many you rejected.
• Est. time saved — reckoned at ~3 minutes per deflected chat, and labelled an estimate.
• Top questions, each with an Add to FAQ button that drops it into your assistant's knowledge.
• Where your assistant was stopped — every time a guardrail kept a wrong reply, sale or change from reaching a customer, counted by kind. Frequent entries point at something to tune: a missing FAQ fact, an expired offer.
• Chat volume by day (Singapore time), with escalations shaded.

Free and Pro

Details worth knowing

• Nothing stopped? It says "the assistant stayed inside its rails" rather than inventing a problem.
• The honesty footer. Under the guardrails card, a line compares sales your assistant closed against how many were later refunded, cancelled or disputed — next to your shop-wide rate. It's a mis-selling check, and it's there to be able to tell you bad news.
• "Where your assistant was stopped" is a feature, not a fault log. Every entry is the guardrails doing their job. A number that's rising is a prompt to fix the cause, not a bug report.

Works with — Improve, What it knows

Your AI usage

The problem — "Why did my bot stop answering?"

Where

Free150 customer replies/month. Pro — effectively unlimited.

What happens at the cap — The assistant hands new chats to your inbox with a friendly holding line and you're nudged to upgrade. Your customers never see a paywall. You're warned at 80%.

Details worth knowing

• AI quotes and the customer assistant never run out mid-conversation — they're excluded from the credit budget entirely. The allowances gate your tools, not your customers' conversations.
• The 150 is replies, not conversations. A ten-message chat is ten.
• Your buyer is never told your plan ran out. They get "Someone will be with you shortly" and a real human — your billing is not their problem, and a paywall in a customer's chat window would cost you the sale the assistant just earned.
• Hitting the Free cap isn't an error and doesn't alert anyone. It's the plan working as designed.
• Paid plans have a ceiling too — around 10,000 replies a month. It is an abuse valve, not billing: no real solopreneur reaches it, so tripping it means something's wrong (a bot sweep, a viral flood) and we look at it rather than charging you.
• Your tools are metered separately — the co-pilot, previews, scans, AI snapshots. Reaching Selza support is not.

Works with — Plan & billing, When the assistant hands over

The co-pilot

Your private helper

The problem — You need help with the console, not with a customer.

Where — The ✦ button, bottom-right of every page

Free and ProDeliberately not gated — it's your only in-app route to a human at Selza, and a plan gate would mean a Free merchant couldn't reach us at all.

What it can do for you — Ask it anything about running your shop. It can create a discount code, send a support ticket to Selza, and file a feature request — each as a proposal you approve with one tap.

Details worth knowing

• Those three are the whole list. Not a summary of a longer one: an action the model invents can be proposed and still never execute.
• It re-checks everything at the moment you tap, not when it proposed — so a co-pilot switched off in between can't slip one through.
• It's a different AI from the one answering your customers, with different knowledge and different powers. The ✦ button knows your console; the Assistant knows your shop.

Works with — Get help from a real person

Get help from a real person

The problem — Something's wrong and you need a human, not a bot.

Where — The co-pilot (✦)

Set it up

1. Describe the problem in your own words.
2. It offers to send it to Selza. Review the draft and tap Send to Selza.

Free and ProThis never uses your AI allowance, even when everything else is spent — you must never be unable to reach us because you ran out of AI.

Details worth knowing

• When your credit is gone, the door doesn't close — it changes shape. The co-pilot can't draft for you any more, so the panel gives you a plain form instead. Same ticket, same queue, no AI involved.
• You approve what we see. The model drafts; nothing reaches Selza that you didn't read and tap send on.

Suggest a feature

The problem — Selza can't do the thing you need, and you've no idea how to tell anyone.

Where — The co-pilot (✦) — say "I wish Selza could…"

Set it up — It first checks it isn't already possible, then drafts the request and names the area. Review it and Add to roadmap.

Free and ProNever uses your AI allowance.

Details worth knowing

• It checks whether it already exists first. Often the answer is "it does, here's where" — which is a better outcome than a roadmap row.
• Asking twice doesn't file twice. The same ask from your shop lands once; you're told it's on the board rather than getting a duplicate.
• Requests are counted by distinct shop, so a hundred merchants asking for one thing reads as a hundred — which is exactly what makes it get built sooner. Asking again doesn't add weight; asking once does.