Kivade Docs

Branding & hostnames

Overview

Settings groups your organization's configuration into sections, each with its own web address you can bookmark or share:

You need the Manage roles permission (owners have it); members without it do not see the Settings entry at all.

Organization branding

Three fields are shown on your landing page:

  • Organization name — also shown in the workspace header.
  • Logo URL — the web address of your logo image, starting with https://.
  • WhatsApp number — the default number used for the landing page's WhatsApp button.

Fill in the fields and click Save branding.

Hostnames

A hostname is the web address where your landing page lives (for exampleyourname.pages.dev). Leads submitted from a verified hostname are routed to your organization, and your branding is served there. Manage them underWorkspace → Hostnames.

Add a hostname

  1. Enter the hostname in the field and click Add hostname.
  2. It appears with an Unverified status and a verification token.
  3. At your DNS provider, add a TXT record for that hostname whose value is the token.
  4. Click Verify. If the record has not propagated yet, wait a little and click it again.

Once verified, the hostname routes leads and branding to your organization. One hostname maps to one organization — a hostname already claimed by another organization is rejected.

Remove a hostname

Click Remove on a hostname row. Leads from that address will no longer route to your organization.

WhatsApp

Pair your number underChannels → WhatsApp to chat with leads from the workspace, receive every reply on the lead's detail page, and send the automatic confirmation from your own number.

  1. Click Connect WhatsApp.
  2. Open WhatsApp on your phone → Settings →Linked devices → Link a device.
  3. Scan the QR code shown on this page. The page updates once the device is linked.

The connected number is shown on the card. Click Disconnectto unlink the number; leads and message history are kept.

New contacts

Choose how unknown senders on your connected number are handled:

  • Review first (default) — messages from people who are not already leads land under Needs review; you save or dismiss each one. Best for a personal number.
  • Add automatically — every new sender becomes a lead immediately. Best for a number used only for business.

Important: this links your number to the platform through a WhatsApp Web session. Use a number you are comfortable linking, and keep the phone online so messages flow.

AI replies

The AI replies section controls AI assistance for your whole organization, across WhatsApp and email:

  • Off (default) — no AI replies are sent. New leads get the fixed confirmation message, exactly as before.
  • Auto — AI answers incoming messages from confirmed leads on your connected channels. New WhatsApp leads also get an AI-written first reply when your provider allows free-form messages (the paired-number connection does); otherwise the fixed confirmation message is sent instead.

Auto replies are capped per lead per day, and AI never answers contacts still waiting under Needs review. Independently of this setting, the AI draft button on a lead's page generates a suggested reply you can edit and send yourself.

Incoming media is part of the conversation the AI reads: photos are understood directly, and voice notes are transcribed when your deployment has transcription configured (ask your operator — it is a deployment setting, not an organization one). When transcription is not configured, the agent knows a voice note arrived but cannot hear it, and a mistake in reading an attachment never blocks or delays a reply.

Before anything is sent automatically, a second pass checks the draft against your product knowledge and your learned rules. If it finds a claim the knowledge does not support, or a rule the reply breaks, the reply is held as a draft instead of sent and anAI escalation: note records the reason — open the lead to review, edit, and send it yourself. If that check cannot run, the reply is sent as normal; a broken checker never silences the agent. Follow-up steps that fail the check are skipped and never replaced by the fixed template.

Turn on Review AI replies before sending if you want the agent to draft but hold every automated reply: nothing goes out until you open the lead and press Use draft (then edit and send). The trade-off is speed — held replies wait for you.

On a lead's page, a suggested or held draft appears as a card withUse draft, Re-draft (tell the assistant what to change), Remember this (turn your comment into a standing rule), and Discard (with an optional reason). Nothing is sent until you press Send. Sent AI replies have a Feedback action: mark what was wrong, optionally add what you would have said, and the assistant learns from it without anything being sent to the lead.

Teaching the assistant

The What the AI learned from you card underLearned is where your corrections become habits. Two kinds of learning live there:

  • Rules — short instructions such as "keep replies under three sentences" or "never promise discounts". Add them by hand in the card, or turn a draft comment into a rule withRemember this in the Inbox.
  • Voice samples — real replies you paste, plus the replies you send yourself, which are collected automatically. The assistant uses the most relevant samples as a tone reference, never as facts.

Rules can be Active, Proposed, or Archived. Rules learned from a comment you explicitly remembered, or from a reply you sent yourself, become active straight away; rules learned from discarding or marking a reply are proposed until you activate them. You can edit, activate, archive, or unarchive any rule — nothing is ever deleted, so provenance stays visible. Give a rule a channel scope if it only applies to WhatsApp or email.

When a draft is generated, Grounded in lists the knowledge entries, rules, and voice samples that shaped it, so you can always see why the assistant wrote what it did.

Practicing with the assistant

Practice is a rehearsal room for the assistant. You play the lead, the assistant answers with your organization's knowledge, rules, and voice samples, and nothing is sent to anyone. It works before your first lead arrives, and every member can use it. Open it fromPractice in the sidebar on desktop or the bottom navigation on a phone, or go straight toPractice.

Start a session with a title and a channel. The channel decides which rules and voice samples apply: an email session uses email rules and rules that apply to every channel, never WhatsApp-only rules. Then type a message as the lead and press Send as lead. The session header and every assistant reply carry a Practice badge, so a rehearsal is never mistaken for a real thread.

Each session begins from a fresh virtual lead: no lead record, no memory, and no earlier turns. Sessions belong to you inside your organization, and you can resume or delete them at any time. A colleague cannot see or open your sessions, and deleting one keeps the rules and suggestions it produced.

Every assistant reply offers four actions:

  • Good reply marks the turn as liked.
  • Discard marks it as discarded. An optional reason is distilled like any other correction.
  • Re-draft takes a comment as the instruction for the next reply. The old reply stays visible, marked as replaced, with your comment beside it, so the revision trail is on screen.
  • Remember this turns your comment into a standing rule right away.

Corrections from practice learn the same way as corrections in the Inbox. A remembered comment becomes an active rule sourced from practice. A re-draft comment you never remember stays proposed: it appears in the Learned list and grounds nothing until you activate it. Business facts arrive as proposed entries at the top ofAI product knowledge. Nothing role-played can reach a lead's memory.

Grounded in under a reply names the knowledge entries, rules, and voice samples it used. An Advisory check line appears when the claim checker finds something the knowledge does not support; it never blocks or changes the reply.

Practice writes only to its own session store. It creates no messages, leads, notes, memory, drafts, follow-ups, or credit ledger rows, sends nothing on any channel, and consumes no per-lead daily reply cap. Practice turns do not use credits, and a daily cap per organization (100 turns by default) applies. Practice generations are excluded from theWhat the AI is learning counts.

Checking a rule before you turn it on

In the Measurements area ofLearned, your own settled replies — the ones you actually sent — are replayed with today's rules and with one change, so you can see whether a rule helps before you activate it.

  • Measure on a proposed rule asks "what if I turn this on?"; on an active rule it opens an inline editor so you can measure a reworded version or measure removing the rule.
  • Run a check measures today's rules alone and reports how much of your voice the assistant reproduces.
  • Each case shows what you sent, what the assistant writes without the change, and what it writes with it, plus a similarity score and whether the claim check flagged anything. The summary is one ofImproved, Regressed, orNo measurable change.

Measurement never sends anything, never changes a rule, and uses no credits. A run walks through its cases one at a time — if you close the tab partway, open the history and press Continue to finish it. Runs are capped per day, and a check needs a handful of settled replies first.

AI actions

While it works, the AI can also do things, not just reply. These actions run during automatic replies and follow-ups; theAI draft button only reads — it never records facts, changes stages, or adds notes.

  • Record facts — preferences, budget, family, past purchases — into the lead's What the AI remembers list, where you can edit or delete them.
  • Move the lead's stage along the pipeline when the conversation clearly justifies it.
  • Ask for a human — adds an internal note starting withAI escalation: so you can take over.
  • Share your booking link when a lead is ready to talk. Add your Cal.com, Google appointment page, or any https://booking URL in theAI actionssection. If it is empty, the AI says it will check with the team instead of inventing a link.

Follow-ups

Follow-upssend a short automatic nudge to new leads who don't reply — a fixed sequence (by default 24 hours, 3 days, and 7 days after capture), off until you turn it on.

  • A reply on any channel cancels the remaining follow-ups, and so does blocking the lead or moving it to Converted orLost.
  • Messages send only during business hours (9:00–21:00 Malaysia time) and are capped per day, so nobody gets spammed overnight.
  • Turning the switch off cancels every step that is still scheduled, and turning it back on does not bring them back — only later captures start new sequences.
  • Follow-ups use AI when your plan allows it and the channel supports free-form messages; otherwise a fixed message is sent. Each AI follow-up uses credits from your balance; the fixed fallback is free.
  • You can see and cancel scheduled follow-ups on the lead's page.

Notifications

Turn on phone or desktop notifications for the device you are using underNotifications. You get a notification when a lead messages you, when an AI draft is held for review, and when a new contact waits in the Review lane — even when the workspace is closed. Notifications name the lead and channel, never the message text, so a lock screen gives nothing away.

  • Enable notifications on this device asks your browser for permission, then registers the device. Each device is enabled separately; turning one off does not affect the others.
  • Send test notification confirms the current device receives them.
  • Turn off this device removes its registration, and no further notifications are delivered to it.
  • Opening a lead clears its notifications, and the red count on the Inbox shows how many are still unread.

On iPhone

Notifications only work for a workspace added to the Home Screen first (iOS 16.4 or newer). Step-by-step install and enablement for both platforms:Install the workspace on your phone. Android and desktop browsers need no install.

AI credits

Every AI-generated reply or draft uses credits from your organization's balance — by default one credit per generation.AI credits shows your plan, the current balance, the monthly grant, and the current period.

  • Credits never expire and unused credits carry over to the next month.
  • Your plan's grant is added automatically at the start of each month.
  • When the balance reaches zero, AI stops: new leads still get the fixed confirmation message, manual replies and every other feature keep working, and drafts explain that the balance is exhausted.
  • Internal steps (checking or shortening a draft) and background learning work (remembering facts, finding repeated questions) never use credits.
  • A generation that succeeds is charged even if the message then fails or ends unconfirmed — the credit covers the AI work, not the delivery. An automatic reply that is interrupted after generating resumes and sends the same result without a second charge. A generation that fails before producing a reply is not charged.

Buy or upgrade your plan and top up credits from the AI creditscard — payment happens in a secure checkout and your balance updates as soon as payment clears. On a paid plan you can also manage or cancel your subscription from the same card. If billing isn't enabled for your deployment, plan changes and top-ups are handled by the team; contact us and we'll adjust your balance.

AI product knowledge

AI product knowledgeteaches your AI assistant what you sell. Add one entry per product, service, or policy: a short title (the product name works well) and the facts customers ask about — what it is, who it's for, price, availability, shipping, policies.

  • Replies and drafts are grounded in the entries most relevant to the customer's message, not the whole list, so a large catalogue stays accurate.
  • Write entries as reference notes. The assistant is instructed to quote only prices, stock, and availability the entries state, and never to invent quantities or promises — if a detail isn't covered, it says it will check with the team.
  • Limits are shown in the card (number of entries and characters per entry). Long sections are split into multiple entries automatically, so importing a long document loses nothing. Changes apply to the next generated reply.
  • A legacy knowledge text field remains available and is used only while an organization has no entries — once you add entries, they take over.

Use Test it to ask a question a customer might ask and see the grounded answer immediately. Nothing is sent to any customer and nothing is stored.

When an AI draft is generated for a lead, the composer showsGrounded in with the entries, rules, and voice samples used to write it — if the answer is wrong, that is the entry to fix.

When feedback or a conversation reveals a missing or wrong business fact, the assistant proposes it at the top of this card as aproposal. Approve it to add it to your knowledge, or dismiss it. Proposals never ground replies until you approve them.

A few shortcuts save typing:

  • Find repeated questions reads your recent conversations and proposes entries for the questions customers keep asking. Review each one and approve what is right — nothing is saved until you approve, and questions already covered by an entry are skipped.
  • Import text splits a pasted catalogue or policy document into entry drafts (one per heading or paragraph; a section longer than one entry becomes several drafts). Edit any draft title, then approve drafts one by one or all at once.
  • Upload a file (PDF, Word, Excel, PowerPoint, txt, md, csv) reads the text in your browser — nothing is uploaded by the import step — and fills the import box for you. WhatsApp chat exports work too: timestamps and names are stripped before splitting. PDFs without a text layer (scanned documents) are not supported through the file door.
  • Fetch page reads a public product or policy page from its URL, server-side, and fills the import box with the page's readable text. Pages that render entirely with JavaScript may return little text; nothing is stored until you approve a draft.
  • Upload an image (PNG, JPEG, WebP) of a price list, catalogue, or screenshot: the image is sent to your deployment's configured AI provider to be read and its text becomes drafts. The import step does not store the image. This is the only file type that leaves your device for reading.

Files in your library

You can attach a file to any knowledge entry with Attach file(images, PDF, Word, Excel, PowerPoint, txt, md, csv, or a short MP4 video — up to 25 MB). Attached files are stored on the workspace so you can send them to a lead without hunting for them again. The AI still reads only the entry's text; attaching a file does not send it to the AI provider.

  • An entry can be just a file — leave the text empty for pure send material like a brochure, price sheet photo, or demo video.
  • To send one, open a conversation, choose Attach file in the composer, pick the file, add an optional message, and send. Media sends work on WhatsApp and email.
  • Sends are capped before anything leaves the workspace: at most 5 files per message, 16 MB per file, and 24 MB combined. Email attachments are also checked against the provider's 40 MB encoded limit. Over-limit sends are rejected up front; a file the channel itself refuses fails on the message with the provider's error.
  • Files are served from an unguessable private link — anyone who has the link can open the file, so keep internal-only documents out of the library.
  • Deleting an entry deletes its stored file once no other entry uses it.

What the AI is learning

What the AI is learningshows how your AI drafts are doing over the last 30 days:

  • Draft acceptance — sending a draft as-is counts as accepted; sending your own edited version counts as edited. A low rate usually means the knowledge entries are missing something.
  • Generations — how many AI answers were generated in the last 7 days, and how many failed.
  • Frequent questions — questions customers keep asking, marked covered or not covered by your knowledge. Not covered? Add an entry or run Find repeated questions.
  • Last measurement — the result of your most recent check: the reproduction rate for a baseline check, or the verdict for a measured rule.