Reviewing and editing AI responses

Read what your Clone has been saying

Open Messages in the left sidebar (under Interactions). Every conversation your Clones hold, on every connected channel, lands in this one inbox. It's a two-pane view: on the left, a searchable list with one row per contact; on the right, the selected conversation's transcript — until you pick one it simply reads Select a conversation to view.

Inside a transcript you'll see the contact's name with their channel's icon, the inbound messages and your Clone's replies with timestamps (replies a workflow sends don't appear there today), and a header reading IcloneU is handling this conversation while the Clone is on duty. If the Clone has a greeting turned on (Identity tab → Greeting), the reply that opens a conversation can start with that greeting, exactly as you wrote it, with the Clone's answer below it in the same message; when the customer only says hello, the greeting goes out on its own. Each person who writes in is also added to your Contacts automatically.

Filters keep a busy inbox manageable: a conversation-type dropdown — All conversations, Handover requests (conversations escalated for human attention), All pending assignment (handovers still waiting for an agent to take or accept them), All active handovers (handovers any agent has accepted), My active handovers and AI only — plus a date-range filter and an account filter.

Customers can send files too. A photo, document or video sent on WhatsApp — or the first photo, file or video of a message in a Facebook or Instagram DM — lands in the conversation: a photo shows inline, anything else as a file card, and a caption as a message of its own. Storing the file uses units. Whether your Clone knows what the file shows depends on its Read photos and documents customers send switch (the Clone's Skills tab; off until you turn it on). With it on, the Clone reads a photo or a PDF before it answers and replies from what it read — the amount and reference on a receipt, say, or the product in a photo — and reading each file uses units. What it read shows in Messages under the photo or file, after an eye icon (hover it for "What the clone read in this file"); it's written in English, with any text in the file copied as written. Videos and other kinds of files, such as a Word or Excel file, are never read, and the switch changes nothing on Telegram, email or web chat, where a file never reaches the Clone. A reading isn't proof of payment: the Clone is told to confirm a payment only if its instructions say a receipt is enough, and otherwise to say a person will check it. When the Clone doesn't read the file — the switch is off, or it's a file it doesn't read — it can't see it: it follows whatever its instructions say about files, photos or receipts — worth adding a line, such as what to do with a payment receipt — and otherwise acknowledges the file and asks the customer, in their language, to describe or confirm the key details in text. Only what the customer typed with the file counts toward that language, so a photo sent with no text keeps the conversation in the language it was already in. That check isn't on every account yet: some accounts created before mid-July 2026 still answer through an earlier system without it, where the Clone judges the language from the whole message, including the note IcloneU adds about the file, which is in English. If a reply to a file comes back in the wrong language there, contact support. The Clone is also told, in a rule your instructions can't override, that it knows only what it was told about a file: it never says what a file it didn't read shows, or agrees with what the customer says it shows. A file hands the conversation to a person only if the Clone has Enable handover on, its Hand over when a customer sends a file the clone can't read switch is on (Identity tab → Human Handover → Configure schedule), the Clone couldn't read the file, and the file arrives within the Clone's handover hours (with no hours set, at any time): the Clone still replies to it, and then the conversation is badged Handover requested. A photo or PDF the Clone read stays with the Clone. Otherwise, handovers follow the Clone's handover settings and instructions. If a workflow picks up the message, it gets the file as {{rv_attachment_url}}, {{rv_attachment_type}} (image, document or video) and {{rv_attachment_name}}.

Review and rate individual replies

Every Clone reply in a transcript carries its own control strip — your reply-by-reply review kit:

• Good response / Bad response — thumbs-up or thumbs-down a reply. Feedback is a tuning signal for the Clone's future answers, so rate freely: it's the cheapest way to steer quality.
• Notes — attach a note to a specific reply.
• Copy — copy the reply's text.
• Classification (the ⓘ control) — opens a popover showing how the reply was classified: an icon, a category and a one-line description, e.g. "💬 Answer — The clone responded the question". The categories use the same vocabulary as the Response authorizations required for list in a connection's settings.
• The edited marker — a small text-cursor icon next to ⓘ on a reply someone changed before it went out. It's a label, not a button; its tooltip reads "Edited — this response was changed from what the AI wrote".
• Delete response (the trash can, last in the strip) — takes a reply that already went out off the platform. It's offered on replies to Facebook, Instagram and YouTube comments in Post replies, and on Telegram replies in Messages for 48 hours after they were sent, your team's replies as well as the Clone's, to members who can manage that surface. Replies in WhatsApp, Messenger or Instagram DMs, by email or in web chat can't be deleted, and neither can a comment reply that went out as a private message. After you confirm ("Delete this response?" — "It will be removed from the platform and marked as deleted here. This can't be undone."), the reply stays in the thread, greyed out, with a Deleted pill and the note "Deleted from the platform"; customers no longer see it.

Take over a conversation yourself

Sometimes reading along isn't enough and you want to answer personally. A conversation can pass to a human in three ways: the AI detects a predefined scenario (the user asks for a person, for example), the AI detects a condition you configured in the Clone, or you simply take over manually. Here's the manual route:

1. Open the conversation

   In Messages, select the conversation. While the Clone is in charge, the header reads IcloneU is handling this conversation.

2. Click Take Over

   Use the Take Over (person) icon in the header. The header flips to Agent handling, tagged with your user, a system line records that an agent joined, and the Clone pauses — it stops auto-replying immediately.

3. Write to the customer

   A composer appears at the bottom — Type your message to the customer... — with attachment, emoji and microphone controls. The microphone records a voice note of up to 2 minutes: listen back, then Send voice note or Discard, and it reaches the customer as an audio message on the conversation's channel. The Clone stays paused while the conversation is in your hands.

4. Return it to the AI

   When you're done, click Return to AI (the circular-arrow button). A Conversation returned to AI toast confirms the hand-back, the transcript logs "You are now chatting with the AI assistant.", the header reverts and the composer disappears — the Clone resumes on the very same conversation. It reads what you wrote as a message from someone on your team: it won't claim or apologise for it, and won't ask again for what the customer already told you. Accounts still answering through the earlier system mentioned above don't have this.

Your own replies can fail, too. On WhatsApp, Facebook and Instagram the platform only accepts messages for 24 hours after the customer's last message. Once that window has closed, an amber strip above the reply box warns you before you type: "This conversation's messaging window has closed — Meta only delivers replies within 24 hours of the customer's last message. Anything you send now will not reach them until they write again." The box stays usable, but a reply you send then is rejected, not held — and the thread says so, instead of rendering your message as though it arrived: it carries a red Not delivered pill. Hover the pill and it explains why — "The platform rejected this reply because its messaging window had already closed — it never reached the customer. You can reply again once the customer sends a new message." A rejection for any other reason shows the same pill with a shorter note, "The platform rejected this reply — it never reached the customer." There's no retry button for either: the message is gone, and on the 24-hour channels the window reopens the moment the customer writes again. The Clone's own replies carry the same Not delivered pill when the platform doesn't deliver them, on every connection. That includes a WhatsApp reply Meta accepted and then reported it couldn't deliver; that pill shows once you open or reload the conversation. Messages a workflow sends don't appear in the thread, so a refused one shows in the workflow list's Runs column in Automations instead. When the platform gave a reason, the tooltip adds it after "Reason given by the platform:", usually in the platform's own words, in English. On a connection with Direct-message review on, a Clone reply refused for a reason other than the closed window also has a retry button, whose tooltip reads Retry sending.

When the Clone escalates a conversation on its own, it also writes an Agent Briefing — a short summary of the case for whoever picks the thread up, drawn from the transcript and whatever it has gathered about the contact. Escalation isn't the only thing that produces one: the Clone also writes a briefing when it captures a contact's details mid-chat, and again every fifth customer message on any longer conversation. So a perfectly ordinary thread that was never escalated can carry a briefing too — its presence tells you the conversation got long enough to be worth summarising, not that anything went wrong. The briefing is kept on the conversation: open the contact in Contacts and look under Recent conversations, where every conversation that has one shows its briefing under the row.

One thing to know before you lean on it: a briefing is refreshed only while the Clone is still answering. Once an agent is actually handling the conversation the Clone steps out of the thread entirely, and the briefing stops updating with it — so what you read is the case as the Clone last saw it, not as it stands after a human took over. Treat it as the handoff note it is, and read the transcript itself for anything that happened since.

While a conversation waits for a human, account-wide settings decide whether the Clone sends periodic wait messages to reassure the contact, how long to wait, and what happens if nobody picks up — handing the conversation back to the Clone or notifying your team. You'll find them under Account settings → Options (avatar menu, top-right).

A handover nobody has picked up doesn't have to wait for that timeout. While a conversation is badged Handover requested or Pending assignment, anyone on your team who can manage conversations can cancel it with the undo-arrow button in the conversation's header — its tooltip reads Cancel the handover and return to AI — without assigning it to themselves first, and the Clone answers again. To clear a backlog, tick the conversations in the list and click Return Selected to AI, next to Delete Selected. After you confirm — the dialog warns "Any agent handling them will be released and the clone will answer again." — every selected conversation in a handover goes back to the Clone, including ones an agent is already handling.

Hold replies for your approval

On comment channels — Facebook, Instagram and YouTube — people don't just message you privately; they also comment on your posts. Replies to those comments are managed in Post replies (left sidebar, under Interactions), and you can require your authorization before a reply goes out — the same per-connection gate that direct messages can opt into (see the callout below). The gate is configured per connection:

1. Open the connection's settings

   Go to Connections, select your Facebook, Instagram or YouTube connection and click the Settings (gear) button. The Account options panel for that account opens.

2. Tick the categories to hold

   Under Response authorizations required for, tick every category that should wait for you — for example Answers, Casual greeting or Unknown answers. You can select more than one, or use Select all. A ticked category is held for your approval; an unticked one is sent automatically.

3. Save

   Click Save. The setting is per connection, so the same Clone can answer freely on one channel and wait for your approval on another.

A held direct-message reply appears in its Messages thread with a green checkmark approve button (its tooltip reads "Authorize response") and a cancel control. On WhatsApp, Facebook and Instagram the platform only accepts replies for 24 hours after the customer's last message, so the held reply also shows an amber hourglass pill with the deadline — approve before 9:15 AM — and the connection settings show the same warning while hold mode is on: "This platform only allows replies within 24 hours of the customer's last message. A reply that isn't approved in time expires and can no longer be sent." A reply that isn't approved before the window closes can no longer be sent: it shows the red Not delivered pill, and its tooltip says the messaging window had closed. On Telegram and Email there is no window, so a held reply simply waits until you act on it.

Auto-send after a delay is the middle ground: the reply parks for the minutes you set (5–240, 30 by default — the field's hint reads "Minutes before the reply sends automatically. You can approve, edit or cancel it during this time.") and then sends on its own. You can send it early, edit it, or cancel it during the window; if the customer writes again first, the parked reply is cancelled and its answer is folded into the regenerated one. On WhatsApp, Facebook and Instagram this mode is the safer choice — a reply you never get to still sends inside the platform's 24-hour window instead of expiring.

A reply that hasn't gone out yet carries a status pill in the thread: blue Scheduled when reply timing parked it, amber In review — In review · sends 9:15 AM — when the connection is set to Auto-send after a delay, and the amber approve before… deadline on held replies on WhatsApp, Facebook and Instagram. Scheduled and In review replies can be sent right away (the send button's tooltip reads "Send now — overrides the schedule"), cancelled, or edited before they go out (text only: a file the Clone sends can't be edited). And conversations with a reply waiting on you — held for approval, or parked in review — show an hourglass badge with a count in the conversation list, on any channel with direct-message review on, until you deal with the reply.

• Automation replies are never held. Responses produced by an automation skip authorization on every channel — even on comments with the toggles on.
• “Continue in private” skips the gate. When a comment is escalated to a DM on Facebook or Instagram, the escalated reply is a direct message and sends without authorization, whatever the comment toggles say.
• Under human control there's nothing to approve. Once an agent is handling the conversation — badged Agent handling — the Clone goes fully silent (the customer's messages still land in Messages). While the conversation is only badged Handover requested or Pending assignment, the Clone keeps answering — with context telling it a person will take over — until an agent actually takes it. On a WhatsApp number in Coexistence mode, replies you send from the phone show in the conversation as team replies with a green WhatsApp Business app pill, and replying from the phone at either of those two stages counts as taking the conversation: it moves to Agent handling and the Clone falls silent. In a normal conversation, the Clone's Replies from your phone setting decides (Identity tab → Human Handover → Configure): Keep answering, the default, doesn't pause the Clone; Pause the clone for a while keeps it quiet on that conversation for the hours you set (1 to 24, 2 by default) after each phone reply — the header then reads Paused: you answered from your phone with the time left, and its play button, Resume the clone now, lets it answer again at once; Hand the conversation to me moves the conversation to Agent handling until someone returns it to the AI.
• The escalation reply itself is always reviewed. On a connection with Direct-message review set to Hold until approved or Auto-send after a delay, the reply that hands the conversation to a human is held (or parked) whatever categories you ticked — it is still AI-written text going to your customer, so it goes through the same gate. The handover doesn't wait for you: the moment the Clone escalates, the conversation is flagged for handover — it shows up under the Handover requests filter — your team is notified and the briefing is written, so reviewing that text never delays the takeover. The one exception is a handover triggered by a customer's file: the Clone's reply to that file isn't the escalation, so it follows your ticked categories like any other reply.

Post replies is where you then track each reply's journey. Pick the account to view and filter by response status. For Facebook and Instagram, comments sync automatically in the background, so the list stays current on its own; Refresh from Platform reloads the view from IcloneU's stored data. For YouTube, Refresh from Platform re-fetches the latest comments from the platform. The statuses you can filter by:

IcloneU doesn't publish per-status timing or retry details, so treat the names as your guide. And if the page shows "No connections available", nothing is wrong — you just don't have a comment-capable channel connected yet (see the FAQ).

Three shortcuts help you work through the list. The round down-arrow button in the toolbar — its tooltip reads Jump to next pending authorization — opens the next post with a reply waiting for your approval; while it looks, the list reads "Looking for the next pending authorization...". When you open a post, its header has an Open post link that opens the original post on the platform, and clicking the post's thumbnail there opens a Post image window — or says "The image for this post could not be retrieved." when the platform won't provide it.

Frequently asked