Helpdesk

Your customers' conversations – from the chat on your site, from your support email, and the calls and visits your team records by hand – in one queue, each conversation bound to the customer record once you know who it is. The agent answering a ticket sees who they're talking to: lifetime value, orders, segment, and every previous conversation.

You'll find it under Customers → Helpdesk. For an overview written for support and CX leads, see the AI helpdesk for support leads.

Where conversations come from

Chat. With Site chat in Test or Live on one of the brand's sites (Website SDK → Site chat), visitors write to you from any page of that site – anonymously, or as a signed-in customer when your site signs who they are (named without a signature, they show as not verified). Replies you send from the Helpdesk appear in their chat within seconds, and reach them by e-mail if they have left. Until the first one arrives, the empty queue says Chat is live on yoursite.com – the first conversation lands here (or Chat is in test on yoursite.com). See Chat conversations below.

Email. Every Helpdesk has an inbound address (support-…@inbound.theaicmo.com). Anything sent or forwarded there becomes a conversation. Four ways to connect your real support address, all under Setup → Integrations → Support inbox:

Forwarding – add your support address and forward it to the inbound address at your mail provider. Replies you send go out from your address whenever your sending setup allows, and customer replies come back into the Helpdesk automatically.
Hosted support address – the complete setup. Delegate one unused subdomain (for example help.yourbrand.com), add the DNS records shown, and the address works in both directions: customers write to it and replies come from it. Your existing company email is never touched.
Connect a Microsoft 365 mailbox – your support address is on Microsoft 365 (Outlook)? Microsoft no longer accepts passwords over IMAP, so connect it here instead. Enter the address, press Connect Microsoft 365 and sign in with your work Microsoft account. An administrator of your Microsoft 365 approves the app for your organisation once – if that isn't you, use copy the approval link, send it to them, and press Check again once they have approved. From then on new mail to that address lands in the Helpdesk within a minute and replies leave from it. Only mail arriving after the connection is read, and no password is stored. Your administrator can limit the app to that one mailbox.
Connect via IMAP – your mail runs on your own servers? Provide IMAP credentials; the connection is tested live before anything is saved, credentials are stored encrypted, and only mail arriving after the connection is read. New mail lands in the Helpdesk within a minute.
Send replies through your mail server (SMTP) – the outgoing half of that connection. Add your SMTP host, port and login (usually the same as IMAP); the login is tested live, and from then on every Helpdesk reply leaves from your support address through your own server. This is support-only: the brand's marketing sending is never touched.

Files and photos in e-mails. When your mail is connected through Microsoft 365 or IMAP, the files a customer sends arrive with their message: a photo pasted into the e-mail or attached to it shows as a picture, and a PDF, Word, Excel, PowerPoint, text or CSV file as its name, type and size; click one to open it. A message that was only a photo shows the photo. Up to 10 files of up to 10 MB each are kept per e-mail. Tracking pixels, small signature icons and your own logos quoted back under a reply are left out; everything inside a forwarded e-mail counts as sent by the customer. A file that is not kept here (a program, an archive, one over 10 MB) is still named on the message in grey, with a line under it saying why; it stays in your mailbox. Files are kept privately and open only for people who can work the conversation. Mail that arrived before this was switched on is not filled in afterwards. Until your organisation's administrator confirms version 1.4 of the AI Support Agent Addendum in Settings › Agents, the AI is told only that a file came with the e-mail – never its name or anything in it – and where the answer depends on the file, it says a colleague will look at it. Once 1.4 is confirmed, and wherever the support employee works the brand, it reads the photos and PDFs a customer e-mails exactly as it reads the ones visitors attach in Site chat (see Files visitors attach): a photo as it is, converted first; a PDF by its text, with names and numbers replaced; never a HEIC photo, a Word or Excel file, or a file only named in grey. A small picture pasted into the e-mail, such as the logo in a signature, is treated as part of the e-mail, not as something the customer is asking about.

Contact forms. Many website forms send their messages from your own address (for example podpora@yourbrand.com) and put the visitor's address in Reply-To. When a message comes from one of your brand's own addresses (your support mailbox, your website's domain or your sending domain) and names someone else in Reply-To, the conversation belongs to that person: replies go to them, their customer record is found or created, and the name the form wrote next to their address is used. A line on the conversation says it came through the form. A normal customer e-mail keeps its sender, whatever its Reply-To says. A form message that gives no Reply-To shows no customer ("Not identified"), the AI does not answer it, and a line asks you to attach the customer; attaching them sets where replies go. The Helpdesk never sends a reply to your brand's own address.

The Replies send from setting controls the outgoing identity when neither Microsoft 365 nor an SMTP server is connected: Auto uses your own address or verified domain whenever possible and the platform address otherwise – customer replies always find their way back to the Helpdesk either way.

Logged by hand. A phone call, someone at the counter, a message that reached you somewhere we do not read – New ticket, top right, records it on the desk like any other conversation. See Logging a call or a visit below.

Everyone who can open the brand – its owner, the organization's owners, admins and managers, and brand members – sees the same inbox, queue and canned replies, whoever connected the mailbox.

Several brands, one desk

If you can open more than one brand – the organization's owners, admins and managers see every brand of the organization; everyone else the brands they are given – the Helpdesk opens on All brands: one queue with the conversations of all of them. Each row names its brand beside the customer, and the open conversation names it first in its header, so you always know which brand you are answering for.

The brand picker at the top right, beside New ticket, says how many conversations are waiting on each brand. Pick one to narrow the desk to it; while you do, the picker shows how many are waiting on the others (+2). Picking a brand here changes the desk only – the rest of the app stays on the brand it was on. All brands brings everything back, and so does opening the Helpdesk again.

Everything around a conversation belongs to that conversation's brand: its canned replies, the support employee and its drafts, lookups, the customer's record and the address the reply leaves from. On All brands, New ticket asks which brand the call or visit was for; Knowledge and Lookups, which each brand keeps for itself, have a selector above them; Reporting counts every brand together – pick a brand for one.

Reporting's Who handled it lists each person who held tickets, with how many they resolved and their typical first reply, plus a line for tickets nobody held. Someone who has since left the team keeps their name when we still have it, and reads "Former teammate" otherwise.

Leaving someone out of Reporting. Not everyone who appears in Reporting answers customers: an owner may hold tickets that teammates really answer. An administrator of the brand can point at that person's row in Who handled it, open the … menu and choose Leave out of reports. Their row goes, and their replies stop counting as the team's in When customers write (the reply bars, the number of people and "the team sent" line). Everything about the customers stays as it was: how many conversations came in and when, the channels, the topics, and the first reply and resolution times, because the customer was answered whoever answered. The choice is the brand's, so everyone who opens that brand's Reporting sees the same report. Under the card a list, Left out of reports, always names who is left out, one row each, with Bring back on each row and Manage in Team under it for administrators. On All brands, an administrator of every brand shown leaves the person out of all of them at once, and someone left out of only some brands is named with those brands. The same choice is in the admin portal's Team page: Counted in helpdesk reports, per brand, right under who takes customer support tickets. Untick a brand there and the person is left out of that brand's Reporting; a change in either place shows in the other.

Any period. Reporting opens on the last 30 days. The date button above the numbers offers the last 7, 30 or 90 days, or pick any other period on the calendar, up to a year: click its first day and its last, then Done. Days are your own calendar days, in your browser's time zone, both included, and every number on the page is that period's: the conversations that came in during it, their reply and resolution times, the hours, who handled them, topics and channels. Waiting for a reply now is about this moment: for a period that has ended, it lists that period's conversations still waiting today. The period is in the page's address, so the browser's Back button returns to the period before, and a link you share opens the same period.

Export. Export, beside the dates, makes a file of exactly what the page shows for that period, brand and the people left out:

Spreadsheet (CSV): every table on the page, one after the other, each with its title and column names, numbers as numbers. It opens in Excel in columns: where your browser writes decimals with a comma (Czech, Slovak, German and others), the file uses semicolons and decimal commas, as Excel expects there; elsewhere commas and points.
PDF report: the same report as a document to share, with a cover, the brand, the period, the time zone and when it was made, in the brand's colours.

Files are named by brand and dates, for example helpdesk-report_loamwood-outfitters_2026-09-01_2026-09-30.pdf. People left out of Reporting are left out of the file the same way, and the file names them.

Reporting by hour

When customers write, under Volume in Reporting, is for planning people and the rota. It follows the period you pick and the brand picker:

24 hours: what came in each hour of the day above the line, and your team's replies below it, with the number of team replies under each hour. Conversations counts new conversations; Messages counts every customer message, follow-ups included. Replies the AI sent on its own are shown apart, as a dashed outline, and never counted as the team's.
By weekday and hour: Monday to Sunday against the 24 hours, with the number in every hour that had one. Point at an hour, tap it or use the arrow keys, and the line under the chart gives its numbers.
First reply, by when they wrote: the typical time to a first reply for conversations that came in during each four-hour stretch, weekdays and weekends apart, with how many were replied to ("2 of 14 replied"). A typical time is shown from 3 replies; with fewer, the stretch says how many were not answered ("12 not answered").

The hours are in your own time zone, the one your browser uses, and the card names it ("Prague time"). If the browser's zone cannot be used, the card says the hours are in UTC. The daily Volume chart uses the same clock. Numbers are totals, not averages; when a period holds some weekdays more often than others, the card says which, and the shading allows for it, so each weekday is compared per day. Tickets logged by hand count as conversations that came in, but not in first reply, because their time is when they were logged. Pick Email, Chat, Social or By hand to see one channel; the choice appears when more than one is in the period. On All brands every brand is counted together, test brands included.

Logging a call or a visit

Support does not stop at the channels we read. New ticket opens a form that takes about twenty seconds:

What happened – Phone call, In person, or Other. (E-mail is not on the list on purpose: mail already arrives on its own, and a hand-made e-mail thread would look like one the customer wrote to while threading nowhere. A mail that reached you elsewhere is Other.)
Customer – start typing a name, an e-mail address, a phone number or a customer number and the brand's records come back as you type, each with what is already on file: company, segment, how many conversations you have had before, and how long they have been on record. Arrow keys move through the list, Enter picks. Nobody matching? Create a new customer takes their details, or log the call with just a name and a number – the ticket keeps both, and you can attach a record later. A customer created here has no marketing consent on record, the same as someone whose first e-mail creates their record – unless they already gave your brand consent before (it stays on their consent history and is never undone). Your replies and other service e-mail reach them, but marketing e-mail does not until they give consent (tick Marketing e-mail on their record in Customers, or they tick a landing page's consent box). On a Support seat the box finds a customer by their exact e-mail, phone or customer number only, and instead of Create a new customer it says who can create the record (see Support seats).
Subject, the first note, priority, assignee and tags. The note is internal, like every note; you add more over time as the case moves.

The ticket opens in the thread. It sits in the queue as Open and is not marked as waiting for a reply – the call was already answered on the phone – until you move it to Pending or Resolved yourself. Reporting counts it under its own channel and never counts it as "never answered".

The rail on a logged ticket shows the channel, when it was logged, who logged it, and how the person on the other end was identified: assigned by hand when a customer record is attached, typed by hand when it is a name and a number the agent wrote down. Neither is a verified identity, and the Helpdesk never claims it is.

Replying to a logged ticket. There is one road out: e-mail. The Reply by e-mail button works when the attached customer has an address and your brand has a support mailbox or a verified sender; the answer leaves as e-mail, and the ticket stays a phone call. When there is no road, the button is off and says in one line why. Adding notes always works.

Working the queue

The desk is three panes and fills the window – nothing here scrolls the page.

Left: the views and the queue. The views answer the three questions a desk asks all day – Mine, Unassigned, Needs reply – then status (Open, Pending, Resolved, All) and the doors work arrives through (Email, Chat, Social, By hand). A door narrows whichever view is picked: with Chat picked, Pending lists the pending chats, Resolved the resolved ones and Mine your chats, every number beside a view counts chats only, and each door says how many of the view in front of you came through it. Click it again for every channel. Every number beside a view is counted live, on every brand the desk shows, not on the page you happen to be looking at.

Each row carries the customer's name, the subject, the last thing said, who owns it and how long the customer has been waiting. A 2px tick down the left edge is the channel: deep green for e-mail, soft green for chat, brown for social, sand for anything logged by hand. The waiting time turns brown after four hours and red after a day – on a chat, where someone is sitting on the page, after two minutes and ten. High priority is red, and a conversation nobody owns says Unassigned.

The queue is ordered by who has waited longest, or by newest – the control at the top right of the list also holds the agent, priority and tag filters (the tag filter offers the brand's tags, the archived ones last). The agent filter lists the people who take tickets on the brands the desk shows, plus anyone still holding an open or pending ticket here, so work held by someone who has left the team can still be found. Search (in the header, or press /) covers the subject, e-mail, name, phone and customer number of every conversation on the brands the desk shows, however far back.

Hover a row for a checkbox. Tick several and a bar appears: assign them, set their status, add a tag, or resolve them all at once. Assign offers the people who take tickets on every brand you ticked; when nobody takes them all, it says so. If some of the tickets cannot go to that person, the rest are assigned, the bar says which were not and why, and those stay ticked.

Middle: the thread. The subject is editable – click it. The status and the assignee sit in the header, because they are what you change most; if a conversation belongs to a teammate the header says so and offers Take it.

Opening a ticket makes it yours. Open a ticket that nobody holds, on a brand you take tickets for, and it is assigned to you: the assignee in the header changes to you and the thread shows the line You took this by opening it. It happens once the thread has been on screen for a moment (just over a second when you walk the queue with j and k), so skimming past a ticket never takes it. Answering a ticket nobody holds makes it yours too (You took this by replying); a note does not. If a teammate opened it a moment before you, it is theirs: the header says Maya Chen owns this one and offers Take it, and nobody's ticket is ever taken from them by opening it. Nothing is taken when you only take tickets for other brands, when the ticket is resolved, when a teammate is writing on it, when a teammate has already replied to the customer on it (opening it only reads it; answering it makes it yours), or for the ticket you have just logged with New ticket (whoever you chose there stays). To hand a ticket back, choose Nobody in the assignee menu; it is not given to you on opening again. While you read a ticket you have just taken, it stays where it is in the list, even in Unassigned, until you move on.

Who can be assigned. The assignee list follows the conversation's own brand, not the brand the rest of the app is on: it offers the teammates who take that brand's tickets. Taking tickets is a choice per person and per brand, whatever their role: someone can open a brand without taking its tickets. A Data Engineer seat is never offered (it cannot open the Helpdesk), and Me and Take it appear only when you take that brand's tickets yourself.

At the bottom of the list, Start taking tickets for the brand (or Stop taking tickets for the brand when you already do) changes your own choice for that brand in one click. Stopping never takes tickets away from you: if you still hold some on that brand, the Helpdesk says how many, and they stay yours until you or a teammate reassigns them. If you had chosen every brand, including brands added later, stopping one brand keeps the others but no longer covers brands added later, and the Helpdesk says so. Owners, admins and managers also see Choose who takes tickets in Team, which opens Team in the admin portal in a new tab, where they choose for anyone on the team. When nobody takes a brand's tickets yet, the list says so instead of showing an empty list. On a ticket a teammate holds, a brand you can open but do not take offers Start taking tickets for the brand where Take it would be, and so does New ticket. New ticket lists the people for the brand you pick and starts on you when you take that brand's tickets, otherwise on Unassigned. A brand nobody can take says so instead of showing an empty list. The Helpdesk checks this itself every time a ticket changes hands, whoever or whatever makes the change: giving a ticket to someone who cannot take that brand's tickets is refused with one line that names them and the brand, nothing changes, and the lists drop them at once. New ticket refused the same way logs nothing and says so: log it unassigned, or for someone who takes that brand's tickets. A ticket someone already holds stays with them after they stop taking that brand's tickets, and the team can still answer, resolve or reopen it, until someone reassigns it; it just cannot be handed back to them. Names on the thread come from the same team list, so someone who has left keeps their name on what they wrote, unless that name is an e-mail address: then, like an account that no longer exists, they read "Former teammate" (the Helpdesk never shows the address of someone who is not on your team). A teammate added while the Helpdesk is open is named as soon as they write, and offered when you come back to the tab. Your own chip carries your initials.

When someone leaves. Removing a teammate from the team, or taking away their access to a brand, hands the open and pending tickets they held there back to Unassigned, and each conversation shows the line "Returned to Unassigned: their name can no longer open this brand". Resolved tickets keep their name. An owner, admin or manager keeps their tickets when a brand given to them one by one is taken away, because their role opens every brand.

The timeline shows four different things, and each one looks like what it is: the customer's messages as tinted cards on the left, your replies as white cards on the right with the name of whoever sent them, internal notes full-width on note paper with a dashed edge, and system events – assigned, reopened, customer attached, a customer logging in on your site during a chat – as one quiet line. A login reads Customer logged in as jane@acme.com – identity verified when your site signed it, and Customer logged in as jane@acme.com – not verified when it did not. The same line opens every chat a customer starts while already logged in, so it says who your site named, not that they logged in just then. An AI draft appears as its own card marked Draft – not sent, with Use it and Discard. Replies, notes and those lines are kept as written: each names the teammate who wrote it, and nobody can change or delete one from the Helpdesk.

An assignment line reads right to whoever reads it: the person who took the ticket reads You took this one, their teammates Jan took this one; the person it was given to reads Jan assigned this to you, everyone else Jan assigned this to Vojtěch. A ticket that became someone's when they opened or answered it reads You took this by opening it to them and Jan took this by opening it (or by replying) to everyone else.

When a ticket is given to you, the bell at the top of the page says who gave it to you and on which brand – Jan assigned you a ticket on CZ Prod – and opens that conversation. It never repeats anything the customer wrote, the subject or an address, so it is safe in the e-mail copy of your notifications too. Taking a ticket yourself rings nothing. Several tickets given to you on the same brand before you look gather into one line – Jan assigned you 3 tickets on CZ Prod – which opens Mine; if your notifications come by e-mail too, the next e-mail carries the new count. Helpdesk bells show whichever brand the app is on, because your tickets span every brand you take tickets for; Marketing Pulse leaves them out. New bells arrive while the page is open.

The composer has two modes. Reply is white and goes to the customer. Note turns the box to note paper and stays with the team – you can see which one you are writing before you write it. Type / for a canned reply, and Save as canned adds the current draft to the list for everyone. Send sits in the bottom corner of the box itself, with a small arrow for Send and resolve and Send and mark pending; in Note mode it reads Add note. ⌘↵ sends.

Files on a reply. An e-mail or a chat reply can carry files: a reply to an e-mail, to a logged ticket that is answered by e-mail, or to a chat on your site. Press Attach in the bottom row of the box, drop files onto the box (it says Drop to attach to the reply), or paste a screenshot. Each file goes up at once and sits in the box as a chip with its name and size; while it uploads, a thin line runs along the chip with the percentage, and Send waits until every file is up. The header then reads Goes to the customer with 2 files (on a chat, Goes to the chat with 2 files). The customer receives them as ordinary attachments of the e-mail, from the same address and in the same thread as any reply; in a chat, the visitor sees them in the conversation (see Chat conversations).

How much. Up to five files and 2 MB per reply, whichever mailbox sends it, and up to three files and 2 MB on a chat reply; hold the pointer on Attach to see it (Up to 5 files, 2 MB in all, or Up to 3 files, 2 MB in all on a chat). The line beside the files says where you are – 2 files · 1.4 MB of 2 MB – and, while the files are past the limit, says so in the same place (2.2 MB of 2 MB. Remove a file to send.) and Send waits. A single file past the limit keeps its chip with its size and the limit, for example 2.6 MB · a file can be up to 2 MB. A photo straight from a phone is often larger; send a smaller copy of it.
What. PDF; pictures (JPG, PNG, WebP, GIF, HEIC); Word, Excel and PowerPoint (.docx, .xlsx, .pptx); OpenDocument text and spreadsheets (.odt, .ods); text and CSV. Each file is judged by what it is, not by its name. Programs, archives, web pages, pictures drawn in code (SVG), documents with macros and the old .doc, .xls and .ppt formats are never sent: such a file keeps its chip with This kind of file is not sent, and the line beside the files reads Remove the file that can't be sent until you remove it.
When something goes wrong. A file that could not upload says Could not upload, with Retry. If the reply itself cannot be sent, the words and the files stay in the box and Send tries again with the same files. A mailbox's server that refuses the size says so in plain words, and nothing is filed. Moving to another conversation while a reply with files is still sending does not stop it.
Where files do not go. Notes, social replies and store replies take no files, and neither does a chat that continues by e-mail (its replies go to an address the visitor typed, which may not be theirs). In Note mode the files you attached stay in the box, which says 2 files stay with the reply. Switch to Reply to send them.; a file dropped on a composer that cannot take it says Files go only with an e-mail or a chat reply, so nothing was attached., and on a chat that continues by e-mail This chat continues by e-mail now, so no files go with a reply. Nothing was attached. If the visitor moves the chat to e-mail while you are writing, the files you attached stay in the box, which says Chat continues by e-mail. Remove the files to send., and Send waits until you remove them.
Afterwards. The files sit on your reply in the thread like the files a visitor sends: a picture as a thumbnail, anything else as its name and type – DOCX · 120 KB. They are kept privately and open through a link the desk asks for each time; a document downloads rather than opening in the browser. Files attached but never sent are deleted after a day; files you sent stay as long as their message, and the files of a conversation go with it when a customer's data is erased.

Right: the customer. Two sections you can fold away, and they stay folded next time: Customer (who they are, what they are worth, their previous conversations) and Details (priority, tags, the AI setting, the channel, when it started, who logged it, the identity level – verified, claimed, assigned by hand, typed by hand or anonymous – the language, the tone, and, once resolved and rated, the CSAT score).

The keyboard


`j` / `k`	next / previous conversation (one nobody holds becomes yours once you stay on it a moment)
`Enter`	open it
`r`	write a reply
`n`	write a note
`e`	resolve
`a`	assign it to yourself (when you take that brand's tickets; otherwise it says why and offers Start taking them, which starts and then assigns it to you)
`x`	tick it for a bulk action
`/`	jump to search
`Esc`	drop the selection, or leave the box
`⌘↵`	send

New messages arrive on their own – an open thread grows as the customer writes, and the queue keeps its counts current without a refresh.

When a teammate has the same conversation open, you see them: their initials appear on the row in the queue, and the thread header says "Jan is here". The moment they start writing it changes to "Jan is replying…" – and if you are writing a reply too, a line inside your own composer says "Jan is already replying – check before you send." It never stops you sending; it just makes sure two people do not answer the same customer without knowing. Nobody is tracked and nothing is recorded: close the tab and you are gone from the desk. Only people on that brand's team, signed in, can see who is on its desk, and only they show up there.

The view and the open conversation are in the address bar, so your browser's Back button works: back out of Reporting into the queue, back out of a thread into the list. A link like /helpdesk?view=unassigned opens straight into that view, and you can send a colleague the link to a conversation.

Chat conversations

A chat is someone on your site right now, so the desk shows what an e-mail cannot.

Whether they are still there. The thread header and the rail say On the site now · /checkout while their page is in use, and Left 12 min ago once it is not; the queue row says on site beside Chat. The rail also lists the page they were last on, the site, their device, their browser's language and their local time – and, when the visitor allowed marketing on your site, where they came from and the page title. These are shown as text and never as links. Their page, device and language arrive with their first message.

Who they are, and how sure. Signed in – identity verified when your site signed who they are – the customer's record sits beside the thread. Logged in on the site – not verified when your site named them without a signature, and Logged in on the site – signature didn't match, not verified when it sent one that did not match. Anyone who loads your site can name anyone, so such a name is a claim: the queue row reads Not verified · jane@acme.com and the thread header jane@acme.com (not verified), and the rail says The site says this is jane@acme.com but doesn't sign it, so anyone could claim it. Confirm who they are before discussing their account. (or The site says this is jane@acme.com, but its signature didn't match, so it isn't verified. Confirm who they are before discussing their account.). No record is attached and nothing treats the name as theirs – no reply e-mail, no account lookups, nothing in your warehouse. Open their record shows the customer the name matches, when you can open Customers; find the customer binds a record by hand once you are sure; and an administrator of the brand also gets Set up signing, which opens the site's Identity tab in the Site chat studio. Reply address typed, not verified when they gave an address in the chat: a place to reply to, not an identity, so no record is attached – find the customer binds one by hand once you are sure. Anonymous otherwise.

Logging in during a chat. When a visitor logs in on your site while they are chatting, the thread says so within seconds, on the chat's next check – no new message needed, and nothing earlier is lost: the line Customer logged in as jane@acme.com – identity verified (or – not verified) appears, and the rail follows. This happens for a chat they wrote in within the last hour; an older one shows the line with their next message. A signed login attaches the customer's record; an unsigned one is a claim, as above. It does not ring as a new message.

A site that does not sign. An administrator of the brand sees one line across the top of the desk – in the Chat view, or with one of the brand's chats open – while one of its sites names logged-in customers without a signature – shop.example names logged-in customers without a signature, so chats show them as not verified. – or sends a signature that does not match – shop.example sends a signature that doesn't match, so chats show customers as not verified. With Require signed identities on, they read shop.example names logged-in customers without a signature, so chats treat them as anonymous. and shop.example sends a signature that doesn't match, so chats treat them as anonymous. Set up signing (or Fix signing) opens the site's Identity tab. It reads what reached the chat from real visits – a test tab never counts – so nothing shows while every customer arrives signed, or when none has reached the chat in the last 7 days.

What they can read again. A customer the site signs in for chat across devices finds their earlier conversations on that site under History in the chat, read only: their messages and your team's replies, never your internal notes or a draft the AI wrote that nobody sent. See Site chat → History for signed-in customers.

Seen. Seen sits under the newest of your replies the visitor has had on screen.

Files. When the site lets visitors attach files (Site chat → Behaviour → Visitors can attach files), the photos and PDFs a visitor sends sit on their message: a photo as a thumbnail, a PDF as its name with its type and size – PDF · 180 KB. A click opens it full size in a new tab. Files are kept privately and open through a link the desk asks for each time, which works for ten minutes, so a file you open tomorrow opens just the same. A photo your browser cannot draw – an iPhone's HEIC photo outside Safari – shows as a tile that still opens. If a link cannot be had, the file says Could not load – try again or Could not open – try again; a file that is no longer kept says No longer kept; and when the browser stops the new tab, Your browser blocked the tab – click again. A message that is only files reads Sent a file in the queue. Anyone who can work the conversation can open its files, support seats included. What the AI may do with them is under What a model sees.

Files from your team. You can send files back in any chat, whether or not the site lets visitors attach them: up to three files and 2 MB per reply, of the kinds an e-mail reply takes (see Files on a reply above). The visitor sees them on your message in the chat: a picture as a picture, and a PDF or any other file as its name, which opens in a new tab (a document downloads). They open through short-lived private links, exactly like the visitor's own files. If the visitor has left and gets your reply by e-mail, that e-mail does not carry the files – the address may only have been typed – and says instead that the reply came with files they can open in the chat. Your files stay as long as the conversation shows the message. A chat that continues by e-mail takes no files.

Typing. While you write a reply – never a note – the visitor sees that someone is typing, and the composer says Goes to the chat · they see you typing. The chat names your team, or your first name when the site's chat signs replies by person.

Where your reply goes. One line in the thread header says it: Goes to the chat – e-mailed to jana@example.com if they have left, Goes to the chat only – nothing is e-mailed, or Goes to the chat – they left no address for e-mail. A reply the visitor has not seen a few minutes after you sent it goes to them by e-mail from your support mailbox; the rail shows E-mail due 14:05 unless they see it, then Replies e-mailed. Their answer by e-mail lands in this same conversation. When the owner of a typed address answers by e-mail, the header says Continues by e-mail with … – the chat shows nothing new: from then on the conversation is e-mail only, so nothing your team writes reaches whoever typed that address into the chat. Your replies – and the AI's – then go to that address by e-mail at once, without the few minutes' wait.

Being told. The people who take a brand's tickets (the Takes customer support tickets tick in Team) hear about it in the bell at the top of the page, and by e-mail if their notifications come by e-mail:

a new conversation – an e-mail, a chat, a social message or comment, a store review that gets a draft – and a resolved conversation the customer reopens while nobody holds it: New e-mail on CZ Prod;
a customer writing back on a ticket you hold goes to you alone: A customer wrote back on your ticket on CZ Prod;
the AI handing a ticket to a person goes to whoever holds it while they take that brand's tickets, otherwise to everyone who does, with the reason in a few plain words (it reached its limit of 3 replies);
a message that needs a person – a request to self-exclude, a sign of harm, a chargeback or a legal threat – goes to everyone who takes the brand's tickets and to whoever holds it: A message on CZ Prod needs a person: possible self-exclusion, harm or a dispute.

A bell names the brand and never repeats what the customer wrote, the subject, a name or an address. More of the same before you look gathers into one line – 3 new e-mails on CZ Prod – and the next notification e-mail carries the new count. When nobody takes a brand's tickets yet, the brand's owner hears instead, and the bell says Nobody takes tickets for CZ Prod yet. Choose who in Team.

The tab title counts the conversations waiting for a reply on every brand you take tickets for – (3) Helpdesk – whichever brand the desk is showing.

Sound. While the Helpdesk is open, a new ticket on one of those brands plays a clear sound, whatever it came in by: e-mail, chat, a social message or a store review the AI drafts for, and a ticket the AI hands to a person. A softer sound plays when a customer writes back on your ticket or on one nobody holds (a Pending one included, and an answer that comes a few seconds after yours), and when a teammate assigns you a ticket. On a conversation the AI answers on its own nothing rings while it does; a message it only drafts an answer for, because its kind of question is not switched on for unattended replies, rings like any other. It plays when the tab is in the background too, at most once every four seconds, and only one Helpdesk tab plays it if you have several open. Your own actions never make a sound, in any of your Helpdesk tabs or on your other computers (taking a ticket yourself included), and neither do a teammate's tickets. Sounds play only while the Helpdesk is open in a tab; on other pages of the app the bell still counts.

The control is on the queue's toolbar, beside the order: Sound on or Sound off (No sound if you take no tickets for the brands shown, since nothing would ring). Click it for the volume, a sample of each sound (Play new ticket, Play reply) and Alerts for: New tickets, Customers writing back, Tickets assigned to you. Sound is on unless you turn it off (if you had switched the old chat chime off, it stays off). After the page is reloaded the browser keeps it silent until you click somewhere on the page; the control then reads Click for sound, and one click anywhere turns it on.

Desktop alerts. In the same panel, Turn on shows a notification on your computer while the Helpdesk tab is in the background, such as New e-mail on SK PROD. The browser asks for permission when you press it, never before. A notification names the brand and the kind of message only, never what the customer wrote; clicking it opens the conversation. Several at once while you are away gather into one notification (3 new e-mails on SK PROD) that sounds once. Your computer must allow notifications from your browser (on a Mac, System Settings > Notifications), and Focus modes silence them. If you once refused, the panel says Blocked in this browser's site settings; allow notifications for the site there. They are not available in phone browsers.

These settings are Kept on this browser only: each person sets their own, on each computer they use.

A brand you can open but do not take tickets for stays quiet: its conversations are still in the queue, they just do not raise their voice. The one exception is a brand nobody takes tickets for yet: whoever its bells go to (its owner) hears it on the desk too.

Regulated – AI paused. When a chat message touches a regulated topic – a request to self-exclude, a sign of harm, a chargeback or a legal threat – no model reads that conversation again until a person allows it. The thread says Regulated – AI paused and the queue row AI paused; a person answers. Allow AI again asks once before it lifts the pause.

While it is paused, the thread also says Until you allow AI again, this person gets none of the brand's marketing e-mails, texts, WhatsApp messages, pushes or promos, and no website message meant for them alone. They are left out of its Meta and Google Ads audiences too, and taken out overnight of those already pushed. Service messages and your replies still go. Each such send leaves that person out, with no time limit, and what they were left out of is not sent to them later. A send's summary counts them as held – sensitive support conversation, never with the conversation's subject. A website message for an audience or the whole site and a popup do not ask, so the person can still see those. A newer conversation does not lift it; only Allow AI again does, within two minutes.

Ad audiences (2 Nov 2026). Pushing an audience to Meta or Google Ads leaves the person out, on every brand, and the push's counts show them as held: sensitive support conversation. Every night we take them out, one by one, of the audiences already pushed, the same way someone who opts out is taken out (see Pushing an audience to Meta or Google Ads). After Allow AI again, the next push of an audience includes them. An audience already pushed does not take them back on its own: a new push makes a new audience on Meta or Google Ads, and that one includes them.

Test chats. A conversation started from Site chat's test link carries the tag test. The AI answers it as it would a real one, but the support employee never learns from it: its answers are not kept as earlier replies to look at, its questions never reach Questions it could not answer, it adds no topic – it can carry one you already have, so a topic you switched on can be tried from the test link – and it does not count towards whether the employee believes it is ready for its next step.

The customer card

The Customer section of the right rail is who you are talking to, and it shows only what is actually known – nothing is padded out with blanks.

First, how to reach them and how sure we are it is them: the name, then a quiet line saying whether they signed in, whether the address is merely claimed, or whether one of your team bound this conversation to a record by hand. Under that, every handle the record carries – customer number, e-mail, phone, company. Click any of them to copy it, which is quicker than reading a customer number off a screen down a phone line. A customer with no e-mail at all is a first-class customer here: a phone number and an account number are a complete record.

Then what your data knows: lifetime spend, orders, segment and last purchase from your connected data, plus how long they have been on record, their company, and any tags on their record.

Then Previous conversations – how many there are, and the last five as one line each with the channel, the status and how long ago. Click one to open it. View all searches the queue for everything they have ever sent.

Find customer (or Change, when one is already attached) searches the brand's records the same way the New ticket form does and binds this conversation to the person – on any channel, not just the logged ones. The thread records who attached whom. Attaching never redirects a reply: on a conversation that arrived by e-mail or social, the address it arrived from stays the address you answer to. On a chat your site signed in, the visitor stays signed in as the person your site verified: a record you attach by hand is never read as that identity, and the thread says so.

When a high-value customer's tone turns negative, the card raises a banner – and Grant a promo lists your active promos for a one-click gesture. Grants respect suppression: a self-excluded or blacklisted identity never receives one, and the outcome lands in the thread as a note either way. Open in Customers takes you to the full record (a support seat does not see it; see Support seats below).

AI in three modes

Each conversation has an AI setting under Details:

Observe – nothing from the AI appears in the conversation and nothing is sent. Click Suggest reply and a draft appears in your composer to edit and send. (The button is not there on a thread that holds only your own notes – there is nothing the customer wrote for a draft to answer.) While the employee is on step 1, Learning, it still writes its own reply to itself on an Observe conversation and compares it with what you sent; you never see it in the thread.
Draft – every incoming message arrives with an AI draft already waiting in the thread. Click the draft to load it into the composer, edit, and send. Nothing is ever sent without you.
Auto – the AI answers on its own, inside strict limits. It only answers what it can ground in the knowledge you marked For customers, your saved replies, or the conversation itself – it never guesses business facts like opening hours or prices. Sensitive requests (self-exclusion, disputes, legal language, a customer asking for a human), anything it isn't confident about, and every conversation after three automatic replies escalate to a person – with an internal note explaining why, high priority, and a draft already waiting.

Suggestions and automatic replies match the customer's language.

Hiring the support employee

Draft and Auto stay unavailable until the Autonomous Customer Support Employee works for the brand. That takes two steps, both by an administrator of your organisation:

Confirm the addendum in Settings › Agents. Read the AI Support Agent Addendum (linked from the card) and confirm it on behalf of your organisation. The confirmation is recorded – the version you confirmed, when, and who confirmed it – and the card keeps the whole history under Record.
Hire the employee for the brand on the Agents page, the same way you hire any other teammate. A brand has no support employee until you hire one, and pausing or archiving it there stops it for that brand at once.

Hiring it does not put it in front of a customer. It starts on the first of three steps and works its way along them, and every move is yours: it advises, you decide. It never moves itself.

Learning. Nothing reaches a customer and nothing appears in your conversations. For every message a customer sends, it writes the reply it would have sent, keeps it to itself, and compares it with what your team actually sent. New conversations start on Observe.
Copilot. Its reply is waiting in each conversation for someone on the desk to approve, edit or discard. Nothing reaches a customer without a person. New conversations start on Draft.
On its own, case by case. It answers by itself only for the topics an administrator has switched on, one at a time. Everything else stays in Copilot: a proposal waiting for a person. New conversations start on Auto.

The employee's profile on the Agents page shows the three steps, which one it is on, and the sentence in which it says whether it believes it is ready for the next one, with the plain numbers behind it ("in the last 30 conversations I would have answered 24 the way your team did"). Each move is a button there. Anyone who can work the brand can move it to Learning or Copilot; only an administrator can let it answer on its own. You can move it on before it says it is ready, and you can move it back at any time.

Topics are the kinds of question your customers write about – deposits, bonus terms, account verification. The employee sorts every conversation into one as it goes, and the same profile lists them with how many of its replies your team sent with little or no change. Each topic carries one switch, Answers these on its own, and the page says what it does: on, the employee answers questions of that kind by itself, without waiting for anyone; off, it writes the reply and a person on your desk approves, edits or discards it. Even with a topic on, it hands the conversation to a person when it is not sure of the answer or cannot back it with what you gave it, and always for self-exclusion, signs of harm, chargebacks and threats of legal action. Only an administrator can turn a switch on, and the switches only count on step 3 – the page tells you which step you are on and what that means for them. A topic it believes it can handle says so beside it, once your team has answered at least 15 of its replies on that topic, sent about 8 in 10 of them with little or no change, and replaced none of the last 5; that is its opinion, and the switch stays yours. A topic it is handing back to a person more often than it used to says that too. You can rename a topic, merge two that turned out to be the same thing, and add one yourself.

The setting on each conversation always wins – anyone on the desk can put one back on Observe, and only an administrator can put one on Auto. A ticket you log by hand always starts on Observe.

Two things run without the confirmation, each with its own switch in the same card, both on unless you turn them off: marking the tone of incoming messages, and Suggest a reply.

What it costs

The employee's work is charged in credits, per piece of work, from the brand it works for – so a brand's credit budget applies to it exactly as it does to everything else.

What it did	Credits
A reply it wrote silently while it is learning	1
A reply you asked it to suggest on the desk	1
A proposal it left in a conversation	2
A reply it sent on its own	3
Reading a document or a page into Knowledge	2
Re-reading a page that has changed	1
A lookup in one of your own systems	5
Reading a message's tone, learning from an answer your team sent, "Ask it", answering a question yourself, correcting a passage	metered – a fraction of a credit each

Three things follow from that, and they are all deliberate:

Work that failed is not charged. A model that did not answer, a reply that never left, a document that turned out to hold no readable text: the credits go back. An unattended attempt that hands the case to a person instead of answering is charged as the proposal it turned out to be, not as a reply it sent.
A page that has not changed costs nothing to look at. The weekly re-read – and a Resync you press – fetches the page and compares it with what it read last time; the charge only follows a page that actually moved.
The small reads are metered, not priced. Marking the tone of a message, learning from an answer your team sent, "Ask it" in the Knowledge view, answering a question yourself and correcting a passage each make one small model call, a fraction of a cent. They are charged on what they use, gathered on the brand's monthly line in Usage, rather than a credit each. The holding reply a regulated message gets and retiring a passage use no model and are not charged.

Every line appears in Usage in the customer administration area, with the brand, the person and what it was for, and a refund is its own line beside the charge it undoes.

When there are no credits left, or the brand's budget is used up, the employee stops rather than working on credit. It says so on its profile on the Agents page in its own words, a conversation it would have answered unattended goes to a person with the reason on the thread, and adding a source to Knowledge answers with the same sentence. It starts again on the next message once there are credits.

Limits and warnings. Two things stop the employee: your credits, and the credit budget you set for the brand. The daily and weekly limits under Agents › Office do not stop it unless you switch that on, under Spending on its profile. Those limits were made for agents that work on their own schedule, and a busy desk would reach a typical daily limit before lunch – a desk your customers are waiting on should not stop there. While it is off, what the employee spends is not counted towards those limits either, so it cannot use up what the brand's other agents are working within. Switch it on and the employee stops with them when a limit is reached; the page says exactly that before it changes. Under those two limits on Agents › Office, one line says which of the two is in force for the support employee, with a button to its profile – so nobody sets a ceiling believing the desk is inside it when it is not.

You are told, in the bell at the top of the page, when:

it has stopped – no credits, the brand's budget, or (if you switched that on) an Office limit. Once a day per reason, with the reason in its own words.
it has used 80% of an Office limit, when it is held to one, with the numbers and both ways out: raise the limit, or take it off the limits.
the credits left will not last it long at the pace this desk is working, worked out from what its replies have cost in the last day.
it is working again – the first piece of work it is charged for after a day it stopped on, saying what had stopped it. It is said once; if it stops again later, you are told again.

Each of these is said once a day at most. Turning alerts off under Agents › Office turns them off here too.

On the desk itself. While the employee is stopped for any of those money reasons, one line sits across the top of /helpdesk, above the queue and the conversation, saying which it is. It is there for the people with customers waiting rather than for whoever set the agents up: nobody has to open the Agents page to find out why no proposals are arriving. Anyone who can open the Agents page gets a button to the employee's profile beside it; a support seat, who cannot, is told to tell an administrator. The line disappears on its own as soon as the next piece of work is charged.

What a model sees

Before the text of a conversation reaches a model, the identifiers we know for the customer – name, e-mail address, phone number, customer number – and anything in the text that looks like an e-mail address, a phone number, a bank account or card number, a national identification number or a document number are replaced with placeholders. The values are put back on our servers when the draft returns. This is automatic and, for free text, best effort: the content of the message itself is still read.

A message that asks for self-exclusion, signals harm, mentions a chargeback or threatens legal action is never sent to a model. It goes to a person, with a short holding reply prepared for them to review. On a chat, the whole conversation is then kept from models until a person allows it again – see Chat conversations.

Since 2 November 2026 this check reads every content language you can choose for a brand (Brand → Profile → Market): English, German, French, Spanish, Czech, Slovak, Swedish, Polish, Romanian, Slovenian, Arabic, Chinese, Russian and Japanese. Every message is checked in all of them at once, whatever your brand's own language, so a customer who writes in another language is caught too, with or without accents. The holding reply is prepared in the customer's language when it is known, otherwise in the language they wrote in or in your brand's. A new content language always comes with this check: it cannot be added without it.

The details a chat visitor's browser reports – their page, device, language and time zone – and the address they typed for replies never reach a model.

Photos and PDFs a chat visitor attaches are read only once your organisation has confirmed version 1.3 or later of the AI Support Agent Addendum – reading them is what 1.3 adds – and only on brands where the employee works. In Settings › Agents the card says so beside the newer version: Version 1.3 lets the employee read the photos and PDFs visitors attach in Site chat; until then it only knows that something was attached. An administrator confirms it there with the button that names the newest version, Confirm version 1.4 since 12 October 2026, which covers this too; everything you confirmed before keeps working either way. Until then a model is told only that a photo or a PDF came – nothing of its name or content – and where the answer depends on it, the reply says a colleague will look at it. Once 1.3 is confirmed, a photo is read as it is, by a model that reads images: it is converted first, which drops what a camera stores with it, such as where it was taken, but a name or a number in the picture cannot be replaced with a placeholder, and the model is told never to copy a personal detail it can only see in a picture into a reply. A photo cannot be checked for a regulated topic before a model reads it, so the model is told to stop at once and write nothing when a photo shows self-exclusion, signs of harm, a chargeback or legal action; the AI is then paused on the conversation exactly as for such a message – best effort, since it depends on the model recognising the picture. HEIC photos are kept and shown to you but never read by the employee: it is told only that a photo came that it cannot see. A PDF is read by its text, which is treated like a message: identifiers replaced, and a regulated topic in it pauses the AI on the conversation like one in a message would. A PDF is read once: the text read from it is kept with the file and deleted with it. Nothing attached is read on a conversation where the AI is paused. A reply that read a file costs what any reply costs.

Photos and PDFs a customer e-mails are read the same way, with the same rules, only once your organisation has confirmed version 1.4 – reading them is what 1.4 adds. The card in Settings › Agents says so beside the newer version, and an administrator confirms it there with Confirm version 1.4; everything you confirmed before keeps working either way. Until then a model is told only that a file came with the e-mail (see Files and photos in e-mails under Where conversations come from).

What is kept

On brands where the agent is on, what the AI drafted is kept alongside what your team actually sent, so you can see how often the two agree. These records are kept for 24 months unless an administrator sets a different period in the card. The text read from a PDF a chat visitor attached is kept with the file, so it is read only once, and is deleted with the file and its conversation. Discard on a draft records that it was rejected. An administrator can withdraw the confirmation at any time, which stops the agent at once, and can delete what it has kept.

An optional closing sentence

If you want customers to be told that an assistant answered, the employee's profile on the Agents page has an optional closing sentence that is added to the replies it sends on its own. It is off by default and the wording is yours, per language – nothing is added until you write it and switch it on.

Knowledge

The AI answers from what you give it. Knowledge is where you give it: documents and pages, on the Helpdesk itself – the link sits beside Reporting at the bottom of the views list. Everyone who works the desk can open it, a Support seat included, because the people answering all day are the people who know what is missing.

Knowledge needs the support employee switched on for this brand: an administrator confirms the addendum in Settings › Agents and hires the employee for the brand on the Agents page. Until that is done, the page says so and offers nothing to add – nothing would be read and nothing kept.

What to add. Start with what your team answers from every day: your terms and conditions, the rules of a running promotion, the payment methods and their limits, the steps to verify an account, and the internal FAQ the team keeps for itself. A document is a PDF, Word, PowerPoint, Excel, CSV or text file up to 10 MB; a page is an address on your own site. Each source is read once, cut into passages, and searched when a customer writes. Reading a source costs 2 credits, and a page that has changed costs 1 when it is read again the following week – see What it costs above.

Who may hear it. Every source is marked, and nothing is marked for you – you choose before you choose the file:

For customers – the employee may say it to a customer, including in a reply it sends on its own.
Internal – it helps a colleague handle the case. It can shape a draft a person reads before sending, and it is marked as internal on the desk. It is never quoted to a customer and never used in an unattended reply.

You can change the audience later. Every passage of that source follows the change, and the page says exactly that before it happens.

Valid until. A source can carry an end date – the Friday a promotion closes. After that day its passages stop being used in an answer and the source reads Expired in the list. Leave it empty when it has no end. A confidently wrong answer about a bonus that ended last week is the mistake this prevents.

Pages are read again every week, so a help-centre article you change follows on its own. Read again re-reads a source on the spot, Pause keeps it without using it, and Remove takes it and its passages away.

Ask it. The box at the top answers one question: what would the employee find for this? Type what a customer would write and you get the passages it would work from – each with its source, where in it, whether it is for customers or internal, and how close the match is. Nothing is sent to a model and nothing is answered; it is how you see what the desk does and does not know. When nothing covers the question, it says so.

Every passage, one at a time. Open a source to read what it was cut into, how often each passage has been used and when it was last used. A passage that is wrong can be corrected in place – it is read again as you save it – and one that should not be used again can be retired, and put back later.

Questions it could not answer. Above the sources, Knowledge lists the questions customers asked that your knowledge did not cover. Repeat questions fold into one row, so the list reads asked 14 times · last 38m ago rather than fourteen rows, with up to three of the conversations it was asked in, one click away. Two buttons on each: Answer it opens a box where you write the answer in general terms – no customer names or numbers – choose who may hear it exactly as you do for a document, optionally give it an end date, and save. It becomes knowledge at once and the employee answers from it on the next message. When a colleague already answered that question in one of those conversations, their words are waiting in the box for you to edit into something general, with a link to the conversation they came from. Not needed takes the question off the list and nothing else – the conversations stay where they are.

It looks before it asks. When a customer's reply leaves the desk, the question it answered and the answer are kept together for this brand, with names, addresses, phone numbers and account numbers already replaced. The question is turned into a search vector for this – one small call, metered at a fraction of a credit – and it is how the employee can look at how your team has answered before writing to you. Two things come of it. A question it could not answer arrives with a proposal where one exists, and the line above the box says which it is: A colleague answered this in this conversation, or A colleague answered a similar question on 12 Sep, with the conversation one click away. And a draft a person will read may lean on one of those earlier replies where nothing you have written down covers the question – marked under the draft as From a colleague's reply on 12 Sep – unconfirmed, check it still applies. An earlier reply is not knowledge: nobody approved it, it may be about one customer's case and it may be out of date. So it never grounds a reply the employee sends on its own, and the question stays on the list until somebody answers it properly.

When the employee has found an answer by itself in what you have already given it, a second list appears, Proposed by the employee. A proposed answer can shape a draft a person reads before sending; it is never used in a reply the employee sends on its own until you keep it, which asks who may hear it in the same way. Retire puts it away.

Sources under a draft. A draft the AI wrote shows what it answered from, under the text: From Bonus terms.pdf · p. 3, with an internal source marked Internal – do not quote. Click a line to open that source and check the passage before you send. The same lines appear under a suggestion in the composer and under a reply the AI sent on its own. They are for the desk – the customer never sees them.

One thing it does not do: it does not go looking for information outside what you have given it – your documents, your pages, how your own team has answered before, and the lookups you define below. It never searches the web. It answers from what is here, writes down what it could not answer, and those questions are yours to fill in.

Training: showing it how something is done

Knowledge is what you can write down. Some of what your desk knows is not written down anywhere – it is a path through a screen that somebody learned by watching a colleague do it. Training is how you show the employee that: record your screen while you do the thing, or drop in a few screenshots, and it writes down what it understood as steps for you to check.

It is on the employee's own profile, under Agents → the Autonomous Customer Support Employee → Training. Everything there is per brand, like the rest of its work.

Be honest about where this is today. What it understands is written down, versioned and kept, and it is not used in a reply or in a lookup yet. This first step exists so you can see what it makes of your own recordings before it is allowed to act on them. Nothing on the page pretends otherwise.

What to record, and how

One thing per scenario. Checking a payment's status is a scenario; everything in the back office is not. A recording that shows one path from start to finish teaches more than a long one that wanders.

Talk while you do it. Say what you are clicking and why, and say what you are reading on each screen and what it means. The narration is what turns a sequence of clicks into steps somebody else could follow. A silent recording is the single most common reason a study comes back with nothing.
Use a test customer. Before anything is uploaded you tick a box saying so, and it is not a formality: what is on the screen is read by an AI model, and a picture cannot be redacted the way text can. Text we send to a model has names, addresses, phone numbers and customer numbers replaced with placeholders first. A screenshot has no such thing – whatever is on it is what the model sees. So record with a test account, and if your systems have no test account, blur or cover the customer's details before you record.
Say what it shows in one sentence when you upload it: how I find a customer's last payment in the back office and what I read there. It is what the employee is told it is watching.
Two kinds: where to look something up (it learns the path through one of your systems) and how a case is handled (it learns what you do and what you tell the customer).

Limits, which the page also shows you: a recording of up to ten minutes and 200 MB, as mp4, mov or webm; or up to thirty screenshots of 10 MB each, as png, jpg or webp.

What it costs

What it studied	Credits
A screen recording	5 for every started minute, at least 10
A set of screenshots	1 each, at least 5

A new version of a scenario costs the same as the first one. Three things follow:

Nothing is charged when you upload. The charge is made when the study runs, on the recording's real length, so a file that turns out to be too long or the wrong kind never costs you anything.
A study that fails is refunded in full, and the page says in plain words what went wrong. Try again puts it back in the queue; that attempt is charged when it runs, and given back in full if it fails again too.
Reviewing, editing, answering its questions, approving and discarding cost nothing. That is you checking its work.

Reading back what it understood

When the study is done the scenario says Ready for you to review, and the review is the heart of it. On the left, the steps it wrote – each one an action, the detail behind it, the moment in the recording it was seen, and the frame it read it from. Select a step and that frame is shown large on the right.

Everything is editable, because what you approve is what is kept: the title, the one-line summary, when it applies, the system and where it starts, every step (you can rewrite them, add one, remove one and reorder them), the lines under It must never, and – for a handling scenario – what the customer is told. A frame you would rather not keep has a Remove this frame button; the picture is deleted there and then.

It also asks about whatever it was unsure of, and each question has an answer box. You can approve without answering them; the page tells you how many are still open.

Approve closes the version. From that moment it is the version the employee was trained on, it is never edited again – a change is a new version – and your recording is deleted. The frames you kept stay, because they are what the steps are checked against. Discard keeps nothing: the recording and its frames are both deleted.

Versions

Every study is a version, so what the employee was working from on any given day can be read back. The list shows each one newest first, with its state, who taught or approved it, when, and what it cost. An approved version opens read-only. If a newer version turned out to be wrong, an earlier approved one has a Make this the current version button – nothing is deleted and nothing is re-studied.

Teach a new version when the screen or the procedure changes: same title, same kind, new recording.

Before you can teach it

Training reads pictures of your systems with a model, which the earlier version of the AI Support Agent Addendum did not cover. So two things have to be true before the tab shows anything: the employee is hired for this brand, and your organisation has confirmed version 1.2 of the addendum. What you agreed to before keeps working; 1.2 is the version that adds training.

Until then the tab does not just say no. It lists what is done, with who did it and when, and gives the one thing to do next its own button – open the agreement, or open Agents and hire it. Anyone who cannot do that step themselves is told who can.

Lookups

Knowledge answers questions about your business. Lookups answer questions about one customer: where is my order, did my payment arrive, what address is on my account. Those answers are not in a document – they are on a screen in one of your own systems. A lookup is the employee opening that screen in a browser, reading it, and reporting what it found.

It exists for the systems we cannot connect to. Where there is a connection, use it. Where there is only a back office somebody signs into, this is the door.

You decide what may be looked up. The employee never browses freely. On the desk, Helpdesk → Lookups, your team defines each one: a name, the system it is in, the address it starts at, how it finds the customer (by customer number, e-mail address, phone number or name), and what to do and read once it is there – in plain words, as steps. Each one has its own switch, so a lookup can be turned off without being deleted. Support seats work here, because the people answering all day are the people who know which screen holds the answer.

Write the instructions the way you would tell a new colleague: search for the customer with the number in the box at the top; open their record and go to Orders; read the newest order – its number, the date, the amount and the status; report those four things and nothing else. A lookup that asks for one thing gives a better answer than one that tries to read a whole account.

It only reads. Every lookup is told never to save, approve, refund, send or change anything, and to stop and report that it was blocked the moment the only way forward would change something. Make that real at your end as well: give the account it runs under read-only rights. An instruction is a rule the browser follows; permissions are a rule it cannot break.

What it does depends on the stage the employee is at (the same three stages as everywhere else, on its profile under Agents):

Learning. It says which lookup the question needs and stops there. The conversation's rail shows It would look this up: Order status with a Look it up button. Nothing runs until a person presses it.
Copilot. It runs the lookup itself, and when the answer is back it writes the proposal with the fact in it, for a person to read and send.
On its own. It runs the lookup itself and does not wait for anybody. It answers on its own only when the topic is one you switched on and the sender's identity is verified – somebody signed in, not somebody who typed an address into an e-mail. For anyone else the reply with the fact waits in the thread as a proposal. An account fact is exactly the thing that must not be read out to whoever claims to be that person.

Anybody working the desk can also press Look it up on any conversation at any stage, and choose which lookup to run. If the conversation carries nothing to find the customer by – no customer number, when that is how the lookup finds them – the press is refused with that sentence, and nothing is charged.

Where it runs. In a browser, under a real login, because that is what these systems need:

A Helpdesk open in a browser with our Chrome extension claims the lookup and runs it there, in a window of its own that nobody is working in, under that person's own logins.
If no such browser takes it within about three minutes, and the brand has a stored login for that system, it gets one attempt in the cloud browser.
If neither is available, the run ends and the desk says so plainly: nobody's browser with the extension is open. The employee never guesses the answer.

So a desk where somebody has the extension installed and the Helpdesk open is a desk where lookups happen while people work. Keeping a login stored for the system is what covers the nights.

What it read is for the desk. The result appears on the conversation, in the same note paper as an internal note, with the system it came from, the time and whose browser it ran in. It goes with the conversation when the conversation is deleted, and it reaches the customer only through a reply a person or the employee writes.

What it costs. A lookup costs 5 credits. A run that ends without an answer – it could not get in, no browser was open, you stopped it – is not charged, and one that was already charged is refunded in full.

It needs its own permission. Lookups are off until an administrator confirms the separate Account Lookups Addendum in Settings › Agents and switches lookups on for the brand. It is a second confirmation on top of the first one, and it is asked for separately on purpose: this is the one part of the employee's work where a model reads a page belonging to a named customer, with their real number or address in hand. Until both are done, the Lookups view says so and offers nothing else.

Ratings

Asking customers to rate a conversation is optional, and it is off for every brand until you turn it on. Each brand decides for itself: in Connections › Support inbox, switch on Ask customers to rate the conversation. The same switch is in the Site chat studio, under Behaviour → Ratings; it is one setting per brand, so turning it on or off in either place changes both. It saves at once – in the studio it does not wait for Publish. Anyone who works on the brand can change it; a Support or Data engineer seat cannot.

While it is on, resolving a conversation asks one question, 1–5, in the customer's language: by e-mail with one-click rating links for an e-mail conversation, and in the chat window for a site chat. One ask per conversation, ever. Other channels (social messages, store reviews, tickets logged by hand) are never asked.

While it is off, nobody is asked: no rating e-mail goes out and the chat shows no rating box. Turning it off keeps every rating already given, and a customer who clicks a link in a rating e-mail sent earlier still gets their rating recorded and a thank-you page, never an error.

Scores appear in the rail and in Reporting under Satisfaction, and flow into your warehouse alongside spend and lifecycle data – a chat rating only when the visitor was signed in and had allowed marketing on your site. With ratings off and none given, Reporting says "Ratings are off for this brand. You can turn them on in Connections." instead of an empty number; averages from before stay visible.

Support seats

Invite teammates with the Support role (Organization → Team). A support seat sees only the Helpdesk – no campaigns, data, or settings – which makes it safe to staff the queue without opening the rest of your workspace. A new support seat takes tickets for the brands you give it; you can change that in its Customer support section, and the seat can start or stop taking a brand's tickets itself from the Helpdesk. It has its own notification bell, and every link in it opens the Helpdesk.

A support seat can log tickets, find a customer and attach them to a conversation, and sees the customer card beside every conversation. It finds a customer by their exact e-mail address, phone number or customer number – the whole of it, never a name or part of one – so the seat reaches the person it was told about rather than browsing your customer list. Case does not matter, and a phone number matches however it is written (spaces, dashes, + or 00), as long as it has the same digits as the record: if the record has a country code, type it too.

Lookups also go at a person's pace. A support seat that finds dozens of customers within a few minutes, or several hundred in a day, sees the search paused for a while with a note saying when to try again. Someone answering calls never reaches that; searches that find nobody do not count toward it.

It cannot create a customer record – that stays with owners, admins and managers, like everywhere else in the product. A support agent who meets someone new logs the call with their name and number, and whoever manages customers creates the record afterwards.

The card never offers a support seat a door it cannot open. Grant a promo is there (granting a gesture is part of the job), but Open in Customers is not, because the seat stays out of the customer pages. When a brand has no active promo, the seat reads that an administrator can create one under Promos, and when a store refuses a saved key on a review, that an administrator can replace it on the Connections page.

Instagram, Facebook, YouTube and Reddit comments and messages

With Instagram, Facebook, YouTube or Reddit connected, comments under your posts land in the Helpdesk as conversations, next to your e-mail, and so do direct messages on Instagram, Facebook and Reddit. Each thread names the person by their handle, and a comment thread shows the post it belongs to with a link. Reply under the comment; on Instagram and Facebook you can also switch to "In a DM" to answer the commenter privately. A message thread answers in the message thread. Meta allows support answers outside a 24-hour window, not marketing; Reddit chats are text only.

Listening on Reddit. Press Listening on the Helpdesk and give it the phrases people use when they look for what you sell, and optionally the communities to watch. Every couple of hours new threads that match land in the Helpdesk as "Reddit thread" conversations, with the thread as the message and a drafted reply from your Social Media Manager. Nothing is ever posted on its own – you send it, edit it, or leave it – and a sent reply appears as a public comment from your account. Check now runs the search on the spot.

If your AI team has a Social Media Manager, it drafts the reply in your voice the moment a thread arrives – send it, edit it, or write your own. When the team is trusted to publish on its own, it answers on its own too, with the same limits as e-mail: sensitive requests are handed to a person, and after three answers in a row a person takes over. In chat, "what's waiting on Instagram?" lists the threads, and "reply to Lena that we do support Shopify" sends the reply you approve. Comment-to-DM automations ("comment LINK and I'll send it") can be set from chat as well.

App Store and Google Play reviews

With App Store Connect or Google Play connected (Integrations → App stores), the reviews of the brand's apps land in the Helpdesk under Social – one conversation per review, next to your e-mail. The apps are the ones saved for the brand that the connection covers – saved on App Store Optimization, in Brand → Profile, or when you connect the store. The first read brings in the last 7 days; after that the stores are read every six hours, and Check for new reviews on a review reads them now. When a reviewer changes their review – new stars or new words – it arrives in the same conversation.

Each conversation is named for its stars and store ("2★ App Store review: Crashes on login"). The details beside it show the reviewer's stars, the app, and the version, territory and device where the store gives them, and whether it has been answered. A reviewer is a store nickname, never matched to a customer: the reply under the review is the only way to reach them. A review already answered on the store comes in resolved, with your answer on it.

The reply is public, under the review, and a person sends it with Post the reply. The AI never answers a store review on its own – on these conversations the AI setting stops at Draft. A new review of 3 stars or fewer rings the bell, and when your support employee drafts, it arrives with a draft waiting; reviews of 4 and 5 stars are filed quietly, with no bell and no draft.

Google Play takes at most 350 characters. The count shows as you type, and the reply cannot be sent until it fits.
The App Store checks a reply before it shows it, so it reads Sent – not public yet until then. Sending again replaces your earlier reply.

In chat, a reply you ask for to a store review is left as a draft in its conversation for a person to send. If the store stops accepting the key, the Integrations card says so with Replace the key, and the brand's bell rings once. Three failed reads in a row ring it too, and the app's row in Brand → Profile says what went wrong.

App Store Optimization reads the stores' recent reviews as well, for what the listing itself should answer.

Helpdesk

Where conversations come from

Several brands, one desk

Reporting by hour

Logging a call or a visit

Working the queue

The keyboard

Tags

Chat conversations

The customer card

AI in three modes

Hiring the support employee

What it costs

What a model sees

What is kept

An optional closing sentence

Knowledge

Training: showing it how something is done

What to record, and how

What it costs

Reading back what it understood

Versions

Before you can teach it

Lookups

Ratings

Support seats

Instagram, Facebook, YouTube and Reddit comments and messages

App Store and Google Play reviews