Squawk Co-pilot

Squawk Co-pilot is the AI assistant built into Studio — you’ll see it as the Co-pilot button in the top bar, and its panel is headed Build with AI. Describe what you want in plain English and the Co-pilot does the building for you: it creates and configures AI agents, links knowledge, sets up actions and flows, and answers questions about how the platform works, from the product documentation, along the way.

Squawk Co-pilot is now generally available to all Studio users — no beta enrollment or plan add-on required.

Open the Co-pilot

Three ways:

Click Co-pilot at the top right of any page except Home and the full-screen chat. When the top bar is narrower than about 1,180 pixels — a small window, or a larger one with the Co-pilot panel open beside the page — the button shows only its icon.
From any page, type your request into the search field at the top of the screen — Ask us anything, or jump to a page… (on a narrow top bar it shortens to Ask us) — or press Cmd/Ctrl + K to jump to it. Your request appears as an Ask AI about … row at the top of the results; press Enter on it. If a conversation is already open, your request is added to it; the Co-pilot starts a new conversation only when it is empty. With the field empty, Cmd/Ctrl + K also lists your Recent chats; choosing one opens it full screen, even from a page where the Co-pilot would normally dock. The Build with AI result under Actions also opens the full-screen chat.
On the Home page, type into the box in the middle of the page. Once you’ve had a conversation, a Threads column beside it lists your past conversations, most recently active first — click one to reopen it, or choose New thread to start over.

Asked from Home, the conversation opens full screen. Asked from any other page, the Co-pilot opens as a resizable panel beside it, so you can keep working while you chat — use the expand icon in the panel’s header to move the conversation to full screen. The conversation follows you between the two. From full-screen chat, following a link to any other page — a result card’s link, or an item in the sidebar — continues the conversation in the panel. Going to Home while a conversation is open in the panel takes you to that conversation in full screen instead of the Home overview. Cmd/Ctrl + K replaces the old Cmd/Ctrl + I shortcut. The panel stays open while you test: Talk and Chat on an agent’s page leave it open, so you can try a change next to the conversation that asked for it. The test chat opens beside the panel, narrowing to fit on smaller screens. A Talk test call covers the whole window until you hang up, and your conversation is still there afterwards. Use New Chat to start fresh, or Threads to open your previous conversations beside the current one — sessions are saved per workspace. Threads lists only your own conversations in the current workspace: your 200 most recent, most recently active first, each titled from its first message. Reopening a past conversation, whether from Threads or by closing and reopening the panel, only displays it — it doesn’t change any agent. In full screen, the conversation you’re in is kept in the page’s address, so reloading the page reopens it instead of an empty chat, and you can bookmark the address. Conversations are private to the person who had them, so the address reopens a conversation only for you. If a conversation can’t be loaded within 15 seconds, a new chat opens with a message telling you to pick the conversation from Threads to try again. The docked panel doesn’t do this. A new, empty conversation, in the panel or full screen, opens under How can I help you today? with the same four suggestions as Home: Build, Analyze, Fix and Reach out. On Home and in full screen, choosing one sends it straight away; in the docked panel, it puts a ready-written request in the message box for you to edit and send. Opening the panel with the Co-pilot button, or starting a new chat in the panel or full screen — with New Chat, New chat or New thread — puts the cursor straight in the message box.

What it can do

Ask for an outcome, not a procedure — “Design a renewals agent for our SaaS product”, “Link our product FAQs”, “Add an order-lookup step to this flow”. The Co-pilot asks follow-up questions where it needs detail; once your request contains everything it needs, it gets to work.

Area	What the Co-pilot can do
AI Agents	Create new agents from a plain-English description — name, greeting, and guidelines drafted for you — and update existing ones: the greeting and sign-off for voice and chat, the email sign-off, guidelines, voice and voice tuning, outcome options, and settings such as languages, Transcription Keywords and timezone. It changes only what you ask for — setting outcome options leaves the agent’s other handoff settings alone, and adjusting one voice tuning value keeps the others. It can also duplicate an agent, with its actions, flows and knowledge links; the copy is named Copy of followed by the original name.
Knowledge	Create knowledge bases, crawl a website into one, add FAQs, and link knowledge to an agent — or unlink it again, which detaches the knowledge base from that agent without deleting anything, so it needs no confirmation. A crawl started from the chat fetches up to 10 pages (5 unless you ask for more); for a larger crawl, set the Limit in Crawl Options on the Webscrapes tab yourself — see Knowledge Base. The Co-pilot can rename a crawled website, but it can’t re-crawl it or change its address: to run it again, use Recrawl (Rescrape for a single page) in the website’s menu on the Webscrapes tab once it has finished, or delete it and add the new address. Attach documents directly in the chat with the paperclip — a file waits in the message box and is uploaded and indexed only when you send. It goes into the knowledge base you have open; otherwise into the first knowledge base linked to the agent you have open. If that agent has no knowledge base yet, a new one named after the agent (for example, Renewals Agent Knowledge Base) is created for the file and the Co-pilot links it to the agent; if the agent already has knowledge bases, the Co-pilot asks before linking any other one to it. With neither an agent nor a knowledge base open, you’re told it will go into a new, unlinked knowledge base. The Co-pilot can read the text of an attached .txt or .md file (up to its first 12,000 characters, and it says when a file was cut short), but not .pdf, .doc, .docx or .csv attachments.
Actions & Flows	Create and update pre-conversation, in-conversation, and post-conversation actions. When it creates an in-conversation action, it also adds a short briefing about that action to the agent’s Guidelines, which appears in the Guidelines editor straight away. It edits in-conversation flows node by node — adding, updating, and removing steps and branches — and the flow editor refreshes in real time as it works.
Handoff	Read how an agent currently hands a conversation to a person, and change it. For voice, that is where a call transfers — a phone number or a SIP URI, including a `{{variable}}` filled in at call time — and the options on the same Transfer Rule card: Enable Handoff Transcript and Enable Handoff Recording for a phone number, Screen Pop and Inline SIP context headers for a SIP URI. It can also set what the agent says just before it hands off, separately for voice, chat and email (and per language on a multilingual agent), and choose which connected helpdesk chat and email escalations go to. It checks the current settings first and changes only what you ask for. Changing the voice transfer settings or the helpdesk connection asks you to confirm first; changing the wording doesn’t. See AI Agents, Chat Escalation and SIP Integration.
Campaigns	Set up an outbound campaign — agent, contacts, and calling window — check its hours against local dialing rules, and launch, pause, resume, or stop it. Ask how a campaign is doing and it reports on the run — contacts reached, whether it is still dialing, and dial volume over time — or lists your campaigns by status. It can also tell you whether it’s legal to call right now in a timezone (and, for the US, a state) — no campaign needed. When you add contacts in the chat, it asks for each contact’s postcode up front, and for the date consent was obtained when you say you hold consent. Contacts go in up to 1,000 at a time, and only while the campaign isn’t live — in draft, stopped, completed or failed status. Uploading a CSV is still done on the Campaigns page. A voice campaign can’t launch until its agent has at least one phone number, so if it has none, ask the Co-pilot to link one first. When you ask it to launch, it first runs the same checks as Launch — agent, a phone number on the agent, calling window, SAN when scrubbing is on, account status and balance — and tells you what’s missing before it asks anything. Then it asks you to confirm that the campaign tells people they’re talking to an automated agent, and shows a card with how many contacts will be called, by which agent, and on which days — warning you if Do Not Call scrubbing is off. Nothing dials until you click Launch campaign. If scrubbing is on and your account has no Do Not Call SAN, give the Co-pilot the number from your National Do Not Call Registry subscription confirmation and it saves it after you click Save SAN on a card that names your account and any SAN it replaces. It never guesses or corrects a number. Removing a SAN stays a job for Remove SAN in your campaign account settings.
Phone Numbers	See which of your phone numbers answers for which agent, and point a number at an agent or detach it — both ask you to confirm first, and the card says which agent will stop receiving calls if the number was already in use. Linking a number to an agent also makes it one of the caller IDs its outbound campaigns rotate through, least recently used first, and detaching it takes it out of that rotation. Campaign calls only go out from numbers bought in the Studio.
Users	List the users in your account, invite a teammate (it asks you to confirm first, because a real email goes out), rename someone — yourself included, which changes your own Full Name — or remove them (asks you to confirm first and can’t be undone).
Variables & Transformers	Create agent variables and build, update or delete Data Transformers. Deleting one asks you to confirm first, and is refused while an agent still uses it — the Co-pilot tells you what’s using it.
Integrations	List available Connected-App integrations and their actions and add them to an agent. Tell it what the integration should do — “book and reschedule appointments” — and it picks the matching actions, adds them, and reports what it added. Asking what an integration can do just lists its actions — nothing is added until you describe a job and ask it to set it up. Integrations and their actions also appear as chips you can click instead of typing — see Voice cards, integration chips, charts and tables. It can also run a live test request against an action to check it works. If an integration isn’t connected yet, the reply carries a Connect … link to its connection dialog in Connected Apps. When you finish connecting from that link, the Co-pilot panel reopens on the same conversation, sends a message on your behalf (“I just connected …”), and carries on by listing the integration’s actions. If the connection doesn’t complete, it sends a message saying so and offers to help you retry. This only follows a link you clicked in the Co-pilot, and only if you come back within about 5 minutes.
Analytics	Answer questions about volume, outcomes, and dispositions across your voice, chat, and email conversations — replies can include charts and tables, not just text (see Voice cards, integration chips, charts and tables). On long date ranges, daily volume is grouped by week past about 45 days and by month past about 120 days, and the answer says so. With no period given, it answers for the last 7 days. A custom range in a cross-channel answer can span at most a year (366 days); for longer, ask for it in pieces. A chart or table by channel splits voice, chat and email.
History & audit	Find calls, chats and emails by date range, agent, outcome or search text, and read a transcript. You can also narrow the list by topic, sentiment, errors only, pinned only or dissatisfied callers. For any one conversation it can read the Action Logs — which of the agent’s actions ran, what each returned, and its status and how long it took, with failures only if you ask — the same record the Action Logs switch shows in Call History. It searches every channel you can access unless you name one, and “the last 7 days” means today and the six days before it. Ask what changed in your workspace, who changed something, or what a person created or deleted — “what changed” covers every kind of change, not only edits. See Audit Logs.
Product questions	Answer how-to and what-does-this-mean questions — what a setting is, where it is in the Studio, how to set something up, what changed in a release — from the product documentation rather than from general knowledge. While it looks, the chat shows Searching in the available knowledge, which becomes Searched available knowledge once it has; this is the product documentation, not your own Knowledge Bases, and the line isn’t saved with the conversation, so it doesn’t show when you open a past conversation from Threads or reload the page. If the documentation doesn’t cover your question, it says so instead of guessing. Questions about your own agents, calls, analytics and numbers are still answered from your account, and replies give the answer without links or citations.
Navigation	While you’re on an agent’s page or one of its sections, it leaves you where you are: its edits to that agent don’t move you to the agent’s overview, and a knowledge base it creates doesn’t take you away, whether you’re on an agent’s page, one of its sections or the agent’s Knowledge Base screen — the result card in the chat links to it. (From the in-conversation flow editor, those requests still open the page they relate to.) Creating or duplicating an agent opens it — except in full-screen chat, where you stay in the conversation and the result card’s link opens the agent. Deleting an agent returns you to the AI Agents list, and outside an agent it opens what it creates. The result card for an agent edit links to the card that changed and is named after it — a greeting or sign-off change opens Greeting & Sign-off on that channel’s Voice, Chat or Email tab — while an edit that changes several cards at once links to Agent. It also surfaces links (such as the Connected Apps connect dialog) when a task needs finishing by hand, and answers can carry result cards that link straight to the page they relate to, such as Campaigns or Call Analytics.

When you’re on an agent’s design page or inside a flow editor, the Co-pilot picks up that context automatically — “add a compliance disclaimer to the greeting” just works, without restating which agent you mean. Each time you send a message there, it reads that agent’s saved name, description, guidelines, greeting, voice and voice tuning, outcome options, the transfer message and rule for voice calls, and timezone — plus the saved flow when you’re in a flow editor. It only sees what has been saved, so after editing something by hand, give it a moment to autosave before you ask the Co-pilot to build on it. On a knowledge base’s own page, it also knows which knowledge base you have open — its name and description — so you don’t need to say which one you mean.

Building an agent end to end

For a from-scratch agent, the Co-pilot works through four stages, confirming direction with you at each step:

Mission & persona — it drafts the agent’s name, greeting, and guidelines from your description of the business goal.
Knowledge — it connects or creates the knowledge the agent should answer from.
Workflow automations — pre-conversation, in-conversation, and post-conversation actions for the tasks the agent must perform.
Handoff — outcome options and escalation behavior for the moments a human should take over.

What it won’t do

It can’t buy phone numbers or make billing changes — it will point you to the right page for those instead.

How changes are applied

The Co-pilot typically describes what it’s about to do and asks follow-up questions when details are missing — but if your message already contains everything it needs, it proceeds immediately, and each executed step appears in the chat as it runs. Each step’s result card has a plain-English title (for example “List assistants”) and a Details button that shows the technical record of the request and response. A badge on the card says how the step went: Action completed; Partly applied, when the request went through but some of the values you asked for were not saved — the Co-pilot tells you which; or Action failed. A card never ends up below the reply that describes it, live or when you reopen a conversation.
The line above a reply counts while the Co-pilot works — Thinking for 7s… — and settles on how long it took, such as Thought for 12s, when the reply starts. Select it to read the reasoning. The line and the reasoning behind it aren’t saved with the conversation: open a past conversation from Threads, or reload the page, and you see the replies without them.
Stopping a reply. While the Co-pilot is replying, the send arrow in the message box turns into a red square Stop button. Stopping ends the reply and adds Assistant response stopped. to the conversation. It doesn’t undo anything the Co-pilot had already changed.
Some requests stop and show a card that names the real campaign, agent, number, document, connection, or person involved, and nothing happens until you answer. The card has two buttons inside it: one named for the action (in parentheses below) and Cancel. The requests are: launching a campaign (Launch campaign), or pausing, resuming, or stopping one (Update campaign); saving your Do Not Call SAN (Save SAN); pointing a phone number at an agent (Link number) or detaching it (Detach number); changing the voice transfer settings, such as where calls go when a caller asks for a person, or which connected helpdesk chat and email escalations go to (Save handoff); changing which outcomes count as contained — which recalculates containment for past periods too (Update dispositions); inviting a teammate, because a real email goes out (Send invite); removing a user (Delete user); and deleting an agent, action, flow step or branch, variable, data transformer, knowledge base, FAQ, crawled website, or uploaded document (Delete agent, Delete action, Delete node for a flow step, Delete branch, Delete variable, Delete transformer, Delete knowledge base, Delete FAQ, Delete webscrape for a crawled website, or Delete document). See Answering a confirmation card for how the card works.
Everything else — creating and editing agents, knowledge, actions and flows, and adding integration actions — happens as soon as the Co-pilot has what it needs, with no approval step or undo. Be specific about exactly what you want before sending large requests.
Screens affected by a change refresh automatically — you can watch a flow update in the editor while the Co-pilot builds it, and an open Knowledge Base document list updates when the Co-pilot deletes a document. On an agent’s page, the card the Co-pilot changes pulses briefly: a greeting or sign-off change pulses Greeting & Sign-off, and when one step changes several cards, they all pulse. If the card that pulses is off screen, the page scrolls to it; when several cards pulse, it scrolls to the first. If the section being changed is already open — Guidelines, Variables or one of the Actions sections — a ring pulses around it. Choosing Use this voice on one of the Co-pilot’s voice suggestions pulses Voice Configuration.
Editors you have open follow the Co-pilot’s changes. An open Guidelines editor or Voice Configuration panel switches to the Co-pilot’s version as soon as it lands. In the Guidelines editor, that replaces anything you haven’t finished typing. On Agent Settings and Handoff, if you’re partway through an edit when the Co-pilot changes the same agent, both changes are kept; if you both change the same setting, yours is saved.
Everything it changes is recorded in Audit Logs under your user as the actor, with Squawk Copilot shown in the IP Address column, so you can always tell which changes came from the assistant.
It only touches workspaces and resources your user can already access, and credential fields (auth tokens, passwords, API-key fields) are redacted from the conversation and its saved history. As a general habit, avoid pasting raw secrets into the chat — connect credentials through Auth Profiles instead.

If you ask the Co-pilot to set an agent’s voice and it picks an Arabic voice, the agent is converted to Arabic — single-language mode, every other configured language dropped. The same rule applies to an agent import, a duplicate and a direct API update, because it is enforced by the platform rather than by the screen. Note that a voice change the Co-pilot makes by instruction switches the language silently, with no on-screen notice; only clicking a voice card in its panel shows the “Language switched to Arabic” confirmation. See The Arabic Language Lock for what the switch changes.

Asking the Co-pilot to find a voice by Arabic dialect works for fusha, najdi, hijazi and emirati. The other dialects in the catalogue are not recognised as search terms — filter by Provider in the voice catalogue instead.

Voice cards, integration chips, charts and tables

Some replies show something to look at or click, not only text. While the Co-pilot is still replying, the buttons and chips are greyed out and the row says Wait for assistant to finish….

Voice suggestions. When you ask for a voice, the reply shows a Suggested voices row of cards, usually six. Each card has a play button that previews the voice. The preview reads your agent’s greeting when you’re on the agent’s page, and otherwise a short line that starts “Hello! I am” followed by the voice’s name. Open the agent first, then choose Use this voice to apply it: a Voice changed to message with the voice’s name confirms it, and the card for the voice the agent already uses reads Selected. The Co-pilot can narrow voices by accent, gender or a search word, but not by provider or language — use the Provider filter in the voice catalogue for those.
Integration and action chips. When the Co-pilot lists the integrations available to your account, each one appears as a chip under Available integrations, which also says how many are showing, such as 1–6 of 12. Point at a chip to read its description. A Connected badge marks an integration that is already connected. Choosing a connected integration asks the Co-pilot for its actions; choosing one that isn’t connected asks it to help you connect it and give you the connect link. Its actions then appear as chips under a heading made of the integration’s name and actions, and an action already on the agent shows an Added badge and can’t be chosen again. Choosing an action asks the Co-pilot to add it to the agent you have open as an in-conversation action, which is the same as typing that request. When there are more than fit on one page, a Show next … chip brings up the next set. If you’ve described what you want the integration to do, the Co-pilot picks and adds the actions itself and shows no chips.
Charts. For questions about volume and outcomes, the reply can include a line, area, bar, horizontal bar or pie chart. A chart with several series can be stacked, so the height shows the total. If a chart can’t be drawn, a short notice takes its place: Couldn’t draw this chart: none of the rows had a number to plot., or that the data it refers to isn’t in this conversation anymore.
Tables. For lists such as calls, chats, emails or audit entries, the reply can show a table of up to six columns. It starts with 20 rows. When there are more, a line under the table says so — for example Showing 20 of 87 — ask for more to see the rest. A cell longer than 60 characters is cut off with an ellipsis.

Answering a confirmation card

A confirmation card has two buttons, inside the card: one named for the action — Delete agent, Launch campaign or Save SAN, for example — and Cancel. They appear once the Co-pilot has finished its reply. Clicking the action button runs exactly the action the card shows. Cancel declines it: nothing runs, No, don’t do that. is posted as your reply, and the conversation carries on, so you can ask for something else. A card for something that reaches outside Studio — launching, pausing, resuming or stopping a campaign, saving a SAN, linking or detaching a number, inviting a teammate, or changing voice transfer or the helpdesk connection — also carries the note This affects something outside the platform. next to the buttons. A card for a handoff change has a Details button that lists what is about to change, field by field.

Answered once. A card can be answered once, and only by your very next message. It stays valid for 60 minutes, after which the Co-pilot asks again.
Typing a reply. If you answer by typing instead of clicking, only a plain yes counts — for example yes, ok, sure, go ahead, do it or confirm, optionally followed by please or thanks. A question mark, or anything extra such as “yes, but call it X”, counts as not approved and nothing runs.
Several cards waiting. A typed yes runs nothing, so use the buttons on the card you mean. When you answer one card with its button, any other card still waiting is shown again.
After you answer. The card’s buttons are replaced by Approved, Declined or, for a typed reply, Answered.
After a reload. A card you never answered and come back to after a reload or a thread switch reads “This request has expired. Ask again if you still want it.” Ask the Co-pilot again.

Tips

One goal per chat. Long conversations eventually trim older context; start a New Chat when you switch topics so the Co-pilot stays sharp.
Big asks are fine — expect stages. Very large requests are handled in stages across a few turns rather than one giant step. The Co-pilot usually summarizes how far it got — if it stops early, just ask it to continue.
Attach source material. Dropping a product sheet or policy document into the chat is usually faster than describing its contents. The file is held until you send, then goes into a knowledge base as described under Knowledge above — with no agent or knowledge base open, into a new, unlinked one, after a notice. The paperclip takes .pdf, .doc, .docx, .txt, .csv and .md files, and the Co-pilot can read the text of .txt and .md files (the first 12,000 characters) but not .pdf, .doc, .docx or .csv. If a file can’t be uploaded, nothing is sent: a notice names the file, whatever that send had already created is removed again, and your message and files stay in the box so you can try again. The Home box has no paperclip — open the conversation first, then attach.
Sending. Press Enter to send your message and Shift + Enter to start a new line.
If a reply stalls. Click the red square Stop button, which replaces the send arrow while the Co-pilot is replying, to end a reply yourself. The Co-pilot also stops a reply on its own after about 90 seconds of silence or about 4 minutes in all, saves the conversation, and says so — ask it to continue. Anything it had already changed is kept, whether it stopped or you did. If an error message appears under your request — for example “Lost connection to the assistant.”, “The assistant took too long to respond.”, “The assistant stopped before it answered. Try again.” or “Sorry, something went wrong. Please try again shortly.” — the conversation is saved, and Try again under the message sends your last request again.
Asking support about a conversation. Once a conversation has started, a Session pill showing the first characters of its ID sits in the header of the panel and of full-screen chat. Click it to copy the full session ID, and include it when you contact support.

Two smaller AI helpers share the Co-pilot lineage on specific screens:

Guidelines Assistant — in the Guidelines editor (open the agent’s Guidelines card), the wand button (Generate Guidelines) opens a side pane titled Guidelines Assistant that drafts and refines the agent’s instructions with you. Its edits appear in a Preview Changes overlay, and nothing in your guidelines changes until you select Accept (Reject discards them). Close Chat closes it. See Draft with the Guidelines Assistant.
Generate Action with AI — while you create an in-conversation action (not when you edit one), the sparkle button in the header of the Create In-conversation Action drawer (tooltip Generate) opens a box where you describe what the action should do. Studio fills in the action for you to review, and nothing is saved until you save it. See Actions.

Get Started

Build

Campaigns

Optimize

Manage

Best Practices

Library

Telephony

Product

Trust & Security

Open the Co-pilot

What it can do

Building an agent end to end

What it won’t do

How changes are applied

Voice cards, integration chips, charts and tables

Answering a confirmation card

Tips

Where to go next

AI Agents

Actions

​Open the Co-pilot

​What it can do

​Building an agent end to end

​What it won’t do

​How changes are applied

​Voice cards, integration chips, charts and tables

​Answering a confirmation card

​Tips

​Related assistants

​Where to go next

AI Agents

Actions

Open the Co-pilot

What it can do

Building an agent end to end

What it won’t do

How changes are applied

Voice cards, integration chips, charts and tables

Answering a confirmation card

Tips

Related assistants

Where to go next