AI teammates: how connected AI agents work in your team

Updated Nov 20, 202616 min read
English
On this page

An AI teammate is an AI agent you connect to Deskwoot yourself, such as Claude, ChatGPT, Grok Bot or Cursor, that works in your team like a colleague. It connects through Deskwoot's MCP server, gets its own name, works the conversations you give it and follows the rules you set. A person approves what it may not send alone, and an administrator can pause or stop it at any time under AI teammates. Fynn, the AI agent built into Deskwoot, is not one of them: he answers your live chat on his own.

Connect an AI agent

Administrators find AI teammates in the rail on the left. Connect an AI agent at the top of the page asks where your agent runs (Grok Bot, Claude, ChatGPT, OpenAI Agents SDK, Cursor or Another MCP client) and shows the steps for it. Most apps sign in with OAuth: you add Deskwoot's address https://deskwoot.com/api/mcp in the app, sign in to Deskwoot and choose on the consent page what the agent may do. Your own code gets a static token instead, which only the workspace owner creates. Every step is in Connect an AI agent with the MCP server, and Grok Bot has its own guide, Connect Grok Bot to Deskwoot.

  • Before the first agent, an administrator accepts the Data Processing Agreement for AI agents once, in the first step of Connect an AI agent.
  • A new AI teammate starts working only once the workspace owner connects or approves it. One the workspace owner connects works at once. When another administrator or a member connects a new one, they choose Request approval on the consent page, and it waits under Requests for AI agents on AI teammates until the workspace owner approves it. Until then it cannot read or do anything. Other administrators see it there as Waits for the workspace owner and can decline it, and the person who asked can withdraw it. Only the workspace owner connects an existing AI teammate again and creates static tokens and Custom Webhook bots.
  • While Members may connect personal assistants and the MCP server are on, a member's or an administrator's own personal assistant connects at once. While the switch is off, it waits for the workspace owner's approval too. Only the workspace owner changes this switch; everyone else sees it locked.
  • A connection that works at once, or an approved request, also switches on the MCP server at the top of AI teammates. While it is off, no agent of the workspace can work.

Roles

On the consent page you choose how the agent works:

  • As its own AI teammate (recommended): it gets its own name in your team, works conversations under that name, writes notes and drafts and, if you allow it, replies to customers. Its role shows as Responder.
  • Specialist, under More options: it reads conversations, writes internal notes and mentions people. It never replies to customers.
  • As your personal assistant: it works in your name, as "your name via the app", with your role and your access, and sees only what you see. Everything it does counts as done by you and shows as done through its app.
  • To set up Deskwoot for you, for administrators: a setup agent that changes your settings in your name and never reads conversations or contacts. See Let your AI agent set up Deskwoot.

An agent's messages to customers are marked as written by AI, in the customer's language: in the live chat with an AI badge next to its name, under an email with the sentence "This reply was written by an AI assistant." and on WhatsApp and the other messaging channels with a line above its first reply that names it as the AI assistant of your workspace. Only the workspace owner can switch this label off for an agent, with AI label on its replies on its page. Your workspace must then tell customers that they are talking to an AI wherever the law requires it. On or off, an agent must never claim to be a person: Deskwoot holds such a reply for a person.

What an AI teammate reads

An agent works only in the inboxes you pick for it: Every inbox, also new ones, or Only the inboxes I choose. There it reads its own work: conversations assigned to it or its team, handed to it, or where someone mentions it. It reads every conversation of those inboxes only when Reads every conversation in its inboxes is on. Leave that off unless the agent needs it, because one message that tricks the agent could then make it read every customer conversation in those inboxes.

It also reads the contacts and files of its conversations, contact notes, your help center articles, the Training Hub, canned responses, the names and teams of your members and your support reports for the team as a whole, never figures per team member. It never reads beyond your plan's history. Conversations and contacts marked Do not share with AI agents stay hidden from every agent, and so do the tickets of such a conversation or contact.

An agent with every inbox, in a workspace with tickets (Enterprise), also reads the tickets and trackers handed to its team, the ones where someone mentions it in an internal comment, the tickets made from a conversation it holds or where someone mentions it, or every one when it reads every conversation. There it writes internal comments, and when a person asks for it in a note that mentions it, it raises a Jira or Linear issue from the ticket in that person's name. It also creates tickets for your team: from a conversation it holds or nobody holds, and from another one when a person's private note there mentions it. It never replies to the customer from a ticket, changes a ticket's status or assigns it. More in Tickets.

Fynn can hand chats to an AI teammate: where you choose who gets the chat after Fynn, pick An AI teammate instead of your team. The agent gets the chat only while it is there and may work in that inbox. A customer who asks for a person always goes to your team.

The onboarding: up to 6 questions

Once the agent is connected, write to it in its own chat: Use the Deskwoot connector: call onboarding_start and go through the onboarding with me. The last step of Connect an AI agent shows this sentence to copy. The agent then asks you, one question at a time:

  1. What is its job: sort conversations and leave notes, also prepare drafts, or also answer customers?
  2. Which topics wait for a person? Billing and payments, legal questions and complaints, account security and data protection wait from the start.
  3. Which tone should it write in?
  4. Who takes over when it escalates a conversation?
  5. What must it never promise?
  6. How should it get its work: only when someone mentions it, also when someone assigns it a conversation, or also new conversations in turn with your team? An AI teammate asks this only when its job includes drafts or answers; with notes only, it works where someone mentions it. A personal assistant does not ask it.

From your answers it proposes its rules, and an administrator approves them on the agent's page under AI teammates. Rules that let it send more than before ask for your password. Until its rules are approved, the agent writes drafts only. To skip the questions, choose Go live with these rules on its page.

How it gets work

Under How it gets work on its page you choose how an AI teammate gets its conversations. Each level does what the one before does:

  • Only when @mentioned: it works only where someone mentions it and answers with notes for your team. Nobody can assign it a conversation, and it never replies to customers.
  • Assignable: a person, Fynn or a rule can give it conversations of its inboxes, wherever you can pick a person.
  • Like a teammate: it also gets new conversations of its inboxes in turn with your team, and the limit of open chats per person counts for it too. While it is paused or cannot be reached, your team gets them.

A higher level asks for your password.

Autonomy and held topics

Each agent has one level for all its inboxes. You pick it on the consent page, where it asks what the app may send to customers, and change it later on the agent's page under Replies to customers:

  • Only drafts (Writes drafts, a person sends them): someone in your team checks each reply and sends it.
  • Replies on everyday topics (Sends replies, held topics become drafts): it answers on its own, and replies on the topics you keep for people wait for your team.
  • All replies (Sends every reply Deskwoot does not hold): it answers on its own, and only replies that Deskwoot's safety checks hold wait for your team.

Under Topics that wait for a person you tick the built in topics (Billing and payments, Legal questions and complaints, Account security, Data protection) and add your own, separated by commas. With none ticked and none of your own, the 4 built in topics wait. Once a customer's message touches a held topic, the conversation stays held until a person clears it. Under Notes for the agent you add your own instructions.

Some situations wait for a person at every level: a crisis such as self harm, violence, abuse or a medical emergency, a request from the police, a court or another authority, a journalist asking for a statement, a customer under 18, a request for legal, medical, financial or tax advice, a promise of money, a refund, a credit or a discount, and a legal threat. When a customer asks for a person, Deskwoot tells them that a person from your team takes over and hands the conversation to your team.

What Deskwoot always holds

Whatever level you choose, Deskwoot checks every reply before it goes out and keeps it as a draft for a person when it:

  • links to an address that is not one of yours (yours are your help centers, your inbox domains, your company website and your connected stores),
  • contains an email address, a phone number, an order or tracking number or another long number that does not belong to this customer or this conversation,
  • asks the customer for a password, a code, card details or other login details,
  • says that something was done, such as a refund, that no person did, or says that the agent is a person,
  • answers a customer whose last message is from more than 72 hours before the agent was connected,
  • comes in a conversation that looks like someone tried to give the agent hidden instructions.

When Deskwoot cannot check a reply, the reply waits for a person too. Agents send plain text only, to the customer of the conversation. They never start a new conversation, approve a draft, mark spam or change a contact's email, phone, identifier, name or company.

Approve, edit or reject a draft

A held reply waits as a draft in its conversation, with the reason it waits. Conversations with waiting drafts collect in the Waiting for approval folder of your inbox, and the bell and a push notification tell you about them, never an email (AI agent replies to approve under Settings > Notifications).

  • Approve sends the reply under the agent's name and with its signature, marked as written by AI while its AI label is on. The conversation stays with the agent.
  • Edit lets you change the text before you send it.
  • Reject asks why, for example wrong facts or the wrong tone, and takes a note. The agent reads both.
  • Only a person signed in to Deskwoot approves. A yes inside the agent's own app never counts.
  • After 4 hours without a decision, the administrators get a reminder. After 24 hours the draft expires, nothing of it goes out, and the conversation goes to your team.

Orders, refunds and cancellations

An agent never moves money. When the workspace owner switches on Look up orders and propose refunds on the consent page while connecting the agent, or an administrator later switches on Order lookups and money proposals on the agent's page, the agent reads the orders and subscriptions of a conversation's customer from your connected stores and Stripe once that customer is verified, and before that only an order the customer names together with its postal code. It can then propose a refund, a cancellation or a shipment. A person opens the proposal in the conversation, sees the order live from the store and makes it with their own account, or rejects it.

Working with an AI teammate

  • Mention the agent in a private note to ask it something: type @ and pick it. A mention is new work for it.
  • When a person takes a conversation over, the agent only writes notes there until someone hands the conversation back.
  • While Fynn is still in a conversation, the agent stays out, so 2 bots never answer at once.
  • Deskwoot wakes an agent with a wake address, such as a Grok Bot routine, the moment new work arrives. Every other agent works when its own app or schedule starts it. Under How often it runs on its page you set how often its routine calls Deskwoot, every 5 minutes when you leave it empty.
  • When an agent stays silent longer than its usual rhythm, its page shows Not checking in, and conversations where a customer waits on it go to your team.
  • On the agent's page, the row Under replies to customers sets its email and chat signature for each of its inboxes, as for a person on your team. Notes never get one. A personal assistant signs with its member's own signature.

Pause, stop and activity

  • Pause: an administrator pauses an agent on its page under AI teammates. It stops with its next call, and Resume lets it work again.
  • Stop: the agent loses every sign in and token at once. Its history stays, and its conversations go back to your team. This cannot be undone; to use the agent again, connect it anew.
  • Deskwoot pauses an agent by itself when it held more than 10 of its texts within 10 minutes. The agent then shows Paused by Deskwoot until an administrator chooses Resume.
  • Activity on AI teammates lists the 10 newest things your agents and bots read, wrote or sent, and why Deskwoot held or refused one. You can show one agent at a time, and Export CSV downloads the last 90 days.
  • Switching off the MCP server at the top of AI teammates stops every agent of the workspace until you switch it on again.
  • Everyone finds the agents they connected under Settings > Security, in AI agents you connected, and can stop them there.

Plans and billing

AI teammates work on Startup, Business and Enterprise and in the 7 day trial. The free Hacker plan has none: its agents wait for a paid plan, and the page still lists them so an administrator can stop them. There is no fee per agent, and your plan puts no limit on how many agents you connect.

Your agents count toward the AI conversations of your plan, on top of Fynn and Copilot, together with replies through the REST API and from Custom Webhook bots. A reply of an agent that reaches a customer counts as 1 AI conversation, however many agents reply in the same conversation, and a draft of an agent that your team approves and sends counts 0.5. Reading never counts. How this adds up with Fynn in the same conversation, what other work of an agent counts and when a conversation counts again is in AI conversations and billing.

Your plan includes 50 AI conversations a month on Startup, 100 on Business and 300 on Enterprise, and each one more is paid from your prepaid balance: $0.07 on Startup, $0.05 on Business and $0.03 on Enterprise. When nothing is left, an agent's reply that would add to the count waits as a draft for your team, and approving it needs a top up first.

Frequently asked questions

What is the difference between Fynn and an AI teammate?

Fynn is built into Deskwoot, takes the first contact in your live chat and answers right away. An AI teammate is an agent you connect yourself. It works the conversations assigned or handed to it like a colleague, starting when Deskwoot wakes it or on its own schedule, and never answers while Fynn is still in the conversation.

Who can connect an AI teammate?

A new AI teammate starts working only once the workspace owner connects or approves it. One the workspace owner connects works at once. One another administrator or a member connects waits under Requests for AI agents on AI teammates until the workspace owner approves it, and cannot read or do anything before then; other administrators can decline it. A personal assistant connects at once while the workspace owner allows it with Members may connect personal assistants.

What happens when a customer asks for a person?

Deskwoot tells the customer that a person from your team takes over and hands the conversation to your team. The agent cannot reply there on its own until someone hands the conversation back.

Can I keep some conversations away from AI agents?

Yes. Mark a conversation or a contact with Do not share with AI agents. No agent sees it, and a contact marked this way keeps all their conversations away from AI agents.

Do AI teammates cost extra?

There is no fee per agent. Your agents use the AI conversations of your plan, on top of Fynn and together with replies through the REST API and from Custom Webhook bots: a reply that reaches a customer counts as 1, however many agents reply in the conversation, and an approved draft 0.5. Reading never counts. The rest is in AI conversations and billing.

What happens to our agents when the trial ends?

If the workspace moves to the free Hacker plan, its agents stop working with their next call and show Waits for a paid plan. Their settings stay, and once you choose a paid plan they work again without connecting anew.

Was this guide helpful?

Keep reading

Ask Fynn about this guide

Get a short answer from this guide and the rest of the help center.

Still need help?

Fynn answers from these guides in the chat, and you can reach the team there too.