A.I. Agents: Talk With Your Agent in Chat

What does it mean to talk with an Agent in chat?

An A.I. Agent is a named teammate your account owns. Until now it worked in the background, running the Skills you gave it. It can now also take a brief directly from you, in the chat you already use, and do the work itself: find a Contact, draft an Invoice, raise a Task, leave a Note, add someone to a Team.

You open a direct message with the Agent, write what you need in your own words, and watch it work. It signs in as you, so it can only find what your own lists show you and only change what your own screens would let you change. Nothing is created, changed or deleted until you press Confirm.

IMPORTANT: An Agent needs A.I. to think with. Your account must have an active A.I. Service subscription or a verified connection of your own under A.I. Connections. Without one, the Agent answers "This Agent cannot work until the account has A.I. An administrator can set it up."

Before you start

• A.I. is switched on for the account, as above.
• The account has at least one active Agent. CLICK HERE to learn how an Agent is created.
• You are a Staff member. Clients and Prospects never see the Agents tab, never get an Agent conversation, and are turned away by every Agent action. An Agent cannot be given to a Client or a Circle.
• You have been given the Agent. A brand new Agent is reachable by administrators only until someone switches other people on.
• You are not impersonating anyone. While you are signed in as someone else, the Agent answers "You cannot talk to an Agent while impersonating another user. Exit impersonation and try again."

How do I start a conversation?

1. Open the chat and go to the DMs tab.
2. Click the + button to open Create Direct Message.
3. Choose the Agents tab. It sits beside Staff, and beside Companies and Contacts if your role may message Clients. It appears only when there is at least one Agent you are allowed to use.
4. Pick the Agent. The conversation opens and then lives among your other direct messages. There is one conversation per person per Agent, so yours is yours.

TIP: Brief your Agent the way you would brief a colleague. You do not need exact record names or ids. It looks things up for you, tells you what it found, and asks when it is not sure which record you meant.

If the tab reads No Agents Yet with the line "An administrator can give you access to an Agent.", the account has Agents but none of them has been opened up to you yet.

What happens while it works

Send your message and a small line appears underneath it: the name of the step the Agent is on, or Working while it is deciding what to do next, together with how long it has been going and a Stop button at the end.

• Stop ends the request before its next step. Steps already taken stay taken, a change that was still waiting is listed as "Not run", and the answer closes with "Stopped as you asked."
• The box you type in is never locked. You can carry on writing while the Agent works, and your new message waits its turn.
• One request runs at a time in a conversation.

Confirming a change

Before the Agent changes a record, deletes one, or creates something a Client would be able to see, it stops and shows you a card headed Confirm before this Agent goes on:. The card names the record and spells out the exact change, including the value as it is now and the value it would become.

• Confirm makes the change. The card then reads "Confirmed".
• Skip leaves that change undone and lets the Agent carry on with the rest.
• Stop ends the request there.
• Open This on Its Own Page shows the same list on a full page titled "An Agent Is Waiting for Your Confirmation", with a link back to the conversation. Useful when a card arrives on a phone, or when you want to read a long change properly.

IMPORTANT: Typing "yes" is not a confirmation. Any message you send while a change is waiting counts as a new request, and the change that was waiting is left undone with "Stopped here because you sent a new message." Only the buttons confirm, and only the person who asked can press them. Nothing written inside a document or a record can confirm anything.

If nobody answers a waiting card for three days, it closes on its own and the change is not made.

Seeing exactly what the Agent did

Above every answer sits one small line: Used 1 Tool, or Used 3 Tools, and so on. Open it and you get one line per step, each naming the record it touched and linking straight to it.

• The answer ends by telling you what it could not do, or says No changes when nothing in the account was altered.
• A step your role does not allow is listed with "You do not have permission for this".
• A step that reached a tool switched off for that Agent reads "Switched off for this Agent".
• Everything the Agent creates or changes carries the Agent's name in the record's history, so the trail shows which work was done by an Agent and which by a person.

When the Agent asks you something

If the Agent needs to know something only you can tell it, such as which of two people called Henderson you meant, it asks in the conversation and a banner appears above the box you type in: "This Agent is waiting for your answer. Your next message answers it:", followed by the question. Your next message is the answer, and the Agent picks up where it left off. An unanswered question closes after three days.

Handing the Agent a file

Agent conversations have a paperclip, Attach Files, that ordinary conversations do not.

• It reads plain text, Markdown, CSV, JSON, XML, HTML, log and YAML files, Word documents, Excel workbooks and PDFs.
• Up to 20 MB per file, and up to 10 files on one message.
• Pictures are pasted into the message instead of attached, and the Agent looks at up to three of them per message.
• It only reads files you uploaded yourself through the chat, in this account. A colleague's upload is named in the answer with "Only files you uploaded can be read".
• Any file it could not read is named with the reason, such as "Too large to read whole" or "It has no text to read, as with a scanned PDF", rather than being quietly skipped.

What can my Agent actually do?

An Agent carries 53 tools, grouped the way your main menu groups the screens they belong to. Every tool behaves exactly like the screen it stands for, and answers to your role in exactly the same way.

You can see the whole list, with a plain description of each one, on the Agent's Tools page.

IMPORTANT: An Agent never sends anything to a Client. An Invoice, Estimate or Proposal it writes is saved as a draft with sending switched off, and a person decides if and when it goes out. An Agent also cannot make anyone an administrator.

Which steps wait for a Confirm, and which do not

This is decided by the tool itself, never by the Agent. An Agent asked to do something that needs a yes cannot talk its way past the question, because it is not the one being asked. There are three cases.

• Needs Confirmation, always. Every update, every delete or deactivate, creating a colleague, creating a Team, adding a colleague to a Team, and adding an Item to your catalog.
• Needs Confirmation when a Client will see it. Creating a Contact, a Company, a Note, a Project or a Task waits only when what it creates would be visible to a Client. The card says who it reaches, for example "will see it in the portal" or "will be invited to the portal".
• No confirmation. Looking things up, and drafting an Invoice, Estimate or Proposal. A draft appears on no Client list, triggers no email and no notification until somebody sends it, so there is nothing for a Client to see yet.

On the Agent's Tools page, a tool that waits carries a small mark beside its name reading Needs Confirmation, or Needs Confirmation When a Client Will See It.

How do I stay in control?

The Agent always acts as the person who asked. In the platform's own words on the Agent editor: "The Agent only does what the person using it can do: their own role's permissions, and only the records they can see. If a Teammate's role can't create invoices, the Agent won't draft one for them."

Giving someone an Agent never widens what they can reach. There are three separate controls, all on the Agent's own editor.

1. Access decides who may use the Agent. With nobody switched on, that means administrators only: Super Admin and Admin can use every Agent by default. Switch on individual Staff, roles or Teams to open it up.
2. Only a Super Admin can take an Agent away from an administrator. Anyone else sees the Admin row locked with "Only a Super Admin can change this." An administrator switched on by name keeps the Agent even if their role is switched off.
3. Tools decides what the Agent may do at all. Every tool starts on. Switch one off and it is off for everyone who uses that Agent, not just for you.

Deactivating an Agent from the Agents list removes it from everybody's Agents tab at once. The conversations people already had stay readable, sending to it answers "This Agent is not available in the chat right now.", and a request already running stops with "This Agent has been switched off, so it stopped here."

Messages you might see

Good to know

• Your conversation is yours. One per person per Agent. Your colleagues do not see your requests, your steps or your answers.
• It follows the thread. The Agent reads back the recent part of the conversation, so "and give each one a due date of Friday" refers to what you were just discussing.
• It looks records up again rather than remembering them. A record mentioned several messages ago comes back to it only as its kind, "the contact", so it searches again before naming one. That is deliberate: it stops an Agent acting on a stale memory of a record.
• Stop does not undo. A step already under way finishes. Stop only prevents the next one.
• Who pays does not change. Agent requests are funded exactly like your other A.I. features: the A.I. Service when the account has it, otherwise your own verified connection.
• There is no account-wide log of Agent runs. Each person sees their own Agent's steps in their own conversation.

CLICK HERE to learn about A.I. Agents and Skills.

CLICK HERE to learn about Live Chat.

CLICK HERE to learn about Staff roles.