Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.elementum.io/llms.txt

Use this file to discover all available pages before exploring further.

Elementum agents are conversational AI assistants you can chat with to get answers, take actions, and move work forward. When you interact with an agent, you’ll want to know where to find one, how to start a conversation, the ways an agent can collect information from you, and what it can see and do on your behalf. If you are creating or configuring an agent, see Building Agents instead.

​
Where you can chat with an agent

An agent is reachable from any of the following surfaces, depending on how an admin has set it up:
  • In your Elementum app — Chat from a managed view (a chat box above the records, or a full-page Agent view), or from an Assign an Agent button on a record.
  • Microsoft Teams — One-on-one chats with the Elementum bot. After your Teams admin has added the app, find Elementum in Teams and start a conversation.
  • Slack — In channels where the Elementum app has been added. Each channel can use a different agent.
  • Phone (Twilio) — Call the phone number assigned to the agent and hold a voice conversation.

​
Starting a conversation

When you open an agent, it greets you with a first message. From there you can:
  • Click a Starting Action — a clickable chip below the greeting — to begin a common task with one click.
  • Type a message to ask anything.
In Microsoft Teams, an Elementum app record is created for your session so the conversation can be associated with your user details and reviewed later. In Slack, behavior depends on how the channel is configured — some channels reply to every message, others only to questions or only to high-confidence matches.

​
Starting Actions

Starting Actions are clickable prompt chips that appear directly under an agent’s opening message before you’ve sent your first message. They are a fast way to kick off common tasks without typing.
  • Actions without variables send the prompt as soon as you click the chip.
  • Actions with variables open a short inline form that collects values (text, number, or dropdown) before sending the prompt.
Starting Actions disappear after you send your first message in a conversation, whether you sent it by clicking a chip or by typing. For the full reference, see Agent Starting Actions.

​
Sharing files in a conversation

You can give an agent a file as part of the conversation so it can read and reason over the contents.
  • Native file inputs — Agents accept .xlsx and .docx files directly in conversation, with no manual conversion required. The platform converts each file into a format that preserves embedded images, signatures, and content before passing it to the AI model.
  • Slack attachments — You can attach files in Slack the same way as any other Slack message. Agents running on multimodal-capable models can use the attachment content in their reply; other agents still receive the message text but do not interpret the file. File attachment support in Slack is available on the Enterprise track.

​
Answering in-chat forms

Some agents use the Ask User Question tool to collect input with a structured form instead of asking multiple back-and-forth questions. When the agent invokes the form:
  • A form is rendered inside the chat with the questions, fields, and any required selections.
  • Fields can be single-select, multi-select, or free text.
  • Dynamic dropdown fields populate their options from live platform data — for example, a list of vendors or projects — and let you search within long option lists.
  • Submitting the form sends your answers back to the agent so the conversation can continue.
Form wording, options, and field layout are generated by the model at runtime, so they may vary slightly between invocations.

​
Understanding what the agent is doing

While the agent works, live status messages appear in chat — for example, “Searching knowledge base…” — that indicate when a tool is running. When an agent uses several tools back-to-back, each one appears as its own status message so you can follow what it is doing.

​
Long conversations

Elementum automatically summarizes earlier messages when a conversation approaches the model’s context limit, so you can keep chatting through long sessions without the agent losing track. Compaction happens transparently in the background — no action is required on your part.

​
What the agent can see and do as you

How an agent acts on your behalf depends on the permissions mode an admin set when creating it:
  • Run as current user — The agent acts under your permissions. It can only see and modify records you can access. Agents attached to managed views always respect the viewing user’s permissions in this way.
  • Run as publisher — The agent acts with a fixed level of access regardless of who is chatting. Two users may see the same answers because the agent reads and writes the same set of records.
  • Run as service account — The agent uses a dedicated controlled account configured by an admin. Common for integrations or processes that require consistent access independent of any one person.
If you are not sure which mode an agent uses, ask the agent’s admin — it affects what data the agent can return to you.

​
Channel-specific notes

​
Microsoft Teams

  • Conversations are one-on-one with the Elementum bot, not in channels or group chats.
  • Only one active conversation per user at a time.
  • After a period of inactivity, the conversation closes and the next message starts a new conversation.
  • Agents may end a session themselves once the request is complete.
See Agent Microsoft Teams Integration for the full reference.

​
Slack

  • Each Slack channel is configured separately and can use a different agent.
  • Some channels operate in Quick Answers mode (short replies, no multi-turn state) and others in Full Conversations mode (full transcript and context).
  • The agent replies to messages according to the Channel messages rule set for the channel — all messages, only questions, or only high-confidence matches.
  • Idle conversations close after a configured timeout; the next message starts a new conversation.
See Agent Slack Integration for the full reference.

​
Phone

  • You speak with the agent over a regular phone call.
  • The agent reads its responses aloud using a voice model selected by an admin.
  • If you go quiet for too long, the agent plays an idle message and may end the call after a configured timeout.
  • Some agents restrict who can call. If the caller’s phone number is not recognized, the agent plays a message explaining that they could not be matched to a user.
See Agent Phone Integration for the full reference.

​
Conversations are recorded

App admins can review the transcript and tool activity of past conversations from the agent’s Conversation History tab in Intelligence. They can also see a step-by-step Waterfall breakdown of how long each part of the conversation took.