Conversations

Review what your agents say across every channel, oversee internal AI chats with privacy controls, flag conversations, read end-user feedback, and use the dashboard Chat to talk to your own agents.

Every interaction with your agents is logged as a conversation: a website chat (typed or spoken), a phone call, a form-driven exchange, or a member chatting with the AI inside the dashboard. botts.ai gives you two related but distinct surfaces for working with them, plus a built-in Chat for talking to your agents yourself.

Members see only the dashboard Chat. The analytics surfaces are for Builders and above.

Channels

Conversations come from these channels, each shown with a colored badge:

WhatsApp, Telegram and Email appear in the channel picker as Coming Soon and can't be deployed yet, so no conversations come from them today. See Deployments.

Internal chats are a first-class channel with their own privacy and oversight rules. See Internal-chat privacy and oversight.

Why review conversations?

Regularly reviewing conversations is one of the most impactful things you can do to improve your agents:

• Catch errors early: Spot incorrect or incomplete answers before they become patterns.
• Improve your Knowledge Base: If the agent can't answer a question, it usually means the Knowledge Base is missing that information.
• Refine your System Prompt: See whether the agent follows your instructions or drifts off-script.
• Understand your customers: Learn what questions people actually ask, which may differ from what you expected.

The Conversations hub

Open Conversations under the ANALYTICS group in the left sidebar. The hub is a two-pane view: a list of recent conversations on the left and a transcript on the right.

The hub lists conversations across all channels, newest first, and loads more as you scroll, so you can reach the full history. Each row shows:

• A channel badge (Voice, Website or Internal).
• A caller label: the end user's phone number, or for internal chats the owning member's name or email, otherwise the channel identifier.
• The start time and a message count ("N msgs").
• A flag icon when the conversation has been flagged.
• A rec marker when a voice recording is attached. Open the conversation to play the recording above the transcript.
• A flask icon with a number when replies in the conversation have been saved as tests.
• For internal chats, an admin-visibility icon: an emerald lock for private, an amber eye for full. See Internal-chat privacy and oversight.

Click any row to load its transcript in the right pane.

Filtering

The hub has four filters above the list:

• Search ("Search messages, callers or sources"): matches message text and titles, the caller's phone number or identifier, the member who owns an internal chat, and the agent or deployment name. It searches the whole history, not just the rows on screen. Message text of an internal chat you aren't allowed to read is never searched.
• Date: All time, Today, Last 7 days, Last 30 days, Last 90 days, or Custom range.
• Channel: Website and the web widget share one "Website" option. Options are built from the conversations loaded so far.
• Source: the agent and deployment the conversation came from.

Clear filters resets all four.

Exporting conversations

Export downloads conversations as a file. In the hub it exports every conversation that matches the current search, date, channel and source filters, across the full history and not just the rows loaded on screen. On an agent's Chats page it exports all of that agent's conversations.

Five formats are offered, in two shapes:

For a very long conversation, prefer CSV or a per-message format: Excel caps a single cell at 32,767 characters, so the per-conversation .xlsx silently loses the tail of a transcript longer than that.

Every row carries the conversation's metadata: id, channel, source deployment, channel id, phone, language, status, start and end time, message count, flag, negative-feedback count, and owner. Message rows add the timestamp, role, content, and any feedback the customer left with its comment.

Transcripts contain only the live customer and agent turns. Tool-call plumbing and superseded versions of edited messages are left out, because they would drown the readable conversation.

Reading a transcript

The transcript shows every message exchanged. For each message you may see:

• The role (the customer/member and the agent).
• Tool-use cards when the agent called a tool: the tool name, its arguments, and its result.
• End-user feedback: a thumbs-up or thumbs-down a customer left on an agent reply, with an optional comment.

Internal chats are not shown in the hub transcript through the analytics path. Their content is governed separately (see below).

Flags and status

Conversations carry two review states:

• Flagged: a marker (amber flag) you set to revisit a conversation later. Toggle it with the flag button on an agent's Chats page. The hub displays the flag but does not toggle it.
• Status: active or closed, shown as a pill on the agent's Chats page. You don't close conversations by hand. A phone conversation is closed automatically when the call ends, which stamps its end time and starts a background memory-extraction task so the agent can carry useful facts forward.

Per-agent Chats and Forms

Open an agent and use its Chats and Forms pages (in the Insights group of the agent navigation, which starts collapsed on a wide screen) to work with that agent's data in detail.

Chats

The Chats page has two views, toggled at the top: Individual (the default, one row per conversation) and Grouped (conversations grouped by person).

The Individual view lists that agent's conversations and adds controls the hub doesn't have:

• A flag toggle.
• A status pill (active / closed).
• A negative-feedback count: a red thumbs-down with the number of replies a customer rated negatively.
• A recording player for voice calls (see Voice recordings).
• Save as Test under each agent reply, which turns the conversation up to that point into a saved test. Once saved, it becomes a Saved as test link, and the row shows how many tests the conversation backs.
• The full transcript, including tool-use cards and per-message feedback.

The Grouped view lists the contacts (end users) this agent has interacted with.

Voice recordings

When a voice call has a recording, a player appears above the transcript, both in the Conversations hub and on the agent's Chats page.

Forms

The Forms page (headed Form Submissions) lists form submissions captured by the agent. It is available to Builder, Admin, and Owner.

• The table is paginated and shows, per submission: the date, the form name, the visitor's name or email (or "Unknown"), and the first three captured fields (with a "+N more" indicator).
• A search filters by form name.
• Export CSV downloads every submission, with one column per captured field.
• The trash button deletes a submission (after a confirmation). Deleting requires the manage level of forms access.

The dashboard Chat

Chat sits at the top of the left sidebar and is available to every role. It is where you talk to the AI yourself, either with a general-purpose model or with one of your deployed agents. It has two modes. Switch between them with the Chat / Agent Chat toggle in the chat header.

Agent mode

Agent mode lists agents that have an active internal-chat deployment and that you're allowed to use. Access follows the deployment's setting: open to the whole organization, restricted by minimum role, or invite-only. Each agent shows its name and logo, and a new chat opens with the agent's welcome message.

A deployed agent runs on the model chosen at deploy time: you can't switch a deployed agent to a different model from the chat. If the primary model fails, or hasn't started answering within 30 seconds, the agent automatically switches to its configured backup model. The reply records the model that actually answered. (When you chat with an agent you're still building and haven't deployed, it runs as a builder test chat where you can pick the model.)

Regular mode

Regular mode gives you a model picker showing the models your organization is allowed to use, ordered with recommended models first. Each option shows its name, a recommended star, intelligence and speed (each as an "N/5" rating), and hosting country. Text usage is billed in credits.

You can change the model in the middle of a conversation: the picker stays available once a chat has started, and the model shown when you press send is the one that answers. The thread carries on without losing anything, and a "Switched to …" line marks where the change took effect. Each reply is recorded with the model that wrote it, so a mixed conversation stays honest about who said what. (Switching between Regular and Agent mode still needs a new chat; an agent's model comes from its deployment.)

What the chat remembers about you

The chat builds a private memory of you from your own regular chats (your role, the projects you name, how you like answers), and every model you talk to uses it, so you don't have to repeat yourself after switching. It is on by default and entirely yours: Settings → Chat memory lists what has been learned, deletes any single item, or forgets everything, and a switch there turns it off (what is stored stays until you delete it). Memory is per workspace, is never shown in the admin conversation views, and is built only from your own chats, not from agent chats and not from scheduled-task threads.

Model availability follows your organization's Server Location setting (see Data residency): only compliant models appear, and picking a non-compliant one is rejected. If an admin switches off Regular Chat under Organization → General, the regular-mode tab disappears and members can only use Agent Chat.

The regular-mode assistant identifies itself as the built-in assistant of botts.ai (a Swiss agent platform), writes Swiss High German spelling (always "ss", never "ß"), and has no access to your files, emails, or anyone else's conversations. It sees only the current conversation, the documents you attach, and, unless you switch it off, your own chat memory. Beyond that, it can look up and change what your role allows in the workspace itself; see Managing the workspace from the chat.

Web search. Regular mode can search the live web for current or general information beyond what the model already knows, and open the most relevant result to read it in full, citing its sources and flagging figures that may be out of date. It's billed in credits like everything else here: per search, plus a per-page charge when it opens a page in full. (This is the same web-search capability agents can use.)

Working in a chat

• Stop generating: while a reply streams, the send button becomes a red stop button; pressing it ends the response cleanly and keeps the text generated so far.
• Queue a follow-up: you don't have to wait for the reply to finish. Send another message and it is queued visibly beneath the composer; the queued messages are answered together in one coherent reply once the current answer completes. A queued message can be removed before it is sent. Only the chat that is busy queues: a new chat or another conversation answers right away, so several chats can generate at the same time.
• Leaving mid-reply: switching to another chat doesn't interrupt a reply. It keeps generating, and a pulsing blue dot in the conversation list marks each chat that is still answering. If you close the tab, the reply stops there and the text written so far is saved in the conversation. If only your connection drops, the reply often finishes anyway (see Connection lost under When something goes wrong).
• Voice dictation: the microphone button records you and transcribes the audio into the input box (it is dictation into the text field, not a spoken conversation). Audio is capped at 25 MB.
• Edit and branch: you can edit one of your own messages; this creates a new branch and re-runs the agent's reply. Arrows let you move between branches when a message has more than one version. Attachments on the message you edit are kept.
• Titles: after your first exchange, a short title is generated automatically. You can rename it.
• Copy: copy a reply or an individual code block. Pasting into Word or Google Docs gives clean formatting rather than the page's own styling.
• Jump around a long chat: once a conversation has three or more exchanges, a rail along the left edge of the transcript shows one tick per exchange. Longer exchanges draw longer ticks, and the one you're reading is highlighted. Hover a tick to see the question and the start of the answer, and click it to jump there. In a scheduled task's thread, hovering shows when that run happened instead of the instruction. The rail appears on wider screens, not on a phone.
• New chat from inside a chat: while the conversation list is closed, as it is by default on a phone, the + in the chat header starts a new chat. With endless chat on, the button isn't there.

Organizing your chats

The conversation sidebar groups your chats by age (Today, Yesterday, Previous 7 Days, Older). Threads of your scheduled tasks get their own Running for you group. The sidebar gives you a few ways to keep the list manageable:

• Search across your conversations, and filter the list.
• Pin to top: pinned chats sit in their own group above the rest. You can also drag a chat onto the pinned area.
• Mark as unread: flag a chat to come back to. Results delivered by a scheduled task arrive marked unread the same way.
• Rename and delete from the chat's options menu.

Attaching documents

You can attach documents to a chat and ask the agent about them. Click the attach button, or drag files onto the composer.

Small documents are read in full and placed directly into the agent's context. Larger ones are indexed instead, and the agent searches them the same way it searches a knowledge base, so a long PDF stays usable without flooding the conversation. Either way the agent tells you which file an answer came from.

Pasting a very long block of text turns it into a text snippet attachment rather than an oversized message, which you can open and edit. Whether a snippet is inlined or indexed is decided by its size, like any other attachment; an indexed one is labelled Knowledge base on its chip.

Uploads count towards your organization's upload storage limit and cost credits to index, exactly like knowledge-base content. Attaching the same content twice to one conversation reuses the stored copy rather than paying to index it again.

Retention is an organization-wide setting, not a per-file choice: Organization → General → Document Retention, from one week to forever. It applies to new uploads; documents already uploaded keep the expiry they were given. See Team & Billing.

Attaching images

Screenshots and photos can be attached the same way, or pasted straight into the composer. Models that understand images carry an Images badge in the model picker; with any other model the attachment is refused before it is sent, and a notice offers a switch to a model that can see it.

Images are stored downscaled to 1568 pixels on the long edge with camera metadata removed, which keeps text in a screenshot readable while bounding what each turn costs. The model sees the images on the last three messages directly; older images stay listed for it, and it can look at one again when the conversation refers to it (shown as Looking at an image while it does).

Image input is billed as input text at the model's rate, roughly 1,000 to 2,000 tokens per screenshot on every turn where the image is shown. Retention follows the organization's Document Retention setting, like other attachments.

Spreadsheets and charts

CSV, Excel and JSON files are analysed rather than read: the assistant runs code in an isolated sandbox against the file and answers with figures, tables and interactive charts. Each chart can be downloaded as a PNG, and when the assistant redraws a chart the old one collapses as Earlier version, replaced. Files it produces (CSV, Excel, JSON or text) are offered for download. Each run costs 0.5 credits, with at most 10 runs per message.

Interactive views

Where it helps, an answer comes with an interactive view under the text: key figures, a chart, a table, tabs to compare options, or a form to fill in. A what-if question ("what if the project takes two weeks longer?") gets sliders, and the figures update while you drag. Questions that are better answered in words, such as an explanation or a draft e-mail, get plain text.

• Buttons and forms reply for you. A button, a suggested follow-up or a submitted form sends an ordinary message with your choice and the values you entered, and the next answer works from it. A button that links to a web page opens that page instead. A view changes nothing by itself: whatever needs doing is done by the reply, as usual.
• Print: the printer icon under an answer with a view prints that answer on its own, without the rest of the chat.
• While the answer is still being written, the view shows Building the view. If a view can't be drawn, it says This view could not be shown. and the text of the answer stays.
• Where owners and admins can read an internal chat (see Internal-chat privacy and oversight), views appear as they were shown but can't be clicked.

Managing the workspace from the chat

In regular mode you can ask the chat to do the setup work you would otherwise do on the Agents, Knowledge and Integrations pages, for example "create a support agent that answers from our product catalogue and collects contact requests", "add our new pricing page to the sales knowledge base" or "how much of our plan have we used this month?". It checks your role for every step, the same way those pages do; the table shows what each role can ask for.

• It proposes before it creates. For a new agent the chat shows the name, the instructions (as a quote) and the features it plans to switch on, so you can correct them first. The agent then appears on the Agents page, where everything can still be changed.
• Features are an agent's on/off switches: web search, advanced knowledge browsing, lead capture with the contact, support ticket or booking request form, conversation history and customer memory. Finer settings, such as custom forms, how much history is kept, how long memory lasts or which sites web search may use, stay on the agent's page.
• Internal chat only. The chat can put an agent on internal chat, so your team can pick it in agent mode. A website widget or phone line is still set up on the agent's Deploy tab (see Deployments). Taking an agent off internal chat deletes nothing; deploy it again and it is back.
• Live agents are changed on their page. While an agent answers on a website or by phone, its instructions, knowledge, connections and features can't be changed from the chat, only on the Agents page. Its name still can.
• Credentials never pass through the conversation. When a connection needs an API key, the chat shows a secure field instead of asking you to type the key into a message. The value is stored encrypted, the assistant never sees it, and the field works once within 24 hours. A credential stored by an owner or admin is shared with the workspace, and they then link it to the connection on the Integrations page (see Authentication). One stored by a builder is used only in their own chats, by a per-user connection that already uses that key. The tools of a new MCP server reach agents only once an owner or admin imports them there.
• The usual limits and costs apply. New agents count against your plan's agent limit and internal-chat deployments against its deployment limit, and at either limit the chat says so. Taking an agent off internal chat frees a deployment, not an agent slot; for more agents, delete one or upgrade. Web pages and text added to a knowledge base use your page allowance and credits exactly as on the Knowledge page. On a plan with a page limit the chat adds one web address per message, because how much a crawl will read isn't known in advance. The usage overview shows pages and uploads against your plan and, for owners and admins, the credit balance; what the credits went to is on the Plans & Billing page.
• Regular mode only. In agent mode you talk to the agent itself, and an agent never gets these abilities, not even for an owner. Scheduled tasks don't have them either.

Long conversations

A conversation that runs long is compacted rather than truncated: instead of replaying every message, the agent works from a running summary of the earlier part plus the recent messages in full. This keeps a long chat inside the model's context window and keeps its cost from climbing with every turn, while preserving what was established earlier.

You don't configure this and it happens automatically. If a single message is itself too long for the model, the chat tells you so and suggests starting a new conversation or shortening the message.

When something goes wrong

The chat reports provider problems in plain language rather than failing silently:

Endless chat

If your organization has endless chat turned on, the experience changes: new messages continue your existing conversation instead of starting a fresh one. Regular chat is a single ongoing thread (changing the model in the picker moves that thread onto the new model rather than opening a separate one), and agent chat keeps one thread per agent. The conversation list starts closed (open it from the icon in the header to reach earlier chats), and conversations can't be deleted.

Credits

The chat checks your credit balance before sending. If your balance has run out, the send is refused and the chat shows This organization is out of credits. Add credits to keep chatting. Usage is billed in credits.

Internal-chat privacy and oversight

Member-to-AI chats are treated as more sensitive than agent analytics, so they have their own privacy model.

Ownership. Each member's internal chats are private to them. The dashboard Chat lists only your own chats, and no one, not even an Admin or Owner, can edit, delete, or continue another member's chat.

Admin visibility. Owners and Admins set the policy under Organization → General → Internal chat visibility:

Each new chat snapshots the policy in force at the moment it's created. Changing the org policy later only affects new chats; it never retroactively exposes older private chats. The member always sees a badge in the chat header showing the snapshot for the current conversation (Private or Admins can read).

Who can oversee internal chats. Only Owners and Admins get internal-chat oversight. Builders are excluded: in the hub, a Builder never sees other members' internal chats at all (they're filtered out of the list, and opening one returns "not found"). A Builder sees only their own internal chats listed, and even those show no message content in the hub. To read their own internal-chat content, a Builder uses the dashboard Chat.

• Metadata is always visible to Owners and Admins.
• Content is visible only when the caller is the chat's owner, or the caller is an Owner/Admin and that chat's snapshot is Full access.

When content is withheld, the hub detail pane shows a lock and a "content hidden" notice instead of the messages, and the title is hidden too.

Who sees what

Common patterns to watch for

The improvement loop

The most effective improvement cycle is:

1. Review: Check the latest conversations and the negative-feedback counts.
2. Identify: Find recurring issues or unanswered questions.
3. Fix: Update the Knowledge Base or refine the System Prompt.
4. Test: Save the problem conversation as a test (Save as Test on the agent's Chats page) and use Run All on the agent's Tests page, or try it in the dashboard Chat in agent mode.
5. Monitor: Check follow-up conversations to confirm the improvement.

Over time, this feedback loop makes your agents significantly more accurate and helpful.

Last updated on October 28, 2026