Chat Data
AIBuild, train, and operate AI chatbots with Chat Data, training on text, URLs, and documents. Agents keep a bot's knowledge current and react when a conversation crosses into territory a person should handle.
What This Integration Enables
Most chatbot platforms hand you an answer and let you decide later whether it was any good. Chat Data hands you the answer and the arithmetic behind it. Get Confidence Score returns the average and per chunk cosine similarity of the retrieved context plus the exact embedding text used, and On Chat Completed carries a confidence score on every completed turn. That single design choice is why this connector belongs in a governed workflow: answer quality stops being a subjective judgment made after a complaint and becomes a value a FlowRunner agent can test at runtime.
The rest of the surface is built for a support team that actually staffs the bot. Chat Data registers real webhooks for completed chats, live chat escalations and lead submissions, so the workflow starts the moment something happens rather than on the next poll. An agent can silence the bot on one conversation with Set Bot Control, reply inside the thread as a human with Send Live Chat Message, then hand it back. It can pause automation on a single WhatsApp number or Facebook page during staffed hours without touching the other channels. It can add agentic actions the bot invokes mid conversation, and it can manage which sub-users on the account are allowed to train which bots. Every one of these exists to make one thing dependable: the bot answers what it can answer and stops when it cannot, which is the whole premise of human-in-the-loop automation.
Without FlowRunner
With FlowRunner
Use Case Scenarios
A knowledge backlog built from what the bot actually got wrong
On Chat Completed fires on every answered turn. The agent reads the confidence score and lets the confident ones pass. For the rest it runs Query Knowledge Base against the same question to capture which chunks were retrieved, then writes a record to Notion holding the question, the answer, the score, and the context the bot leaned on. By the end of the week the support lead is not reading a satisfaction survey, they are reading a ranked list of the topics the corpus does not cover, each one traceable to a real visitor question. That list becomes the retraining plan rather than a guess.
Escalation that arrives with the thread already in hand
On Live Chat Escalation fires the moment a visitor asks for a person. The payload already carries the conversation history, the lead's details and a handling URL. The agent calls Set Bot Control to silence the AI on that conversation so the bot does not talk over the human who is about to arrive, generates an admin link with Get Admin Live Chat Link, and posts the whole package into Slack: the customer's name, what they were asking about, and one link that opens the live thread. The agent is not summarizing the escalation for someone to go find. It is handing the conversation over.
Chat leads that reach the CRM only when they are real
On Lead Submission delivers the contact details a visitor typed into the widget. The agent runs List Customers to check for an existing record, then calls Attach Lead To Conversation so the transcript and the person stay linked. Submissions carrying a usable email or phone sync straight into HubSpot. Submissions whose only identifier is a platform username do not, because Chat Data cannot match those deterministically and a lead's UUID can never be changed once set. Those go to a rep to qualify by hand. The agent creates the records it can create correctly and leaves the ambiguous ones to a person rather than seeding the CRM with duplicates nobody can merge.
Human-in-Loop Highlight
The gate sits on Retrain Chatbot, and specifically on its Deletes list. Retraining a Chat Data bot is not additive housekeeping. Re-crawling a URL replaces that page's earlier content rather than duplicating it, and the Deletes list removes ingested sources outright. There is no version history to roll back to. So the moment a docs page is published half finished, or a policy PDF is superseded by a draft, a scheduled re-crawl can quietly overwrite the correct answer and every visitor from that second forward gets the new one. Nothing errors. Nothing looks broken.
So the refresh flow does the work and holds the trigger. It uploads the new material with Upload Training File, runs Get Confidence Score across a fixed set of canary questions the team cares about, and posts the comparison to the knowledge owner: "Re-crawl of 14 documentation pages is staged. Confidence on 11 canary questions is unchanged or better. Confidence on 'what is the refund window' drops from 0.81 to 0.44 because the pricing page now redirects. Approve the retrain, or exclude that URL?" The owner approves, or pulls the one page out. Only then does the agent call Retrain Chatbot. The agent handles the crawling, the scoring and the comparison. A person owns the change that rewrites what the company tells its customers.
Agent Capabilities
36 actionsChatbots and Training Lifecycle
8- Create Chatbot Creates and trains a new bot, optionally seeded with raw text, a list of URLs to scrape and Q&A pairs. Training runs asynchronously, so this is the start of a lifecycle rather than a finished object.
- Retrain Chatbot Adds knowledge to an existing bot or removes knowledge already in it. Re-crawling a URL replaces its earlier content, and the Deletes list removes previously ingested sources, which is why flows put an approval in front of it.
- Get Chatbot Status Reports training progress from 0 to 100. Poll it after Create Chatbot or Retrain Chatbot so a workflow only queries a bot whose knowledge base is actually ready.
- List Chatbots Returns every bot on the account with its ID, name, model, visibility, data sources and conversation count. The usual way to discover the IDs every other operation needs.
- Get Chatbot Retrieves one bot's full configuration including base prompt, greeting and suggested messages, rate limits, widget styling, lead capture and notification settings. Used to audit a setup before changing it.
- Update Chatbot Settings Changes behaviour and widget configuration: system prompt, retrieval model, temperature, greeting, per visitor rate limiting, domain allow listing, chat history and IP retention, and lead capture prompts. Only the fields supplied are changed.
- Update White Label Settings Rebrands a bot for resale by rewriting or hiding the credit line and registering a custom embedding domain and sender address. When a custom domain is supplied the response returns the DNS records that must be published to activate it.
- Delete Chatbot Permanently deletes a bot along with its knowledge base, conversations and collected leads. This cannot be undone, so flows export first.
Conversations
4- Send Chat Message Sends a message to a bot and returns its answer plus the retrieved source chunks. Pass a Conversation ID to continue a thread server side, or override the model and base prompt for one call only.
- List Conversations Retrieves conversations with filtering by time window, traffic source, attached lead, live chat state and visitor rating. Returns a total count so a flow can page through a full history.
- Get Conversation Retrieves one conversation in full with every message, the attached customer, visitor country and browser, the page the chat started on and the visitor's rating.
- Delete Conversation Permanently deletes a conversation and its message history. Used to honour data deletion requests.
Live Chat Handover
5- Send Live Chat Message Posts into an existing conversation as a human agent, as the assistant, or on behalf of the end user, optionally with attachments. The response reports whether the visitor was online or an offline email went instead.
- Set Live Chat State Opens or closes the human live chat session on a conversation, which is how a thread is handed back to the AI once a person is finished.
- Set Bot Control Silences or re-enables the AI on a single conversation so a person can take over without ending the thread. The silence can be time boxed in minutes.
- Get Conversation Portal Link Generates a shareable URL that renders a transcript for anyone holding the link. Used to attach a thread to a ticket or forward it to a colleague.
- Get Admin Live Chat Link Generates an admin portal URL where an agent can take over and reply to a live conversation. Pairs directly with the On Live Chat Escalation trigger.
Knowledge and Confidence
5- Query Knowledge Base Runs vector similarity search over a bot's trained knowledge and returns the matching context chunks without generating an answer. Used to inspect what the bot would retrieve, or to debug why an answer was wrong.
- Get Confidence Score Scores how well a bot's knowledge covers a given conversation, returning average and per chunk cosine similarity plus the exact embedding text used. This is the value flows gate on before letting an answer stand.
- Upload Training File Downloads a document from a URL and uploads it to a bot as a training file. Returns a storage URL carrying a time limited access token; Retrain Chatbot folds it into the knowledge base afterwards.
- Delete Training File Removes an uploaded training file from a bot's storage by its storage URL. Retrain Chatbot is still needed if the content already reached the knowledge base.
- Generate File Access Token Issues a fresh access signature for a stored file so it can be downloaded without an API key. Used when an earlier upload URL has expired.
Leads and Customers
2- Attach Lead To Conversation Associates contact details with a conversation, creating the lead when no match exists and updating it when one does. Matching uses every identifier supplied together, so an email or phone gives deterministic results.
- List Customers Retrieves the leads a bot has captured with filtering by time window, traffic source and any contact identifier. Returns a total count, which makes it the basis of a CRM sync.
Reporting and Plan
2- Get Chatbot Analytics Returns the statistical breakdown for a bot over a window: conversations, messages and credits per day, volume by hour, visitor country distribution, detected topics and sentiment, votes, ratings, captured leads and live chat volumes.
- Get Current Plan Returns the account's subscription and remaining allowances including message credits, training characters and chatbot count. Checked before a bulk training or messaging run.
Channels
2- Send WhatsApp Message Sends an outbound WhatsApp Web message from a bot to a phone number, optionally with attachments and optionally as synthesised audio. Requires the bot's WhatsApp Web integration to be connected.
- Set Channel Bot Activation Turns automated replies on or off for one connected WhatsApp number or Facebook and Instagram page, leaving the bot's other channels untouched. Used to pause automation during staffed hours.
Chatbot Actions
4- List Chatbot Actions Returns every agentic action configured on a bot in the order they are evaluated: the API calls, form submissions and tool integrations it can invoke mid conversation.
- Get Chatbot Action Retrieves one action's full definition including endpoint, headers and parameter schema, addressed either by name or by its index in the list.
- Add Chatbot Action Adds an agentic action so the bot can call an external API, collect structured data or trigger a workflow during a conversation.
- Delete Chatbot Action Removes an agentic action so the bot can no longer invoke it. The deleted definition is returned, so it can be restored if the removal was wrong.
Team Access
4- Create Sub-User Invites a team member to the company account and grants access to specific bots, with permissions ranging from read only through training to full management.
- List Sub-Users Returns every sub-user with their assigned bots, permissions and visible analytics sections. Credentials are never returned.
- Update Sub-User Changes a sub-user's name, password, bot assignments, permissions or visible analytics. Assignments and permissions replace the existing lists rather than merging with them.
- Delete Sub-User Removes a sub-user and revokes their access to every assigned bot immediately. The bots and their data are unaffected.
Triggers
3 triggersEvent Triggers
3- On Chat Completed Fires after the bot answers a visitor, delivering the question, the generated answer with its confidence score, the conversation ID, the traffic source and the credits the turn consumed. FlowRunner registers the webhook on the selected bot and removes it when the flow is disabled. This is the trigger that turns answer quality into something a workflow can act on.
- On Live Chat Escalation Fires when a visitor asks to be handed to a human, delivering the conversation history, the lead's details, the notification email and a handling URL an agent can open to take over. Used to page an on call responder the moment the AI is out of its depth.
- On Lead Submission Fires when a visitor submits contact details through the bot's lead form, delivering the captured name, email, phone and platform identifiers along with the conversation they came from. Used to push new leads into a CRM or notify sales.
Frequently Asked Questions
What can FlowRunner do with Chat Data?
FlowRunner agents can run Create Chatbot, Retrain Chatbot, and Get Chatbot Status in Chat Data, plus 33 more actions.
Does connecting Chat Data to FlowRunner require OAuth?
No. Chat Data connects to FlowRunner with an API key, no OAuth flow required.
Can Chat Data trigger a FlowRunner workflow automatically?
Yes. Chat Data supports 3 triggers that can start a FlowRunner workflow automatically.
Start building with Chat Data
$100 in credits. No card required. Connect in minutes.