Summary
Actions are Dolores AI’s ability to do things beyond answering in text: assign a conversation to an agent, label a customer, trigger a chatbot, send a request to an external system. Each action is defined with a plain-language description that Dolores uses to decide when to trigger it. In this article you’ll get to know the ten available actions, how to describe when to trigger them, and how to wire them into a workflow. If you’ve set up Dolores but haven’t set up actions yet, you’re using only half of what it can do.
Actions versus responses
Without actions, Dolores only responds in text. With actions, it can make changes in the account:
| Without actions | With actions |
|---|---|
| Dolores answers the customer | Dolores answers and performs an action |
| An agent has to handle every case | Dolores handles defined cases automatically |
| Customers labeled manually | Automatic labeling based on conversation content |
| No integration with external systems | Sending an HTTP request at Dolores’s discretion |
The ten available actions
In the Actions tab on the Dolores AI page, you can configure the following actions:
- Assign to agent. Assign the conversation to a specific agent.
- Assign to team. Assign the conversation to an entire team.
- Change status. Update the conversation status (New, In Progress, Closed).
- Add to list. Label the customer with a list.
- Remove from list. Remove a list from the customer.
- Add a note. Write an internal note on the customer card that only agents can see.
- Trigger a chatbot. Hand the conversation off to a specific chatbot.
- Remove assignment. Cancel an existing agent or team assignment.
- Turn off chatbot. Stop a bot that’s running on the customer.
- Send an HTTP request. Call an external system with data from the conversation.
The structure of an action
Every action is made up of two parts:
1. Description: when should Dolores do this?
Free text that explains to Dolores in which situations to trigger the action. This is the most important part. The more precise the description, the more the action triggers at the right times.
Example of a poor description: "When the customer wants to buy"
Example of a good description: "When the customer explicitly says they want to order, asks to see payment options, or asks about shipping. Not when they're only asking about price or requesting product information."
The sharpness of the description is the difference between an AI that works and an AI that triggers actions at the wrong times.
2. The action: what Dolores will do
Choose one or more of the ten actions above. You can combine several actions into one (for example: label, assign, and add a note all at once).
How to write descriptions that work
A description is an instruction to a new employee. It should be:
- Specific. “When the customer is confused” is too vague. “When the customer writes ‘I don’t understand’ or ‘confused’, or asks the same question twice” is better.
- Unambiguous. No “if it seems like…”. Instead: “When the customer writes X or Y”.
- Includes exceptions. “This action should not run when…”. The contrast helps the AI understand the boundary.
- Short where possible. 2-3 sentences. A whole-paragraph description becomes vague.
Examples of working descriptions
| Poor | Good |
|---|---|
| “When the customer is interested in a product” | “When the customer asked about a specific product by name, asked to order, or asked how to buy.” |
| “When they need to talk to an agent” | “When the customer writes ‘connect me with an agent’ or ‘I want to talk to a human’, or asked 3 questions without getting a satisfactory answer.” |
| “When the customer is unhappy” | “When the customer requests a refund, complains about a specific product, or uses aggressive language like ‘wake up’ or ‘this is unacceptable’.” |
| “On a purchase” | “When the customer sent their order number along with the question, or when they talked about a product they already bought (for example ‘I ordered but didn’t receive it’).” |
Examples of actions that work
Action 1: Route hot leads to the sales team
- Description: “When the customer shows a clear indication they want to buy (asks for a quote, asks to schedule a meeting, asks to talk to a salesperson), or when they’ve asked at least 3 in-depth questions about a specific product.”
- Action: Assign to the sales team + add to the “Hot Lead” list + add a note “Classified as a hot lead by Dolores”.
Action 2: Automatic classification by organization size
- Description: “When the customer explicitly said they’re at an organization of 50+ employees, mentioned job titles like ‘CEO’ or ‘VP’, or talked about budgets of thousands of dollars per month.”
- Action: Add to the “Enterprise” list + assign to the Enterprise team.
Action 3: Hand off to a human agent on frustration
- Description: “When the customer expresses frustration: wrote ‘this isn’t working’ twice, used an exclamation like ‘enough already’ or ‘just tell me!’, or explicitly asked ‘connect me with a human’.”
- Action: Assign to the support team + change status to “Urgent” + add a note “Customer frustration, needs a human response”.
Action 4: Create a lead in an external CRM
- Description: “When I’ve collected all four details from the customer: name, email, business type, and main need.”
- Action: Send an HTTP request to
https://yourcrm.com/api/leadswith the data collected during the conversation.
Action 5: Close a conversation when handling is done
- Description: “When the customer said ‘thanks’, ‘okay’, ‘that’s enough’, or when the conversation reached a natural end with no further questions over a few turns.”
- Action: Change status to “Closed” + remove assignment + add a note “Conversation wrapped up by Dolores”.
The HTTP action in depth
The most powerful action is Send an HTTP request. It lets Dolores pass data from the conversation to external systems.
What the request includes
The flexibility lies in the fact that Dolores can send specific data for each case. If you want to send details you collected during the conversation to a CRM, Dolores will identify those details from the conversation and include them in the request.
Typical uses
- Creating a lead in a CRM (HubSpot, Salesforce, Pipedrive).
- Opening a ticket in a support system (Zendesk, Freshdesk).
- Adding a customer to a newsletter (Mailchimp, ActiveCampaign).
- Logging an interaction in a Google Sheet.
- Triggering a business process in an internal ERP system.
Setting up the connection usually requires an authentication key (Authorization header) on the receiving system. Make sure the receiving system knows how to authenticate the request and that the data structure matches what it expects to receive. See Webhooks in Mumble, the complete guide for details on securing connections.
Priority order between actions
If you’ve set up several actions, there may be situations where more than one action fits the same conversation. That’s why it’s important to order the actions by priority and test them in the playground before turning them on with real customers.
Implications:
- Put the specific actions at the top of the list, generic actions at the bottom.
- If one action changes the conversation status, make sure it doesn’t conflict with other actions.
- Think ahead about which actions shouldn’t run together.
Testing in the playground
Every time you add a new action, test it in the playground before you turn it on with real customers. Send a sample conversation that should trigger the action, and check:
- Did Dolores trigger the action?
- Was the action performed correctly (the label added, the webhook sent)?
- Did other actions that shouldn’t have triggered stay off?
Best practices
- Start with just one action. Define and test one action before adding more. An AI turned on with many untested actions may do the wrong things.
- Log every automatic action in a note on the customer card. Every automatic action by Dolores should create a note on the customer card that explains what happened and why. Otherwise an agent looking at the conversation later won’t understand why the customer was labeled.
- Don’t over-split your actions. 3-5 main actions is optimal. 20 actions start to complicate Dolores’s logic.
- Use a webhook for essential actions, not nice-to-have ones. An essential action like creating a lead needs a control mechanism, a retry, or a periodic check. A non-essential action can stay as a helper process.
- Monitor the success rate. Once a week, review the last 20 conversations the system identified as a match for action X. Were they all genuinely a match? If 5 out of 20 were mistakes, the description needs sharpening.
Common issues
The action doesn’t trigger at all
Common causes:
- The description is too vague. Rephrase it more specifically.
- Dolores AI isn’t active. Check that the page header says “Active”.
- The conversation didn’t match the description even though it seemed like it should. Try describing it differently.
The action triggers at the wrong times
This is the clearest sign of a vague description. Add the areas where it should not trigger. For example: “…not when the customer is only asking about price and hasn’t explicitly asked to order.”
The webhook doesn’t reach my system
Check:
- The URL is correct and runs over HTTPS.
- The receiving system actually expects a POST, not a GET.
- The data structure you’re sending matches what the system expects.
- There’s no firewall blocking Mumble’s IP.
For more information, see Webhooks in Mumble.
Dolores performs an action but doesn’t respond to the customer
In certain situations an action can run without an additional response to the customer. For example, if the action is “turn off AI”, Dolores may turn itself off and not continue the conversation. If you want it to respond before it turns off, build the description like this: “…tell the customer they’ll get a human response, and then turn off the AI.”
I want Dolores to do something before it answers, not after
If the action needs to happen before the response (for example, pulling information from a CRM in order to answer correctly), use an HTTP node in the chatbot canvas before the conversation reaches Dolores. See Nodes in the chatbot canvas.
Related articles
- Dolores AI, the getting-started guide
- Nodes in the chatbot canvas
- Webhooks in Mumble, the complete guide
- Automations in Mumble
- Managing agents and teams
Bottom line
Actions turn Dolores from an answering program into a working system. The ten available actions cover most needs, and the HTTP action lets you connect to external systems. Success depends on the quality of the description. A vague description creates mistakes. A specific, unambiguous description with examples and exceptions creates a system that works. Start with one action, see that it works, and then add more.