Dolores AI: A Getting-Started Guide

Summary

Dolores AI is an artificial-intelligence agent built into Mumble that can answer customers automatically, understand the context of a conversation, and perform actions in your account. This article explains what Dolores does, the three ways you can run it, how to set it up, and what you need to know before turning it on in a live account. It is not a replacement for a human agent, and it is not magic. But used correctly, it makes your customer-service team available 24/7 without hiring anyone.

What Dolores AI is

Dolores AI is an AI agent that lives inside your Mumble account. When a customer sends a message on WhatsApp, Dolores can:

• Understand the request and reply in Hebrew (or any other language the customer writes in).
• Use the specific knowledge you give it about your business (products, prices, policies).
• Perform actions in the account, such as assigning a conversation to a team, changing a status, or adding a customer to a list.
• Read images and voice notes that customers send.
• Send messages with buttons and lists, not just free-form text.

It is just as important to know what Dolores is not:

• Dolores is not a person. It will make mistakes in complex situations and will not replace human judgment.
• It does not arrive with ready-made knowledge of your business. You teach it.
• It is not free. Every message it sends costs an AI Credit. See the pricing section below.

Technically, Dolores runs on OpenAI’s GPT models. The difference between it and ChatGPT is that it is built into Mumble: it sees your WhatsApp conversations, knows your customers, and can perform actions in the account.

Who can use it

Dolores AI is available only on the Pro plan and above. On the Standard plan the feature is not available at all (it is not just that there are no credits—you cannot set it up). Monthly credits by plan:

Every message Dolores sends equals one credit, regardless of message length. An average conversation is 5–15 messages from Dolores, so 1,000 credits are enough for roughly 70–200 conversations a month. Exceeding your credit quota costs $0.029 per additional credit.

For up-to-date pricing, see the pricing page.

Three ways to run Dolores

Dolores does not have to handle every incoming message. You can run it in three different modes, including in combination:

1. As a standalone agent, answering every incoming message

Dolores receives every WhatsApp message and replies automatically. Suited to:

• 24/7 service without staffing nights and weekends.
• Frequently repeated questions that need a consistent answer.
• First-line handling of a high volume of inquiries beyond your team’s capacity.

2. Inside a chatbot, after a certain step

A traditional chatbot runs first (collecting details, routing), and at a certain point hands off to Dolores. Suited to:

• Structured collection up front (name, phone, inquiry type) followed by open questions handled by the AI.
• Flows that start in a structured way and move into open discussion.

3. As a single node inside the chatbot canvas

The AI node in the canvas toolbar runs Dolores for one step inside a complete flow. Suited to:

• A single judgment call inside a flow (for example: “classify the lead as hot/cold based on what they said”).
• Using a different prompt for different parts of the bot.

Most businesses start with mode 2 or 3 (a chatbot with Dolores as one step), because it gives more control. Mode 1 (standalone agent) is suited for once you feel confident that the prompt and knowledge really work.

Where you configure Dolores

In Mumble’s main menu there is a section called Dolores AI (available only on the Pro plan and above). This section is the central configuration for your AI.

The settings page has three tabs at the top:

The order of setup matters: settings first, then knowledge, and actions only at the end. An AI that doesn’t know who it is or who it is talking to will not perform actions correctly even if you have configured them.

Tab 1: AI Settings (the prompt)

In the first tab you tell Dolores who it is and what it knows. The prompt is divided into sections:

1. Business details. Name, email, basic information.
2. Products and services. What you sell, with the details worth knowing.
3. Price list. Pricing tiers, costs, discount terms.
4. Frequently asked questions. Recurring questions and your canonical answers to them.
5. Sales approach. How Dolores opens and structures sales conversations.
6. Additional instructions. Everything else, including the AI’s name, gender, and role.

Mumble provides 16 ready-made templates by business type (clinic, online store, real estate, SaaS, and more). You can choose a template as a starting point and edit it, or start from scratch. Each section has two helpful buttons: Tips (what to include in this section) and Help me write (generates a draft automatically).

Important: the templates fill in generic content in several places. After choosing a template, you must go through every section and edit it to fit your business. A “SaaS” template that leaves the words “Mumble” or “WhatsApp Business API” in place when you are a different business will give Dolores the wrong identity.

Tab 2: Knowledge Base

In the second tab you give Dolores the material it will rely on. Three input types:

• Paste text. A free-text area for anything Dolores needs to know: questions and answers, product specs, policies.
• Links. URLs you want Dolores to crawl and use—for example your website, pricing page, terms page.
• Documents. File uploads. Supported formats: PDF, DOC, TXT, CSV, XLSX, JSON, MD, HTML. Up to 5MB per document is recommended.

An important point worth understanding: the knowledge base is loaded into the prompt, not retrieved at conversation time. In other words, Dolores does not “search” the knowledge base every time someone reaches out. All of the information in the knowledge base is added to its prompt. The practical implications:

• There is a practical size limit. A huge knowledge base bloats the prompt and scatters the model’s attention.
• Quality beats quantity. 10 precise questions and answers will work better than 100 shallow ones.
• There is no real-time update. If you change a document on an external site, Dolores will not see the change until you reconfigure it. Dynamic content that changes often (price list, inventory) is not suited to the knowledge base; it is suited to actions that pull a value in real time.

Tab 3: Actions

In the third tab you define what Dolores can do beyond answering with text: assign a conversation to an agent, transfer it to a specific team, change a status, add a list, turn off a chatbot, send an external HTTP request, and more.

Every action is made of two parts:

1. When should Dolores do this? A plain-language description of when the action should be carried out. For example: “when the customer asks about a specific product price.”
2. Perform the following actions. One or more actions that run together.

This section requires more thought than the others, because the quality of the description is everything. A vague description like “when the customer wants to buy” will cause Dolores to trigger the action at the wrong times. A precise description like “when the customer explicitly says they want to make a purchase and is not just asking for information” will work much better.

As of now, the list of possible actions: assign to agent, assign to team, change status, add to or remove from a list, add a note, start a chatbot, unassign, turn off a chatbot, send an HTTP request.

Turning Dolores on

After you have configured everything, Dolores is still not active. You need to explicitly click “Activate AI” in the page header. Only after clicking does Dolores start replying to customers, and only then do credits begin to be consumed.

When it is active, the header shows an “Active” status in green/purple. To turn it off, there is a “Stop AI” button in red.

This is the most common mistake: customers set everything up and don’t click “Activate AI,” then think Dolores isn’t working. Always check that the status reads “Active” before you expect it to reply.

Testing in the playground

Before turning Dolores loose on real customers, you can—and should—test it. On the settings page there is a “Playground” button (in the bottom-left corner). Clicking it opens a small chat window where you can talk to Dolores as if you were a customer.

What to check in the playground:

• Tone. Does Dolores speak in natural Hebrew? In line with the identity you defined?
• Knowledge base. Ask questions whose answers are in the knowledge base and see whether it uses them correctly.
• Actions. Describe scenarios that should trigger actions and see whether it carries them out.
• Boundaries. Ask questions that are not in the knowledge base and see how it handles them (ideally: it admits it does not know and offers to route to an agent).

Testing in the playground after every significant change (a new prompt, a document added to the knowledge base, a new action) saves trouble with real customers. An hour of testing before going live turns into days of happy customers.

Recommended setup order

If this is your first time setting up Dolores, the recommended workflow:

1. Choose a Pro plan or above. The feature is not available on the Standard plan.
2. Go to the Dolores AI section. In the first tab, choose a template that fits your business type.
3. Edit every section of the prompt. Don’t leave generic template text. Add your business name, its real details, and the language you want it to speak.
4. Add an initial knowledge base. At least one of the three: a website link, a document, or text. Without knowledge, Dolores guesses.
5. Test in the playground. Send at least 5–10 messages that simulate real conversations.
6. Add one essential action. For example: “when the customer says they want to talk to an agent, assign them to the support team.” Start with one action and test, before adding more.
7. Turn Dolores on. Only after you are satisfied with the test results.
8. Monitor the first 50 conversations. Read them, check the quality, and update the prompt or knowledge as needed.

Common mistakes

• Forgetting to turn it on. Everything is configured, but “Activate AI” is never clicked. Dolores stays silent and the business thinks something is broken.
• An empty knowledge base. A prompt with no source of facts causes Dolores to invent answers. An empty knowledge base = Dolores is not ready.
• Vague action descriptions. “When the customer wants to buy” will trigger the action even when the customer is only asking for a price. A description should read like an instruction to a new employee.
• Using sensitive fields without oversight. Medicine, law, finance. Dolores can give inaccurate answers, and that is dangerous in these fields. In these cases, limit Dolores to routing to a human agent only.
• Mixing template text with partial edits. You choose a “SaaS” template, change one field, and forget the rest. Your customers get answers that mix their business with the template’s example business.
• Language mismatch. Dolores replies in the language the customer wrote in. If most of your customers write in Hebrew, make sure the prompt and knowledge are written in Hebrew too.

Common issues

I set Dolores up but it isn’t replying to customers

The most common reason: you did not click “Activate AI” after configuring it. Check that the header shows an “Active” status in green. If not, click “Activate AI.”

Dolores replies, but gives wrong or made-up answers

A clear sign the knowledge base is missing or low quality. Add content specific to your business, especially the information customers ask about (prices, products, hours). Test in the playground after each addition.

Dolores triggers actions at the wrong times

The action description is too vague. Rewrite the description more specifically, including what not to trigger on. For example: instead of “when the customer wants to buy,” use “when the customer explicitly says they want to order, and is not just asking about the price.”

My credits ran out mid-month

You may have had higher traffic than planned, or long conversations. Options:

• Upgrade to the Scale plan (5,000 credits).
• Pay overage ($0.029 per additional credit).
• Temporarily turn Dolores off, or narrow the scenarios in which it is active.

Dolores replies in English even though my customers write in Hebrew

Check that the prompt is written in Hebrew. If the prompt is in English, Dolores starts thinking in English and sometimes replies in English even to customers who speak Hebrew. Write the prompt in the language you want it to reply in.

Dolores isn’t reading a document I uploaded

Check that the format is supported (PDF, DOC, TXT, CSV, XLSX, JSON, MD, HTML) and that the size does not exceed 5MB. If the document is valid but Dolores does not refer to it, it may have been written in a foreign language or in an unreadable format (for example, a scanned PDF without OCR). Try uploading a plain-text version.

Related articles

• AI tools in the inbox
• Building a chatbot in Mumble
• Chatbot triggers
• Upgrading or changing your plan
• Mumble pricing page

The bottom line

Dolores AI is not a switch, it is a system. Its success depends on the quality you invest in the setup: a precise prompt, a focused knowledge base, and well-described actions. A poor setup gives poor answers and disappointed customers. A good setup gives you 24/7 availability at a marginal cost compared to hiring a single agent. Invest the first couple of hours in setup and playground testing, then turn it on with peace of mind.