Digital AI (also known as LLM Chatbot) is a new way to build your AI chatbot in Talkative. Instead of creating a step-by-step decision tree chatbot, you write a single prompt that tells the AI how to behave, and it handles the conversation naturally from there. This feature is currently in beta.
Think of it this way: with a traditional chatbot, you're drawing a map of every possible route a conversation could take. With Digital AI, you're giving the AI a set of instructions and letting it figure out the best route in real time. Right now it's available for web chat, with more channels to follow.
Digital AI currently runs on GPT-4.1, which is hosted in US data centres. There is no option to customise the model or choose a different region at this time. If your organisation requires data to stay within the EU, this feature may not be suitable for you right now. Speak to your CSM for the latest on timelines.
How is this different from the traditional chatbot?
With the traditional chatbot, you build a visual decision tree based on the decisions your customers make. You can add an AI node that connects to your Knowledge Base, giving the bot a more flexible, hybrid feel. However, the AI node can only answer from the Knowledge Base or fail, it can't decide which branch a customer should go down next based on the conversation. This means many customers end up building chatbots with hundreds of nodes to cover every possible path, which quickly becomes unmanageable.
Left: Traditional AI node in chatbot builder.
Right: Prompt-driven chatbot builder.
With Digital AI, there are no nodes or branches. You describe what the chatbot should do and how it should handle different situations in a single prompt, and the AI manages the conversation flow itself.
When to use the Digital AI:
- Conversations don't follow a neat, predictable path
- You need the bot to handle many topics without a branch for each one
- Your current chatbot is getting large and hard to maintain
When to stick with the traditional chatbot:
- You need very precise, controlled flows with no room for interpretation
- You don't want AI handling the full conversation
- You need features like chips (clickable buttons)
🛠️ How to set up a Digital AI agent
Step 1: Create the chatbot
Search for "LLM Chatbot" in the settings bar at the top of the platform and look for the option to create a new Digital AI. Add required info and click to create.
- Queue Label — this is the internal name, used to identify the chatbot in your system. The customer won't see this.
- Chatbot First Name — this is the name shown to customers during the chat.
Step 2: Write the Prompts
The Intro Prompt controls how the chatbot introduces itself at the start of a conversation. This is the first message the customer sees, so keep it short and clear.
The Main Prompt is where you define everything about how the chatbot behaves. We'll cover how to write a good prompt in detail below.
Step 3: Connect your tools
Click the Tools tab to add the tools your chatbot needs. The most common ones are:
- Knowledge Base Search — lets the chatbot search your AI Knowledge Base for answers. If your chatbot needs to answer questions from your content, you need this.
- Queue Transfer — lets the chatbot transfer the conversation to a human agent when needed. For example, if the visitor asks to speak to someone or the bot can't help with their question.
Other tools are available depending on your setup (e.g. Interaction Data Capture if you want to collect information during the chat). Add what's relevant to your use case.
AI Knowledge Base: Complete Guide — how to build and manage the Knowledge Base.
Step 4: Test in Preview
Click Preview your chatbot. Run through the most common conversations your customers have and check that the bot handles them well. Pay attention to:
- Does it ask the right questions before giving answers?
- Does it use the correct links from your Knowledge Base?
- Does it handle edge cases (vague questions, off-topic requests, requests to speak to a human)?
- Does it stay in character and follow your prompt instructions?
Step 5: Publish
Once you're happy, click Publish to save your chatbot. You can also use Edit Metadata to update the chatbot name, avatar, transcript email settings, and inactivity timeout.
Important: Publishing the chatbot alone doesn't make it live for customers. You also need to add it to your Chat Widget and publish the widget. Go to Chat Widgets in the left-hand menu, open your widget, add the Digital AI to it, and publish the widget. Without this step, your chatbot won't appear on your website.
✍️ How to write a good Main Prompt
The Main Prompt is the brain of your Digital AI agent. The better your prompt, the better your chatbot performs. Here's how to approach it.
If you're converting from a traditional Talkative chatbot
If you already have a working chatbot with an AI Knowledge Base, you've already done a lot of the thinking. Your existing Knowledge Base prompt is a starting point, but you'll need to expand it significantly because the Main Prompt now handles everything, not just the question-answering part.
Look at your existing prompt and ask yourself:
- What does my decision tree do that the AI Knowledge Base prompt doesn't cover? (e.g. greeting, routing, handling "speak to agent" requests, ending chats)
- What are the different paths a customer can take, and what should happen on each one?
- Where does the bot currently get stuck or give unhelpful responses?
Write all of that into your Main Prompt as clear instructions.
Prompt structure that works well
A strong Main Prompt typically covers these areas:
1. Who the chatbot is and what it does
Start with a clear identity statement. What organisation is this for? What does the chatbot help with? What can't it do?
2. Key context the AI needs to know
Is there anything the AI might misunderstand about your organisation or service? For example, if you're a complaints-handling body, the AI might confuse "I want to complain" (to you, about someone else) with "I want to complain" (about you). Spell out these distinctions clearly.
3. How to handle the main use cases - step by step
This is the most important section. For each common scenario, tell the AI what to do in order. Be specific. For example, don't just say "help with complaints." Say:
- Step 1: Ask what the complaint is about
- Step 2: Check if it's in our remit
- Step 3: If yes, ask if they've already complained to the organisation directly
- Step 4: Based on their answer, direct them to the right next step
4. What to do when things don't fit
What should happen if the AI can't find an answer? If the customer asks about something outside your scope? If they get frustrated? Write instructions for these situations too.
5. Rules about links and information
If your chatbot shares links, be very explicit about the rules. LLMs will sometimes make up URLs that look plausible but don't exist. Include clear instructions like:
- Only share links from the Knowledge Base or listed in this prompt
- Never guess or construct a URL
- If you don't have a link for something, say so
6. Tone and style
How should the chatbot sound? Professional? Friendly? Brief? Should it use bullet points? Should it avoid jargon? Set the tone here.
7. Safety and boundaries
What should the chatbot do if someone is distressed? If someone is abusive? If someone shares personal information? Cover these edge cases.
Top tips:
- Be specific, not vague. "Be helpful" means nothing to an AI. "Ask what the customer's issue is before providing any links or next steps" is much better.
- Tell it what NOT to do. If there's a behaviour you've seen that you don't want, call it out explicitly. For example: "Do not share the complaints page link until you understand what the customer's issue is."
- One thing at a time. If your chatbot needs to gather information before responding, tell it to ask one question, wait for the answer, then move on. Otherwise it may dump everything in one long message.
- Test with real scenarios. Think about the five most common things customers ask, and the five most awkward things they might say. Test both.
- Repeat the important rules. If something really matters (like not making up links), say it more than once in the prompt. LLMs respond well to repetition on critical instructions.
- Iterate. Your first prompt won't be perfect. Test, spot the gaps, update, and test again. This is normal and expected.
❓ Frequently Asked Questions
Can I use the Digital AI and a traditional chatbot at the same time?
Yes. They're separate chatbots and can run alongside each other. This is useful if you want to test the Digital AI on one use case before fully switching over.
Do I still need a Knowledge Base?
If your chatbot needs to answer questions from your website content, FAQs, or documents… Yes, you still need a Knowledge Base. The Digital AI searches it using the Knowledge Base Search tool. The difference is that the Main Prompt now controls the overall conversation behaviour, and the Knowledge Base provides the content the AI refers to.
My chatbot isn't following the prompt instructions properly.
Check that your instructions are specific and unambiguous. Vague instructions like "be helpful" or "handle complaints" leave too much room for interpretation. Try rewriting them as step-by-step actions. If a specific behaviour keeps going wrong, add an explicit "Do NOT do X" instruction.
Are you removing the traditional chatbot builder?
The traditional chatbot isn't going anywhere yet. If your current setup is working well, there's no pressure to switch. However, if your chatbot flow is getting complex, lots of branches, hard-to-maintain nodes, or you need more natural conversations, the Digital AI might be a better fit. You can run both side by side.