To train Papo's artificial intelligence agent (https://papo.global/) with real conversation examples, export transcripts of previous calls or chats, convert them into text documents, and add them to the assistant's Knowledge Base in the Prompt and AI module. This process teaches the AI to recognize the vocabulary, questions, and tone your customers actually use, increasing the automatic response rate and reducing transfers to humans.
Summary of the training process with real conversations
| Step | Action | Expected result |
|---|---|---|
| Obtain conversations | Export audio from recorded calls or copy chat texts (WhatsApp, Instagram, etc.). | Raw material with customers' real language. |
| Transcribe (if needed) | Convert audio to text with tools such as Google Speech-to-Text, Azure, or Whisper. | Searchable text ready for analysis. |
| Prepare documents | Clean sensitive data and save as .txt or .pdf. | Anonymous, organized content for the AI. |
| Add to the Knowledge Base | Upload the documents in the agent's Documents section in the Prompt and AI module. | The assistant starts consulting real speech to formulate responses. |
| Create FAQs from patterns | Identify frequent questions in the transcripts and register them as FAQs. | Direct answers to the most common questions that appear in conversations. |
| Track improvement | Check in the service reports whether the automatic response rate has increased. | Continuous adjustments based on real data. |
Why use real conversations to train the AI?
The interactions your team has already handled — whether by phone, WhatsApp, Instagram, or chat — reveal exactly how customers express themselves. Unlike static documents or hypothetical questions, transcripts show slang, synonyms, common objections, and even typos that the AI needs to interpret. By feeding the Knowledge Base with this material, you reduce the number of messages the assistant does not understand and that need to be transferred to a human agent.
This method is especially useful for businesses that already have a significant volume of recorded conversations and want to accelerate the AI agent's learning curve without relying only on assumptions.
How to obtain real conversation examples
Papo offers different ways to access interaction history, depending on the channel:
- Phone calls: If your account has call recording enabled, you can export the audio files directly from the call history in the Calls module. Then use an external transcription tool (such as Google Speech-to-Text, Microsoft Azure Speech, or Whisper) to convert the audio to text.
- Chat conversations (WhatsApp, Instagram, Facebook, etc.): You can manually copy the text of relevant conversations directly from the inbox. For larger volumes, Papo's API allows you to extract message history programmatically, making it easier to collect many examples.
- Email and other channels: The same principle applies: copy the body of messages exchanged with customers and save them as text.
Regardless of the source, the important thing is to gather interactions that represent real service, sales, or support situations.
Step-by-step to train the AI agent with real examples
- Gather the conversations. Select interactions that cover your business's main topics: product questions, complaints, quote requests, appointments, etc. Prioritize conversations in which the customer used natural language, with regional variations or informal terms.
- Convert to clean text. For audio, transcribe and review errors. For chats, copy the text and remove sensitive personal information (document numbers, cards, addresses). Save each conversation or group of conversations in a .txt or .pdf file.
- Add the documents to the Knowledge Base. In the Papo panel, access the Prompt and AI module, locate the agent you want to train (usually the Captain) and go to the Knowledge Base section. Upload the transcription files as documents. The platform processes the content and makes it available for the AI to consult during automatic responses.
- Create FAQs based on the patterns found. Analyze the transcripts and identify questions that repeat frequently. Register these question-and-answer pairs in the same agent's FAQs area. Example: if several conversations ask “What is the delivery time for the countryside?”, create an FAQ with that exact question and the correct answer.
- Refine the agent's prompt (optional). In the Prompt and AI module, you can review the main prompt text to incorporate common expressions that appeared in the conversations. This helps the AI adopt the tone and vocabulary your customers expect.
- Track the results. After a few days, check the Prompt and AI module statistics and the service reports. See whether the automatic response rate has increased and whether the number of transfers to humans has decreased. Adjust the documents and FAQs as needed.
Who this training is for
- Companies that already have a large volume of recorded calls or historical chats and want to accelerate the assistant's setup.
- Operations that want to capture customers' real vocabulary, including slang, synonyms, and regional expressions.
- Teams that notice the AI agent does not understand some questions because they are phrased differently than expected.
- Businesses that want to reduce the manual work of writing hypothetical FAQs and go straight to what customers actually ask.
Frequently asked questions
Can I use chat conversations instead of calls?
Yes. The method works with any text from real interactions: WhatsApp, Instagram, Facebook, email chats, or even notes from in-person service. The important thing is that the content reflects customers' natural language.
Do I need to manually transcribe every conversation?
No. For audio, use automatic transcription tools. For chats, simply copy the text. If the volume is very large, Papo's API can help extract message history programmatically.
Does the AI learn automatically from conversations in real time?
Papo's AI agent uses the Knowledge Base and FAQs you provide. It does not learn on its own from each new conversation; training is done by you when you add documents and adjust the prompt. That is why feeding the base with real examples is essential.
How many conversations are needed to see improvement?
There is no fixed number. Even a few dozen well-selected transcripts covering the main topics can already make a difference. Ideally, include a variety that represents the different types of interaction in your business.
Can I use conversations from channels other than WhatsApp?
Yes. The Knowledge Base is shared by the AI agent, regardless of the inbound channel. Conversations from Instagram, Facebook, Telegram, email, or SMS are equally valid for training the assistant.
How do I know the training worked?
Track the Prompt and AI module statistics and the service reports. An increase in the percentage of successful automatic responses and a reduction in transfers to human agents indicate that the AI is understanding questions better.
Read also
- How to optimize the Knowledge Base to improve the Captain's RAG search
- How to view the Prompt and AI Module statistics
The first user on the Silver plan is free forever, no credit card required — start training your AI agent with real conversations today.