Instagram Chatbot with an AI Information-Retrieval System and Voice Transcription

The problem

A sports supplements company with 22 specialized products was handling its Instagram DMs entirely by hand. The catalog demanded detailed technical knowledge: ingredients, dosages, contraindications, recommended combinations.

The team spent 3 to 4 hours a day answering questions that were often the same. Replies took hours. And messages arrived at all hours — as text and, very often, as voice notes.

The real bottleneck: it wasn't the volume of messages. It was consistency. Every person who replied gave slightly different information. Some messages simply went unanswered.

The solution

We built a chatbot powered by an AI system with information retrieval that lives inside an automated workflow and uses a vector database. The flow works like this:

1. An Instagram DM comes in (text or voice note)
2. If it's a voice note, a speech-to-text model transcribes it in seconds
3. The message (or transcription) enters the information-retrieval system: it's embedded and searched against the 22 catalog PDFs
4. Claude generates a contextual answer, citing the exact product and the relevant page
5. The reply goes back to the customer through the Instagram inbox
6. If the model's confidence is low, the message is escalated to a human with the context already prepared

The chatbot doesn't improvise: it only answers with what's documented. If it doesn't know, it says so and escalates.

Why we chose a different transcription provider

The best-known provider charges per minute of audio: for this volume, between $0.006 and $0.012 USD per minute. The one we chose runs the same base model at $0.04 USD per hour of audio — that's literally 89% cheaper.

The quality is identical because it's the same model. The difference is the inference infrastructure, which is also significantly faster.

Results since launch

The system has been in production since November 2025. The team went from 3-4 hours a day handling DMs to under 30 minutes of review (only the escalated cases). Response consistency improved visibly — it no longer depends on who's on shift.

The client can now scale the catalog (add more PDFs) without touching code. The system updates itself: upload the new document to the knowledge base and re-run the embeddings pipeline.