An activation agent for an online course uses a hybrid architecture where a deterministic code-based automaton manages the conversation flow, while an LLM handles natural language responses. This approach ensures the agent stays on topic and follows the curriculum while maintaining a conversational tone.
The cost per dialogue ranges from 2 to 5 RUB, averaging 3.8 RUB. This efficiency is achieved by limiting the agent to a maximum of two LLM calls per turn. The system uses DeepSeek-V4-Pro as the narrator and DeepSeek-V4-Flash as the classifier to optimize costs.
During a pilot study with 88 users, the agent handled up to 71 concurrent users without failures. Out of 69 users who started the course, 7 completed the program, and 10 out of 33 users who saw the offer clicked 'Interested'. The total cost for LLM usage during the entire pilot was 169 RUB.
The agent operates within a six-state automaton: introduction, training, finale, offer, refusal, and stop. The code controls the user's position in the flow, preventing the model from deviating from the intended path. A spending limit of 40 RUB per session acts as a safety mechanism.
This activation agent targets the 'consideration gap' where potential customers are interested but hesitant to purchase. With an average online school check of 10,000 RUB, a single sale covers thousands of such dialogues, making the tool highly cost-effective for converting leads.
