Challenge
The author of a personal development programme wanted to guide participants through 30 days in a WhatsApp conversation, personally and in her own voice. With several hundred people at once this cannot be done by hand, and an ordinary chatbot would destroy what people pay for: the feeling that someone is listening.
The channel added its own limits. Meta lowers sending limits for numbers with a low quality rating, and block reports can end in a ban. The conversations are about emotions, so the system had to recognise crisis situations and hand them to people straight away.
Approach
We based the answers on the author’s knowledge base (hybrid search, semantic and full-text, with reranking) and on a short- and long-term memory of each participant. A second, cheaper model scores every answer. When confidence is low, crisis words appear or attempts keep failing, the conversation goes to the coaches’ queue.
Before launch we analysed how WhatsApp scales. Meta’s limits apply to the company’s whole portfolio, not to a single number, so adding numbers does not increase throughput. We stayed with one provider, with a second number only for redundancy, and set communication hours so that nobody is messaged in the evening or at night.
Implementation
The backend collects messages in a queue and merges a series of short messages from one person into a single statement before replying. The team dashboard shows live conversations, the escalation queue, broadcasts with targeting and a knowledge base with document upload. A testing module checks resistance to attempts to pull the model out of its role.
After the first week we reviewed 747 messages. The audit found, among other things, repeated phrases and no escalation after many failed attempts. We rolled out the fixes while the programme was running, without interrupting conversations.
Results
The programme covered several hundred participants, and the system handled tens of thousands of messages. AI resolved 92% of cases without human involvement. The rest went to the coaches, with the full context of the conversation.
In the first 11 days the median response time was 4.8 s, and on the busiest day the system accepted 516 messages with no drop in quality. AI cost in that period was 13.20 USD, or about 0.022 USD per answer, including search and quality scoring.
