To turn years of WhatsApp conversations into knowledge an AI agent can trust, KINTO classifies them with a closed taxonomy, transcribes voice notes and applies one admission rule: a fact enters the agent’s knowledge only if it appears in at least two distinct, recent conversations. The goal is to build the agent from what customers and staff actually say.
Why mine the history first
Chat history is the most honest record of demand a company has. In a real estate firm’s case, KINTO analyzed over a hundred thousand messages across more than four years (measured), classified with a closed taxonomy, with voice notes transcribed. Some conversations in the prior year had ended because the inventory had nothing that matched. The history also turned what the team knew into verifiable facts for the agent.
What the mining produces
- A map of real questions, so the agent’s skills follow demand and not an org chart.
- By-products with their own value. In the real estate case, a list of neighborhoods with unmet demand and a catalog data-quality report.
- Test cases. In a food manufacturer’s project, review of real conversations, including a lost sale, turned each error into a testable rule, and a customer simulator with dozens of scenarios runs before every release.
- A business reading. In a sports e-commerce brand, the audit over hundreds of conversations found that the bottleneck was not the agent’s conversion: most conversations left no data behind.
A warning about automated judges
Mining at scale needs automation, and automation can overstate. In the sports brand’s case, the automated judge overstated: only one in five “high” alerts was a real error, so no finding was accepted without human verification. KINTO uses a different model as judge from the one that wrote the answer, and still verifies by hand.
What it does not tell you
Chat history shows what was asked, not what would have been asked of a better service, and it says nothing about customers who never wrote. Mining also does not replace measuring results: in the real estate case there was no sustained traffic yet to measure the agent’s commercial impact.
Next step: AI agents for WhatsApp describes how this fits a project; Product hacking and growth covers using the same data to find growth levers.
