14.8 million visitors drawn: sales agent Javo does as well as human salespeople for 1.8% of the cost

Out of 100 e-commerce sites, an agent trained on 4.6 million sales conversations increases revenue per visitor by 10% and converts significantly better than a state-of-the-art generalist model.

Does an agent specializing in sales do better than a large generalist model, or even a human salesperson? Benjamin Tannenbaum presents Javo, an agent refined on 4.6 million anonymized sales conversations collected with consent, and put him to the test in real conditions. Between May 1 and September 30, 2026, 14.8 million visitors from 100 sites, spread over 14 sectors (e-commerce, tourism, expensive services), were randomly assigned to three situations: no agent, Javo, or an agent built on a state-of-the-art model with the same interface, the same catalog and the same safeguards.

The Results

Compared to no agent, Javo increased revenue per visitor by 10%. For visitors who engage in conversation, its conversion rate is 21.6% higher than that of the generalist agent. Out of 20 sites with a human-run chat, it is equal (11.26% conversion vs. 11.20%), with no significant superiority, but for 1.8% of the cost per conversation. Return, complaint, and unsubstantiated product claims rates remained within predetermined limits.

It is a pre-print signed by a single author, who presents the product he or she is evaluating: the figures deserve to be independently confirmed.