I built this model (Mada, my co-founder, wrote the article with me). It uses 3-7 independent Opus-tier analyst agents to read the entire deal (emails, transcripts, CRM) and propose credit allocations, and then a consensus judge weighs their reasoning to determine the final report. Buyer actions (form fills, replies, event attendances, etc.) always carry 0% credit because it flows to whatever got the buyer interested enough to engage.

During development, there were multiple occasions when I'd feel sure the agents were hallucinating, only to realize they were consistently discovering true things that even I would have missed (this was right around the Opus 4.6 launch). But I'm still concerned this might be too much of a break from traditional attribution for most teams, even if it is 'more accurate'.