Token Cost per Lead

Once chatbots, product advisors or outbound agents produce a meaningful share of leads, classic cost per lead falls short, because variable cost no longer sits in the media budget but in the inference bill. TCPL aggregates all model cost of a conversational path and divides it by the number of qualified leads. Clean attribution is essential: only sessions belonging to the funnel may count, and abandoned conversations count as cost without return. Only this metric shows whether an elaborate agent with six reasoning steps really converts more cheaply than a lean form with good retargeting.

TCPL decides whether AI funnels scale. Without it, conversation volume and model cost quietly grow faster than lead quality, and a seemingly successful assistant becomes uneconomic at scale. With it, model choice, context size and conversation length can be optimised deliberately against return.

A B2B software vendor generates 320 qualified leads per month via a website agent at EUR 1,180 inference cost — a TCPL of EUR 3.69. After moving the qualification step to a smaller model with cached context, the value falls to EUR 1.42 with unchanged lead quality according to sales feedback.

Until 2025 inference cost counted as a rounding error in marketing projects. With the shift to reasoning models and multi-step agents, cost per conversation rose noticeably through 2025, in some cases tenfold versus simple chat calls. In the first half of 2026 performance teams therefore began treating token cost as its own line in funnel economics. TCPL established itself as a standard reporting metric after several vendors introduced usage-based billing per conversation, making cost attributable at session level.