LinkedIn·Tuesday, 25 August 2026·1d ago

Gartner, same analyst: token prices down 95% by 2030, cost per agent workflow up 500% by 2028. They call it the Inference Paradox. I call…

Andrea Balducci

Fractional Revenue Operations & AI Builder @ SiSu | ex. Sortlist (Deloitte Fast 50 winner)

Gartner, same analyst: token prices down 95% by 2030, cost per agent workflow up 500% by 2028. They call it the Inference Paradox. I call it the line nobody budgeted for. A chatbot reads one question and answers once. An agent reads, plans, calls a tool, checks its own work, finds a gap, and goes again. Gartner puts the floor at 5x the cost of a simple chat interaction, before the task gets hard. Picture an agent that loops 12 times to close one CRM task. Cut the token price in half. You still pay for 12 runs to get one result. 𝗧𝗵𝗲 𝘂𝗻𝗶𝘁 𝗼𝗳 𝗰𝗼𝘀𝘁 𝗶𝘀 𝘁𝗵𝗲 𝗳𝗶𝗻𝗶𝘀𝗵𝗲𝗱 𝗷𝗼𝗯, 𝗹𝗼𝗼𝗽𝘀 𝗶𝗻𝗰𝗹𝘂𝗱𝗲𝗱. RevOps already thinks this way. A rep's salary means nothing until you divide it by meetings booked, SQLs, or closed deals. Agents deserve the same line in your unit economics: 𝗰𝗼𝘀𝘁 𝗽𝗲𝗿 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗲𝗱 𝗼𝘂𝘁𝗰𝗼𝗺𝗲. Three things I do before an agent goes live at a client: → Tier the models. A small, cheap model does the mechanical steps (classify, extract, format). The reasoning model only runs where judgment is needed. → Cap the loop. Hard limit on retries and self-checks per run. Past the cap, the task goes to a human, context attached. → Log cost per outcome in the workflow itself. Every run writes tokens, steps, and result to a table. The dashboard shows euros per updated deal, next to what a human costs for the same task. Without that last number, you are steering on token prices. Gartner has a name for that too: the token-deflation illusion. 𝗪𝗵𝗮𝘁 𝗱𝗼 𝘆𝗼𝘂 𝘁𝗿𝗮𝗰𝗸 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝘁𝗼𝗱𝗮𝘆? Token spend, cost per completed task, or nothing yet? P.S. Source: Will Sommer and Sabine Zimmerhansl, Gartner. Links in the first comment. #RevOps #AIAgents #GTM #SalesAutomation