Share with your CMO
Gartner’s August 2026 forecast punches a hole in the standard agentic AI business case: even as token prices fall roughly 95% by 2030, inference costs per agentic workflow are forecast to rise more than fivefold through 2028. The mechanism is compounding multipliers: a reasoning agent costs roughly 2.5 times more to train, five times more to run, and consumes five to thirty times more tokens per task than a simple chatbot. That arithmetic produces a per-task cost up to 150 times higher than a deterministic flow for the same job, and customer service, with its millions of annual contacts, absorbs the hit first.
What this means for your business
The exposure question isn’t whether your organization uses agentic AI; it’s whether your approved business case has ever seen the 150x number. Most haven’t. The standard pitch runs on token deflation, the assumption that falling per-token prices will gradually improve unit economics. Gartner’s analysts call this the token-deflation illusion, and the forecast makes it concrete: a 95% drop in token cost still produces a fivefold rise in total workflow cost because each generation of capability consumes more, and more expensive, reasoning cycles. Organizations that signed multi-year agentic automation commitments on a simple cost-per-interaction model are holding an assumption that the research directly contradicts.
The 150x figure deserves unpacking because it isn’t a worst case, it’s the output of compounding ordinary design choices. Hardware cost for a medium-sized reasoning model runs 2.5 times higher than a comparable chatbot. Inference for that model costs five times more. Then the model uses five to thirty times as many tokens to complete the same task. Multiply those together and you get a range whose top end is $1.50 for work a deterministic script does for a cent. Customer service concentrates this risk for four structural reasons: interaction volume dwarfs every other enterprise function, conversations involve clarification loops that each trigger fresh reasoning cycles, failed agent attempts burn the reasoning budget before handing off to a human who then costs full price, and the industry-wide shift from flat-rate subscriptions to consumption pricing transfers all of that variance directly onto service budgets.
The recoverable version of this situation runs through what the Gartner analysts call inference tiering, routing each contact intent to the cheapest model class that can actually resolve it. Retrieval questions don’t need a reasoning agent. Password resets don’t need planning-class inference. The cost of an interaction is set by the reasoning tier invoked, not by the vendor’s rate card, which means it responds to design decisions your team controls right now. The organizations that come out ahead won’t be the ones who spent least; they’ll be the ones who can answer, for each of their top twenty contact reasons, which reasoning tier is handling it and what it costs per resolved contact. If that number doesn’t exist in your next operating review, the 18-month clock the article describes is already running.
Concept deep-dive: Inference tiering
Inference tiering means routing each AI task to the least computationally expensive model class capable of handling it correctly, rather than sending everything through a single powerful reasoning system. Think of it like a hospital triage protocol: a sprained ankle shouldn’t consume an ICU bed. In agentic AI deployments, basic lookups, summaries, and scripted resolutions sit at the cheap end; open-ended planning and multi-step judgment sit at the expensive end. Matching task to tier is the primary cost lever available to service operations teams.
Based on reporting from The Inference Paradox: Will Cheaper Tokens Make Customer Service Cost More?, originally published 2026-10-26 17:42:00.
