Why Gartner thinks AI-agent bills will quintuple even as token prices keep falling

Per-workflow cost for agentic AI is on track to climb more than fivefold by 2028, according to Gartner, even as the per-token price of running a model keeps falling. The analyst firm calls the gap the Inference Paradox.

Agentic AI is the term for systems that take actions and chain tasks across tools and APIs, rather than just answering a prompt. A basic chatbot produces one short reply. An agentic workflow might draft a plan, query a database, write code, and email the result, paying for dozens of model calls along the way.

Gartner's Aug. 17 forecast is a projection, not a measured result, and it depends on adoption pace, model mix, and hardware cost curves holding on the trajectory the firm assumes. The firm framed it as a planning problem for product leaders. "Product leaders cannot rely on more efficient token economics to rationalize AI costs," Will Sommer, senior director analyst at Gartner, said in the release. "Defaulting to generic autonomous intelligence will result in unbounded costs orders of magnitude higher than those of optimized product ecosystems."

Three forces drive the curve. First, the unit cost of inference keeps dropping, the same trend the same firm predicted in March 2026 when it forecast that running a 1-trillion-parameter model will cost providers 90% less by 2030 than in 2025. Second, cheaper inference unlocks more capable and more expensive models, which then get used on higher-value tasks. Third, an agentic workflow consumes far more tokens than a chatbot exchange. The same analyst told CIO Dive in March that each query in an agentic assistant "costs five to 30 times more tokens" than a chatbot query.

The two Gartner forecasts do not contradict each other. They describe different denominators. The March release is about the cost per token. The August release is about the cost per workflow. An agentic task can use an order of magnitude more tokens than a chatbot exchange, so a steep per-token decline still leaves a per-workflow bill that climbs.

The August release puts a per-task number on the gap: routing a request to an agentic reasoning model costs "at least 5x" a basic chatbot interaction, and "often much more" as task complexity rises. The fivefold-by-2028 forecast is layered on top of that 5x baseline, not a replacement for it.

Gartner's remedy in the release is "inference tiering, routing and orchestration": send easy tasks to cheap models and reserve the expensive ones for cases where the answer has to be right. The full client report is titled "The Inference Paradox: Inference Tiering Is Critical to Protect Margins." That title is the buyer's question. Which tier does each call land on, and who decides when a task escalates?

The forecast lands on three different ledgers. Enterprises running agents absorb the bill directly as workflows grow. Vendors absorb it when they price per-seat or per-call and lose money on heavy users; Sommer told CIO Dive in March that labs are still losing money today, so customers will not see the full unit-cost decline. End users feel it through per-seat pricing on agent products, or through thinner features when vendors throttle expensive tiers.

The August forecast also sits next to a June 2025 Gartner call that over 40% of agentic-AI projects will be canceled by end of 2027, on cost, unclear value, or weak risk controls. A Forbes column in July 2026 framed that 40% as a management and governance problem, not a model-capability one, the same shape as the new forecast.

Market pressure points the other way. Major labs have cut API prices through 2026, and alternate accelerators are undercutting the dominant GPU on suitable workloads, the unit-cost trend the March forecast describes. The August forecast is the warning that those savings do not survive a workflow that does ten times the work.

The number worth pressure-testing is not the price-per-token figure. It is the routing layer. Any vendor selling an "AI agent" product should be able to answer three questions: which tasks land on which model, who decides when a task escalates from a cheap tier to an expensive one, and what would change the answer.