Wharton and Accenture report finds most AI agent tasks don’t need frontier models A joint report from the Wharton School and Accenture, titled "What Does Agentic AI Really Cost?" and published Monday, found that more than 90% of tasks AI agents could take over are routine enough for cheaper, less powerful models, based on an examination of more than 18,000 tasks and work activities across 923 occupations in 19 industries. Across all 19 industries, only 6% to 11% of agent-capable work is complex enough to typically require a frontier model, and the report cites 2026 analysis from McKinsey, Deloitte, EY and Gartner indicating token costs account for only 20–33% of the total cost of owning and operating an AI agent system, with the rest going to orchestration, human oversight and integration. Fewer than one-third of organizations report sustained, board-level business value from agentic AI, the research found. Wharton and Accenture report finds most AI agent tasks don’t need frontier models A study of more than 18,000 tasks suggests companies may be overpaying for AI horsepower they rarely use Companies love to deploy the biggest, newest AI model for every job their agents touch. A new joint report from the Wharton School and Accenture https://cryptobriefing.com/markets/accenture/ suggests that habit may be expensive and mostly unnecessary. The research, titled What Does Agentic AI Really Cost? , was published Monday. Its central finding: more than 90% of the tasks AI agents could take over are routine enough for cheaper, less powerful models. What the report actually measured The scope is large. Researchers examined more than 18,000 tasks and work activities spread across 923 occupations. Those occupations sit inside 19 industries. That breadth matters, because it means the finding isn’t a quirk of one sector with unusually simple workflows. Across every one of those 19 industries, only 6% to 11% of agent-capable work is complex enough to typically require a cutting-edge “frontier” model. Frontier models are the top-tier systems at the edge of what AI can currently do, and they tend to carry premium pricing. Tokens are only part of the bill The report lands alongside a broader body of 2026 analysis from firms such as McKinsey, Deloitte, EY and Gartner that looks at what agentic AI costs to run in practice. That analysis indicates token costs account for only 20–33% of the total cost of owning and operating an AI agent system. Tokens are the units AI models use to process text, and they are how most providers bill for usage. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. So where does the rest of the money go? Largely into orchestration, human oversight and integration work. Orchestration means coordinating multiple models, tools and steps so an agent completes a task correctly. Integration means wiring the agent into a company’s existing software and data. The research also points to cost overruns as organizations try to move AI agent projects from pilot stage into full-scale deployment. Productivity gains, but a value gap at the top Accenture’s survey work has highlighted significant productivity improvements tied to agentic AI. Workers and teams are getting things done faster with agents in the loop. Yet the research also finds that fewer than one-third of organizations report sustained, board-level business value from agentic AI. Why the AI agent cost debate matters now Agentic AI refers to systems that don’t just answer questions but take actions: completing multistep tasks, using software tools and working with some degree of autonomy. The Wharton and Accenture findings challenge that instinct with data. If more than 90% of agent-eligible tasks don’t need frontier capability, then routing everything through a frontier model means paying top-tier rates for work that doesn’t call for them. What this means for enterprises and AI providers For companies deploying AI agents, the most direct implication is a shift toward matching model capability to task complexity. That means reserving frontier models for the 6% to 11% of work that actually demands them and sending routine tasks to cheaper alternatives. The token-cost data adds another layer of nuance. If tokens make up just 20–33% of total ownership costs, then even a dramatic reduction in model spending addresses only a portion of the overall bill. Human oversight and integration remain substantial expenses regardless of which model sits underneath. The board-level value gap may be the most consequential signal of all. With fewer than one-third of organizations reporting sustained business value at that level, executives are likely to scrutinize AI budgets more closely. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .