Field report SE-78 / R3 · sheet /operations

Operations & RestructuringFiled from the territory

Most AI Work Needs Cheaper Models, Wharton and Accenture Find

A Wharton-Accenture study of 18,000 tasks finds over 90% of agent work fits cheaper models, even as five-year costs for large fleets reach $520 million.

Grid
SE-78 / R3
Surveyed
November 16, 2026
Reading
4 min read
Correspondent
Priya Raman
Most work doesn’t need the Ferrari of AI, Wharton and Accenture find, but companies may be paying for it anyway
Fig. SE-78 / R3Most work doesn’t need the Ferrari of AI, Wharton and Accenture find, but companies may be paying for it anyway — dlohner / pixabay

Field notes

  1. More than 90% of agent-capable tasks in the framework fit less expensive models rather than frontier systems.
  2. Across 19 industries, only 6% to 11% of the work agents could perform typically needs a frontier model.
  3. The study covered more than 18,000 tasks across 923 occupations and scored 11,339 tasks.
  4. For 100,000 employees running 35 agents, estimated five-year costs range from $155 million in health care to $520 million in insurance.
  5. Wharton and Accenture published the report as the third installment in their joint series.

By Priya Raman

More than 90% of the tasks AI agents could take over in the Wharton-Accenture framework do not require a frontier model (the most expensive, most capable class of AI system), according to the study published Monday. The decision follows: reserve those systems for the complex minority and route routine work to cheaper models.

The report, What Does Agentic AI Really Cost?, examined more than 18,000 tasks and work activities across 923 occupations. In each of 19 industries, only 6% to 11% of the work agents could perform was complex enough to typically require a frontier model.

That spending pattern runs against a common buying habit: companies deploy their newest, most powerful system across the entire workflow. They often pay a premium for capabilities that routine tasks do not need.

Eric Bradlow, Wharton’s vice dean for AI and analytics and a report coauthor, doubts that many companies make this distinction. He told Fortune: “I don’t even think most companies are thinking about deploying differential models for different parts of the workflow. I think they’re taking possibly their most sophisticated model and applying it to all of the workflow.”

Bradlow described the appeal of the latest, often best-marketed model as the shiniest object rather than a necessity for routine work. Most tasks do not.

How much agent work really needs a frontier model?

Researchers classified tasks an agent could undertake, weighing the human interaction and physical activity involved. That process left 11,339 tasks, each scored on five dimensions of complexity:

  • Interdependent steps: how many linked actions the task requires.
  • Legal, regulatory and audit obligations: whether formal requirements constrain the work.
  • System reach: how many computer systems the agent must affect.
  • Sensitive-data use: whether the task involves confidential information.
  • Data demands: how much information the model needs to complete the task.

Accenture’s James Crowley, the firm’s Americas industry and enterprise lead and a coauthor, said complexity reaches beyond raw computing power. A task can become difficult because of “the workflow navigation, or the number of systems you touch, or the regulations.”

In life sciences, the report identified developing quality-control systems for production and drafting legal documents such as patent applications as complex tasks. Enrolling people in programs and drafting copy scored lower.

Crowley said the findings do not argue against frontier models, which remain essential for complex, specialized, high-stakes work.

Where do frontier models still earn their cost?

The report found that in capital markets, technology and aerospace, higher-volume tasks tended to be more complex and more suitable for frontier models. In each sector, the model has to handle the surrounding operating constraints, not merely generate an answer.

Aerospace and defense companies trying to raise production rates must closely track certifications, quality and the availability of large quantities of parts across complicated supplier networks. Crowley described that work as extremely complex and high value.

Bradlow also cited health-care research and development, where AI can identify promising combinations of molecules. Those tasks carry high complexity and cost, but successful deployment could lead to billions of dollars in revenue.

What drives the bill?

Model choice matters because AI compute (the computing capacity used to run AI systems) ranks among the largest costs in most of the report’s industry estimates. That helps explain differences across sectors. For a company with 100,000 employees operating 35 agents, estimated five-year costs range from $155 million in health care to $520 million in insurance.

Falling model prices do not guarantee a smaller overall AI bill, Bradlow said. In his example, a task from a year ago costs less today, while the workload now covers 100 tasks it did not cover a year ago. The result: per-task spending falls as total spending rises.

Bradlow and Crowley recommend tracking AI spending alongside the value it creates through an AI P&L (a profit-and-loss statement for AI costs and returns). Bradlow said this discipline changes behavior immediately because treating a large language model as free ignores its compute bill.

How should companies choose models?

The report recommends deciding how much model capability each agent needs at the design stage, before engineers build it. Dynamic model routing (sending tasks to different models as work proceeds) remains difficult to monitor in live operations, and few enterprises can operate it at scale.

This is the third report in a series published jointly by Wharton and Accenture.

Switching models is technically possible, Crowley said, but model selection often comes too late. Teams tend to use the model they already have access to rather than the one best suited to the task.

Bradlow expects AI systems eventually to accept a request, split it into parts and route each part to the appropriate model. He said that capability is not yet fully mature, but should arrive sooner rather than later as customers demand greater return on investment and lower costs through less sophisticated models.

The operating rule is plain: match the model to the task rather than buying the best system once and spreading it across everything. For most companies, routine work offers the larger cost-saving opportunity, while frontier models remain justified for the narrow set of complex, high-stakes tasks that require them.

via fortune.com/feed/ (Source)

Share this article:

More from Priya Raman

Priya Raman

Show full bio

Senior reporter covering technology and product news in corporate strategy and chief-executive leadership at CEO Frontier.

460 articles

Filed under

Adjacent sheets

« Previous articleNext article »

End of sheetSE-78 / R3 · November 16, 2026