Skip to content
← Back to Adoption

Wharton and Accenture: only 6–11% of work AI agents could do needs a frontier model

More than 90% of tasks suit cheaper models, the authors say. Accenture sells AI services, and the costs are modelled, not measured.

6–11%of work agents could do that needs a frontier model, per industry

Wharton’s AI & Analytics Initiative and Accenture built a model that prices AI agents task by task. It starts from the US government’s O*NET job database, which lists more than 18,000 tasks and work activities across 923 occupations. Two language models, ChatGPT 5.6 Sol (Pro) and Claude, scored the 11,339 tasks judged suitable for software agents on five complexity measures, such as how many systems a task touches and how much regulation applies.

Most of that work is routine. More than 90% of tasks fall in the lower-complexity bands that the authors say cheaper models can handle. The complex tasks that tend to need a frontier model make up 6% to 11% of agent-ready work volume in each of the 19 industries studied. The report’s examples of frontier-level work are designing experimental operational models with no existing data and drafting patent applications. Preparing invoices and compiling hiring statistics sit at the other end.

The model then adds up eight upfront and five running cost components, with compute among the largest in most industries, Fortune writes. For a company of 100,000 employees running 35 agents, Fortune reports five-year costs of about $155 million in health care and about $520 million in insurance. The report itself says two firms of similar size can differ more than threefold, mostly in compute. About a quarter of the 35 agents account for half of the cost in its life sciences example.

What it means

By this report’s count, a company paying top-model prices for all its agent work would overpay for most of it. The authors advise choosing the model tier for each workflow when it is designed, since routing between models live is hard to run at scale. Total cost depends on the company’s own mix of work, so the report offers no single price per agent.

Caveats

  • Accenture sells AI consulting and co-wrote the report, and its staff are among the authors.
  • The costs are modelled from the authors’ assumptions, not measured spending by companies. The report calls them a cost floor: they leave out consulting and integration fees and change management, and it suggests adding 15–25% as contingency.
  • Which tasks count as agent-ready, and how complex they are, was judged by two language models, not by testing agents on the tasks. A separate Accenture study cited in the report used a decision-tree classifier and reached a similar result.
  • The $155 million and $520 million figures come from Fortune’s article. They appear in the report as a chart, and the text we could read does not state them.
Suggest a correction