Wharton and Accenture: 90% of work AI agents could do fits a cheaper model
A Wharton-Accenture study finds 6% to 11% of agent-ready work needs a top-tier model. AI programme leads should set a model tier for each agent before it is built.
More than 90% of the tasks that AI agents could take over are routine enough to run on a less expensive model, according to a report from the Wharton School and Accenture published on 16 November 2026 and [reported by Fortune](https://fortune.com/2026/11/16/ai-ferrari-profit-loss-wharton-accenture/). In each of the 19 industries it studied, only 6% to 11% of the work agents could do is complex enough to typically need a frontier model, meaning a vendor's newest and most expensive one.
For whoever owns the AI budget, that makes the model each agent runs on a line to decide, not a default. Eric Bradlow, Wharton's vice dean of AI and analytics and a co-author, told Fortune he doubts most companies assign different models to different parts of a workflow: "I think they're taking possibly their most sophisticated model and applying it to all of the workflow."
What the researchers counted
The report, "What Does Agentic AI Really Cost?", covers more than 18,000 tasks across 923 occupations. The authors first set aside tasks an agent could not take on, judged by how much human interaction and physical activity each involves. That left 11,339. They scored each on five measures: how many interdependent steps it has; whether legal, regulatory or audit requirements apply; how many computer systems it touches; whether it needs sensitive data; and how much data it needs.
In life sciences, Fortune reports, enrolling people in programmes and drafting copy scored low, while developing quality-control systems for production and drafting patent applications scored as complex. In capital markets, technology and aerospace, the high-volume tasks tend to be the complex ones, so top-tier models fit more often. "It's not an argument against frontier models," said James Crowley of Accenture, a co-author.
The bill: $155 million to $520 million over five years
For a company with 100,000 employees running 35 agents, the report estimates five-year costs from about $155 million in health care to about $520 million in insurance. Divided out, that is $310 to $1,040 per employee per year ($155 million ÷ 100,000 ÷ 5; $520 million ÷ 100,000 ÷ 5), or about $0.9 million to $3 million per agent per year ($155 million ÷ 35 ÷ 5; $520 million ÷ 35 ÷ 5). Fortune says AI compute, the metered charge for running the models, is among the largest costs in most of the industry estimates. Bradlow's account of why bills rise as prices fall: "My total cost is going up" even as "my per-task cost is going down", because the company asks for 100 things it did not ask for a year ago.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
What the report cannot show
These are scores on a task framework and modelled estimates, not measurements of live deployments. Fortune's account does not say how much a company saves by moving routine work to a cheaper model, what share of the estimates is compute, which models count as frontier, or how the complexity scores were checked against real results. It gives no figure for smaller companies and nothing on staffing: who chooses the tier, or who reviews a cheaper model's output. Accenture sells AI consulting.
The authors also advise against the obvious shortcut. Dynamic model routing, where software sends each task to a different model as the work runs, is hard to track in live operations and few enterprises can run it at scale, the report says. Crowley told Fortune the model choice "often happens too late": people use the model they already have access to.
What to do: set a tier for each agent
- AI programme leads: for every agent in build or planned, record the model tier it needs and why, using the report's five measures, before a vendor or engineer picks one. Start with the highest-volume routine jobs.
- Finance: ask for a monthly statement per agent showing model usage charges beside the value claimed, which the report calls an AI P&L.
- IT: ask each agent vendor which model its product runs on, whether you can choose a cheaper one, and what happens to the price if you do.
Take this to work
Forward this
A Wharton and Accenture report published on 16 November 2026 found that more than 90% of the tasks AI agents could take over are routine enough for a less expensive model, and that only 6% to 11% of that work in any industry needs a top-tier one, as Fortune reports. Before the next agent is built, it is worth recording which model tier it needs and reporting its monthly cost beside the value it returns.
Questions to ask
- Which model does each of our agents run on today, and who chose it?
- What did each agent cost in model usage last month, and what value do we count against that?
- Can we choose a cheaper model inside this vendor's product, and does the price per seat or per resolution fall if we do?
- Who checks output quality when a routine task moves to a cheaper model, and what error rate sends it back?
- Which of our agents' tasks fall under legal, regulatory or audit requirements or use sensitive data?
The arithmetic
five-year agent cost ÷ employees ÷ 5 = cost per employee per year; five-year agent cost ÷ number of agents ÷ 5 = cost per agent per year
The report's estimates for 100,000 employees and 35 agents: $155 million ÷ 100,000 ÷ 5 = $310 per employee per year in health care; $520 million ÷ 100,000 ÷ 5 = $1,040 in insurance. Per agent: about $0.9 million to $3 million a year. Fortune does not give the share that is compute or the saving from a cheaper model.
Questions this article answers
How much AI agent work needs a frontier model?
A Wharton and Accenture report published on 16 November 2026 found that only 6% to 11% of the work AI agents could do in each of 19 industries is complex enough to typically need a frontier model. More than 90% of agent-ready tasks are routine enough for less expensive models, as Fortune reports.
What does the Wharton-Accenture report say agentic AI costs?
For a company with 100,000 employees running 35 agents, it estimates five-year costs from about $155 million in health care to about $520 million in insurance. That works out to $310 to $1,040 per employee per year.
Should a company route tasks to different AI models automatically?
The report recommends deciding what model capability each agent needs at the design stage instead. It says dynamic routing between models is difficult to track in live operations and that few enterprises can run it at scale.
What is an AI P&L?
It is the report's name for a profit and loss statement that tracks what AI costs alongside the value it generates. Co-author Eric Bradlow told Fortune it ends the assumption that using a language model is free.
GXO targets $50M-$70M from warehouse AI, and more from tracking labour
GXO told investors warehouse AI and robots should add $50M-$70M a year by 2030, less than labour tracking. Warehouse leads: measure labour first.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.