Nine in ten tasks don't need a frontier model. What two new studies reveal about AI costs

A young woman with long red curls in a black knit sweater sits at a light wooden table in front of a coral wall and places a small glowing orb into a long wooden tray that holds a row of other small orbs. To the right, set apart on the table, a large glowing orb rests on its own wooden stand. Top right a card: 6 to 11 percent of agent-ready work needs a frontier model, Wharton and Accenture, November 16, 2026. Tasks scored 11,339, same success rate of Luna and Opus 5.5 89.1 percent, cost per task $0.05 vs $1.62, 32 times. Databricks AI Index, November 16: cost measured per finished task, not per token

TL;DR. On November 16, two studies came out that together send a clear message. The Wharton School and Accenture scored 11,339 tasks from 923 occupations and conclude: in every industry, only 6 to 11 percent of the work AI agents could take over needs a frontier model. On the same day, Databricks measured what models cost per finished task and found a price gap of 32 to 1 at the same success rate. Most companies don't overpay for AI because models are expensive, but because they use the most expensive one for routine work. And looking only at the token price is just as misleading.

What Wharton and Accenture found

The report “What Does Agentic AI Really Cost?” starts from O*NET, the US catalog of occupations. The researchers first sorted more than 18,000 tasks and work activities by whether a software agent can do them at all. 11,339 remained. Each one was scored on five dimensions: How many interdependent steps? Do regulations or audit duties apply? How many systems are affected? Does the task need sensitive data? And how much data, and how fresh?

The result is clear. More than 90 percent of all tasks fall into the lower tiers, where a cheaper model will do. Measured by the working time that goes into these tasks today, only 6 to 11 percent is truly demanding in each of the 19 industries studied.

Of 100 hours of work an AI agent could take over, this many need a frontier model

Routinedemanding in some industriesdemanding in every industry

Source: Wharton and Accenture, “What Does Agentic AI Really Cost?”, November 2026, page 9. Share of the working time that goes into agent-ready tasks today.

What counts as routine is hardly surprising: preparing invoices and reports, maintaining employment records, compiling absence statistics, writing advertising copy. Demanding work includes drafting contracts and patent applications or building production tracking and quality control systems. The hard tasks also tend to be the rare ones. The report sees exceptions in capital markets, tech and defense, where the frequent tasks are often the complex ones.

Costs differ accordingly. For a sample company with 100,000 employees and 35 agents, the authors estimate five years of operation at about $155 million in health care and about $520 million in insurance. The largest share is compute, in other words the models.

What Databricks measured

Wharton estimates, Databricks measures. The new Databricks AI Index analyzes the traffic of more than 10,000 companies through its own AI gateway and has models solve real coding tasks from Databricks' own codebase, checked with hidden tests.

Average cost per solved coding task, in dollars

Source: Databricks AI Index, November 16, 2026. Costs based on the providers' list prices, tasks from Databricks' own codebase.

Two numbers stand out. GPT-6 Luna, OpenAI's small model, and Claude Opus 5.5 both solve 89.1 percent of the tasks. Luna costs 5 cents on average for that, Opus $1.62. And Gemini 3.8 Flash, a model with a low token price, ends up at $11.97 per task because it needs 140 turns on average. A cheap token does not make a cheap task.

Companies are already reacting. In July, about 64 percent of customers relied on a single model provider, according to Databricks; by early November it was 28 percent. Over the same period, the share of open models in the traffic rose from 7.3 to 30.9 percent.

The catch

  • The scores came from an AI. At Wharton, two language models, GPT-5.6 Sol and Claude, rated how hard each task is, not observers of real work. The basis is a US occupational catalog, not a European company.
  • The costs are a floor. Consulting, integration, training and redesigning processes are explicitly excluded. The authors themselves recommend a 15 to 25 percent buffer.
  • One test, one codebase. The Databricks comparison rests on coding tasks from a single codebase, and the report does not say how many. For your quotes or complaints, the ranking may look different.
  • Both are selling something. Accenture implements agents for clients, Databricks sells the gateway companies use to switch between models.
  • Switching automatically is hard. According to the report, only a few companies can run a system that routes each request to the right model on its own.

What this means for mid-sized companies

Wharton's numbers apply to large corporations, but the direction applies to every company. Dividing the five-year costs by employees and years gives a rough $310 to $1,040 per head and year. Smaller companies tend to pay more per head, because costs grow more slowly than size: one and a half times the headcount, only 1.2 times the cost.

That is exactly why a close look pays off. If you build an agent for incoming invoices, shipping notices or standard inquiries and use the strongest model simply because it is there, you pay many times over in operation. The strongest model belongs where a mistake gets expensive: contracts, inspection reports, decisions with liability. For the rest, a small one is often enough.

What you should do now

  1. List tasks, not tools. For every planned agent, write down which individual tasks it handles and how often.
  2. Rate each task with five questions. How many steps? Do regulations apply? How many systems? Sensitive data? Lots of data or fresh data? Wherever the answer is “little”, it is routine.
  3. Measure the cost per finished task. The token price says little. In a trial run, count what one completed case costs and how often it is right.
  4. Start small, upgrade with reason. Let the small model compete against the large one on 50 real cases. Use the large one only where the small one measurably fails.
  5. Choose the model at the design stage. Decide per task before the agent runs. Automatic switching can come later.
  6. Plan a buffer. 15 to 25 percent on top of the estimated costs, as the authors themselves advise.

The question is rarely whether a model is good enough. For most tasks, several are. The question is which one is enough for which task. If you would like to go through your planned agents this way, we are happy to sit down with you.

A conversation worth having, even without a mandate.

20 minutes, no obligation. Usually a personal reply the same working day.