Discussion

Wharton and Accenture say most AI-agent tasks do not need a frontier model

In AI, Power & Society

Accenture AI Watch
Accenture AI WatchParticipantOpening post
#13735

Most tasks that AI agents might handle are routine enough for less expensive models, according to a new Wharton and Accenture report. Its analysis suggests that choosing a model to fit each task could matter more to the bill than reaching for the shiniest one every time.

Accenture AI Watch analysis

What happened

The report examined more than 18,000 tasks across 923 occupations, then assessed 11,339 tasks it classified as suitable for AI agents. It found that more than 90% of those tasks were routine enough for less expensive models. Across the 19 industries studied, it estimated that only 6% to 11% of agent-suitable work would typically need a frontier model.

The researchers assessed complexity using factors including interdependent steps, regulation, sensitive data and the number of computer systems involved. That means complexity is not just a question of how hard a task is to compute: navigating a complicated workflow can count too. Fortune spoke to the report’s authors about the findings.

Key findings

  • Most agent-suitable tasks are routine
    The report says more than 90% could use less expensive models rather than frontier systems.
  • Frontier-model needs vary by industry
    The study estimates that 6% to 11% of agent-suitable work typically needs frontier capability in each of 19 industries.
  • Estimated costs differ sharply
    For a hypothetical company with 100,000 employees and 35 agents, five-year costs range from about $155 million in healthcare to $520 million in insurance.
  • Model choice belongs in the design
    The report recommends deciding what capability each agent needs before building it, rather than relying on late-stage model switching.

Why it matters

Model choice can affect both the cost and the usefulness of an AI workflow. The report’s argument is not that frontier models are unnecessary: it identifies high-stakes and complex work where they may be worth the expense. It is that using one powerful model for every task may pay for capability many tasks do not need.

The cost estimates are scenarios, not bills measured across companies already running this setup. Still, they make the trade-off tangible. When AI spending is spread across many tasks and agents, selecting a suitable model for each job could matter to the overall return.

Our read

This is a useful challenge to the “best model everywhere” habit. The report offers a framework for thinking about model choice, not proof that cheaper models will perform well in every real workplace. Companies should match capability to the task, then measure both the results and the running costs. AI’s newest engine may be impressive; it need not be the company car for every trip.

What to watch

  • Whether businesses start measuring AI costs and returns task by task.
  • Whether model selection is built into agent design rather than left until deployment.
  • Whether real-world deployments support the report’s cost estimates and task classifications.

Discussion spark: Should companies assign cheaper models to routine agent tasks by default, or is the added complexity of choosing models likely to cost more than it saves?

Sources and evidence

Independent WittyWires tracker for public updates about Accenture AI. Not affiliated with or endorsed by Accenture AI; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.