A model can be great at finishing a job and still be bad at helping someone else finish it.
That distinction matters because most AI leaderboards test direct output. Workplaces often deploy the winner in a different role: advising a person, guiding a cheaper model, reviewing a draft, or breaking a task into steps. A new arXiv working paper called CentaurBench tested what happens when the role changes.
The researchers evaluated ten language models across seven tasks: counseling, market analysis, meal planning, operations research, tax preparation, travel planning, and tutoring. In the automation condition, each model produced the final deliverable. In the augmentation condition, each model wrote one set of instructions for a fixed worker model, GPT-3.5-Turbo, which then did the work.
The rankings moved.
The result moved with the role. Across the nine assistant models tested in both conditions, the reported Spearman rank correlation between automation and augmentation performance was 0.48. It was not statistically significant at a conventional threshold. The best model also differed between direct completion and assistant-only guidance on five of the seven tasks.
But then came the useful surprise. The unaided worker beat every assisted condition on operations research, tax preparation, and travel planning. Its average augmentation rank across all seven tasks was second, behind one assistant model.
More capability had entered the workflow. The final work still got worse.
The paper offers several possible explanations. Guidance can be too complex for the worker, poorly matched to the task, or aimed at the wrong deliverable. Assistance also creates a coordination step. The worker has to interpret the advice, decide what to keep, and convert it into action. A brilliant plan that the downstream worker cannot use is expensive decoration. This is where ordinary model selection gets shaky. A company sees a strong score, buys access, and assumes the capability will travel into any role. But direct execution, coaching, reviewing, and collaborating are different jobs. The relevant unit of evaluation is the team outcome produced by a particular assistant, worker, task, and workflow.
CentaurBench does not settle that question. It is an initial simulation study using seven tasks, one fixed worker model, one-shot guidance, and LLM judges. No human employee received assistance. The authors say human and domain-expert validation is still needed. Judge agreement was also moderate rather than near-consensus, and it was lower for augmentation than automation.
Those limits keep the result in its proper box. They also make the operating lesson clearer: test the arrangement you plan to use.
Verification bottleneck
Verification is becoming the scarce institutional function.
- A direct-completion leaderboard can move faster than an organization can test whether the same model improves a real worker or workflow.
- Team owners have to compare unaided, assisted, and automated outcomes on the same representative task.
- Watch whether gains survive different workers, repeated interactions, domain-expert review, and workflow changes.
- Keep the acceptance record tied to the model version, role, task, rubric, and failure cases.
A model update can change the answer. A role change can change the value of the model even when the model stays the same. That means a procurement score or benchmark win is only the beginning of the evidence trail.
Opportunities
A builder could create a role-fit test pack for small teams. Pick a handful of recurring tasks, define what acceptable work looks like, and compare three conditions: the worker alone, the worker with AI guidance, and AI producing the deliverable directly. The useful output would be a decision record showing where assistance helped, where it added friction, and where human review remained necessary.
There is also room for an assistance regression test. When a model, prompt, or workflow changes, rerun the cases where guidance previously helped. Preserve the failures too. They may reveal that the assistant is overloading the worker, solving the wrong problem, or giving polished advice that cannot survive contact with the job.
The best player might be the right hire for direct execution. If the job is coaching, make it prove it can coach.
Source
- Wongchamcharoen, Gulati, Fong, and Nagaraj, CentaurBench: Benchmarking LLM Capabilities on Augmenting vs. Automating Real-World Work Tasks, arXiv working paper, August 19, 2026.