← News

Agents & MCPSources: arXiv · 18 Nov 2026

ATLAS-Finance benchmark shows AI agents struggle with realistic banking work

Researchers introduced ATLAS-Finance, a benchmark of 100 end-to-end investment-banking tasks across 13 simulated firms, and report that today's best frontier agents pass fewer than 11% of tasks.

In plain words

A research team built a test that asks AI agents to do the kind of work investment bankers do, such as digging through email, chat and files to build financial models. Even the strongest models completed much of the work but rarely produced correct final deliverables. The benchmark's tasks normally take human bankers 15-30 hours each.

What was announced

The benchmark's 100 tasks were created by experienced bankers based on their real deal work, and each task normally takes a human 15-30 hours to complete. Mirroring professional workflows, agents must search email, chat, calendars, notes and files to identify requirements, gather information, and reconcile conflicts or updates before building complex financial models. The tasks span 13 simulated firms and reflect assignments where information is spread across workplace applications and may change over time.

Deliverables are graded against expert-authored rubrics with a median of 124.5 criteria, covering financial rationale and results, spreadsheet construction, completeness and formatting. According to the authors, despite completing substantial portions of the requested work, current agents rarely produce correct deliverables: the best-performing frontier models, Claude Opus and Fable, pass fewer than 11% of tasks. The findings come from the benchmark's own abstract, so they reflect a single evaluation framework rather than independent confirmation of the models' broader capabilities.

What it does not show

  • Evidence is limited to the paper's abstract; full methodology, task examples and per-model results are not available in the supplied passages.
  • Pass rates describe performance on this specific benchmark and do not establish how the models perform in real banking jobs.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.