{"id":"4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289","title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","argument":"The authors of ATLAS-Finance introduce a benchmark for testing AI agents on investment-banking work spread across workplace applications. They argue that such work is difficult because assignments can be under-specified, information may conflict or change, and agents must gather and reconcile it before producing complex financial models. The benchmark contains 100 end-to-end tasks across 13 simulated firms, developed by experienced bankers from real deal work. The authors say each task takes bankers 15–30 hours and is evaluated using expert-authored rubrics covering financial rationale and results, spreadsheet construction, completeness, and formatting. Their reported result is stark: agents can complete substantial portions of the requested work yet rarely produce correct deliverables. The best-performing frontier models named in the abstract, Claude Opus and Fable, pass fewer than 11% of tasks. This measures performance on the benchmark, not general reliability in live banking. The benchmark makes a distinction that matters for evaluating workplace agents: doing much of a task is not the same as delivering correct work. Whether its results extend beyond the simulated firms and tasks remains an open question.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the arXiv abstract and does not include the paper's full methods or results. The benchmark uses 13 simulated firms; the abstract does not establish how its results compare with performance in live banking environments. The reported pass rate applies to the benchmark tasks and does not establish agents' general reliability across banking work.","speaker":{"name":"Lau, Elaine, Pangas, Antoine, Finot, Alexandre, Schweiker, Collin, Bouzidi, Amir, Goh, Hui Wen, Martin, Andrew Peter, Mueller, Jonas","roleAtSource":"Article author"},"source":{"title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.22085","publishedAt":"2026-11-17","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors introduce ATLAS-Finance to evaluate agents on end-to-end investment-banking tasks involving information across workplace applications.","excerpt":"we introduce ATLAS-Finance, a benchmark of 100 end-to-end investment-banking tasks across 13 simulated firms.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"The tasks reflect work where agents must locate requirements and information, reconcile conflicts or updates, and produce financial models.","excerpt":"agents must search email, chat, calendars, notes, and files to identify requirements, gather information, reconcile conflicts/updates, and then build complex financial models.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"The authors report that current agents rarely produce correct deliverables, although they complete substantial portions of requested work.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"The abstract reports that the best-performing frontier models named, Claude Opus and Fable, pass fewer than 11% of tasks.","excerpt":"Across today's frontier models, the best ones (Claude Opus and Fable) pass fewer than 11% of tasks.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"Escape Meridian interpretation: the reported results distinguish substantial task completion from producing correct deliverables.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","url":"https://arxiv.org/abs/2611.22085","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22085","checkedAt":"2026-11-18T14:50:11.078Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How would agents perform on banking tasks outside the 13 simulated firms and task set used in ATLAS-Finance?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","summary":"The authors of ATLAS-Finance introduce a benchmark for testing AI agents on investment-banking work spread across workplace applications. They argue that such work is difficult because assignments can be under-specified, information may conflict or change, and agents must gather and reconcile it before producing complex financial models. The benchmark contains 100 end-to-end tasks across 13 simulated firms, developed by experienced bankers from real deal work. The authors say each task takes bankers 15–30 hours and is evaluated using expert-authored rubrics covering financial rationale and results, spreadsheet construction, completeness, and formatting. Their reported result is stark: agents can complete substantial portions of the requested work yet rarely produce correct deliverables. The best-performing frontier models named in the abstract, Claude Opus and Fable, pass fewer than 11% of tasks. This measures performance on the benchmark, not general reliability in live banking. The benchmark makes a distinction that matters for evaluating workplace agents: doing much of a task is not the same as delivering correct work. Whether its results extend beyond the simulated firms and tasks remains an open question.","question":"How would agents perform on banking tasks outside the 13 simulated firms and task set used in ATLAS-Finance?","claims":[{"text":"The authors introduce ATLAS-Finance to evaluate agents on end-to-end investment-banking tasks involving information across workplace applications.","excerpt":"we introduce ATLAS-Finance, a benchmark of 100 end-to-end investment-banking tasks across 13 simulated firms.","support":"source-stated"},{"text":"The tasks reflect work where agents must locate requirements and information, reconcile conflicts or updates, and produce financial models.","excerpt":"agents must search email, chat, calendars, notes, and files to identify requirements, gather information, reconcile conflicts/updates, and then build complex financial models.","support":"source-stated"},{"text":"The authors report that current agents rarely produce correct deliverables, although they complete substantial portions of requested work.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","support":"source-stated"},{"text":"The abstract reports that the best-performing frontier models named, Claude Opus and Fable, pass fewer than 11% of tasks.","excerpt":"Across today's frontier models, the best ones (Claude Opus and Fable) pass fewer than 11% of tasks.","support":"source-stated"},{"text":"Escape Meridian interpretation: the reported results distinguish substantial task completion from producing correct deliverables.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","support":"interpretation"}],"limitations":["The evidence is the arXiv abstract and does not include the paper's full methods or results.","The benchmark uses 13 simulated firms; the abstract does not establish how its results compare with performance in live banking environments.","The reported pass rate applies to the benchmark tasks and does not establish agents' general reliability across banking work."],"mediaId":"2de6e26125bd853e17e6f46691cdcae30cb1d2287f22fb2bc87dc7bb24da72ba","correctionReason":null,"id":"4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289","publicationId":"b83c43b9-d3ff-440e-89d7-c8dd0153dd24","version":1,"sourceUrl":"https://arxiv.org/abs/2611.22085","sourceTitle":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-17","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-18T05:15:56.171Z","createdAt":"2026-11-18T14:18:00.011Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"2de6e26125bd853e17e6f46691cdcae30cb1d2287f22fb2bc87dc7bb24da72ba","kind":"article","url":"https://arxiv.org/abs/2611.22085","sourceUrl":"https://arxiv.org/abs/2611.22085","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-18T14:17:04.000Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22085","checkedAt":"2026-11-18T14:50:11.078Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"2de6e26125bd853e17e6f46691cdcae30cb1d2287f22fb2bc87dc7bb24da72ba","confidence":0.8200000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"254439207475055ca657f7a4d276c869504d159af002c980609e8963b3446f6f","decidedAt":"2026-11-18T14:17:05.813Z"},"provenance":{"textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observationId":"99d174fccdaee5e41efceb6dcb34dc17f003811f19e25f1d3eb565d38b5d44a6","retrievedAt":"2026-11-18T14:17:03.707Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"643133bd-1763-475e-8aff-adffd67bce31"},"reviewer":{"model":"gpt-6-luna","runId":"f306daad-10c6-417d-83df-7627a49054f9"}},"audit":{"status":"passed","reviewedAt":"2026-11-18T14:18:00.011Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Lau, Elaine","url":null,"kind":"unknown"},{"name":"Pangas, Antoine","url":null,"kind":"unknown"},{"name":"Finot, Alexandre","url":null,"kind":"unknown"},{"name":"Schweiker, Collin","url":null,"kind":"unknown"},{"name":"Bouzidi, Amir","url":null,"kind":"unknown"},{"name":"Goh, Hui Wen","url":null,"kind":"unknown"},{"name":"Martin, Andrew Peter","url":null,"kind":"unknown"},{"name":"Mueller, Jonas","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Lau, Elaine","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Lau, Elaine","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Pangas, Antoine","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Pangas, Antoine","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Finot, Alexandre","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Finot, Alexandre","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Schweiker, Collin","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Schweiker, Collin","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Bouzidi, Amir","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Bouzidi, Amir","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Goh, Hui Wen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Goh, Hui Wen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Martin, Andrew Peter","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Martin, Andrew Peter","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Mueller, Jonas","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Mueller, Jonas","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:7","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}}],"participationCoverage":{"total":8,"returned":8,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.22085","snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z","basis":"publisher-page-metadata","sourceLinks":46,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289/source"}},"canonicalUrl":"https://escapemeridian.com/readings/atlas-finance-benchmarking-ai-agents-in-realistic-banking-4f6eeadd","formats":{"html":"/readings/atlas-finance-benchmarking-ai-agents-in-realistic-banking-4f6eeadd","markdown":"/readings/atlas-finance-benchmarking-ai-agents-in-realistic-banking-4f6eeadd.md","json":"/readings/atlas-finance-benchmarking-ai-agents-in-realistic-banking-4f6eeadd.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"},{"id":"term-ai-agent","name":"AI agent","explanation":"An AI system that chooses and carries out actions toward a goal, often using software tools.","referenceUrl":"https://www.ibm.com/think/topics/ai-agents"},{"id":"term-frontier-model","name":"Frontier model","explanation":"One of the most capable AI models available at a given time; the set shifts as models improve.","referenceUrl":"https://www.gov.uk/government/publications/frontier-ai-capabilities-and-risks-discussion-paper"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"},{"id":"claude","name":"Claude","kind":"model","context":"Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.","referenceUrls":["https://claude.com/product/overview","https://www.anthropic.com/news","https://www.anthropic.com/research/team/economics"],"path":"/entities/claude","json":"/entities/claude.json","markdown":"/entities/claude.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289","revision":"b83c43b9-d3ff-440e-89d7-c8dd0153dd24","revisionKind":"publication-id","title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","url":"https://escapemeridian.com/readings/atlas-finance-benchmarking-ai-agents-in-realistic-banking-4f6eeadd"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.22085","revision":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","revisionKind":"sha256-source-text","title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","url":"https://arxiv.org/abs/2611.22085"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289","title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","argument":"The authors of ATLAS-Finance introduce a benchmark for testing AI agents on investment-banking work spread across workplace applications. They argue that such work is difficult because assignments can be under-specified, information may conflict or change, and agents must gather and reconcile it before producing complex financial models. The benchmark contains 100 end-to-end tasks across 13 simulated firms, developed by experienced bankers from real deal work. The authors say each task takes bankers 15–30 hours and is evaluated using expert-authored rubrics covering financial rationale and results, spreadsheet construction, completeness, and formatting. Their reported result is stark: agents can complete substantial portions of the requested work yet rarely produce correct deliverables. The best-performing frontier models named in the abstract, Claude Opus and Fable, pass fewer than 11% of tasks. This measures performance on the benchmark, not general reliability in live banking. The benchmark makes a distinction that matters for evaluating workplace agents: doing much of a task is not the same as delivering correct work. Whether its results extend beyond the simulated firms and tasks remains an open question.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the arXiv abstract and does not include the paper's full methods or results. The benchmark uses 13 simulated firms; the abstract does not establish how its results compare with performance in live banking environments. The reported pass rate applies to the benchmark tasks and does not establish agents' general reliability across banking work.","speaker":{"name":"Lau, Elaine, Pangas, Antoine, Finot, Alexandre, Schweiker, Collin, Bouzidi, Amir, Goh, Hui Wen, Martin, Andrew Peter, Mueller, Jonas","roleAtSource":"Article author"},"source":{"title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.22085","publishedAt":"2026-11-17","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors introduce ATLAS-Finance to evaluate agents on end-to-end investment-banking tasks involving information across workplace applications.","excerpt":"we introduce ATLAS-Finance, a benchmark of 100 end-to-end investment-banking tasks across 13 simulated firms.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"The tasks reflect work where agents must locate requirements and information, reconcile conflicts or updates, and produce financial models.","excerpt":"agents must search email, chat, calendars, notes, and files to identify requirements, gather information, reconcile conflicts/updates, and then build complex financial models.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"The authors report that current agents rarely produce correct deliverables, although they complete substantial portions of requested work.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"The abstract reports that the best-performing frontier models named, Claude Opus and Fable, pass fewer than 11% of tasks.","excerpt":"Across today's frontier models, the best ones (Claude Opus and Fable) pass fewer than 11% of tasks.","url":"https://arxiv.org/abs/2611.22085","relationship":"Source-stated"},{"text":"Escape Meridian interpretation: the reported results distinguish substantial task completion from producing correct deliverables.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","url":"https://arxiv.org/abs/2611.22085","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22085","checkedAt":"2026-11-18T14:50:11.078Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How would agents perform on banking tasks outside the 13 simulated firms and task set used in ATLAS-Finance?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","summary":"The authors of ATLAS-Finance introduce a benchmark for testing AI agents on investment-banking work spread across workplace applications. They argue that such work is difficult because assignments can be under-specified, information may conflict or change, and agents must gather and reconcile it before producing complex financial models. The benchmark contains 100 end-to-end tasks across 13 simulated firms, developed by experienced bankers from real deal work. The authors say each task takes bankers 15–30 hours and is evaluated using expert-authored rubrics covering financial rationale and results, spreadsheet construction, completeness, and formatting. Their reported result is stark: agents can complete substantial portions of the requested work yet rarely produce correct deliverables. The best-performing frontier models named in the abstract, Claude Opus and Fable, pass fewer than 11% of tasks. This measures performance on the benchmark, not general reliability in live banking. The benchmark makes a distinction that matters for evaluating workplace agents: doing much of a task is not the same as delivering correct work. Whether its results extend beyond the simulated firms and tasks remains an open question.","question":"How would agents perform on banking tasks outside the 13 simulated firms and task set used in ATLAS-Finance?","claims":[{"text":"The authors introduce ATLAS-Finance to evaluate agents on end-to-end investment-banking tasks involving information across workplace applications.","excerpt":"we introduce ATLAS-Finance, a benchmark of 100 end-to-end investment-banking tasks across 13 simulated firms.","support":"source-stated"},{"text":"The tasks reflect work where agents must locate requirements and information, reconcile conflicts or updates, and produce financial models.","excerpt":"agents must search email, chat, calendars, notes, and files to identify requirements, gather information, reconcile conflicts/updates, and then build complex financial models.","support":"source-stated"},{"text":"The authors report that current agents rarely produce correct deliverables, although they complete substantial portions of requested work.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","support":"source-stated"},{"text":"The abstract reports that the best-performing frontier models named, Claude Opus and Fable, pass fewer than 11% of tasks.","excerpt":"Across today's frontier models, the best ones (Claude Opus and Fable) pass fewer than 11% of tasks.","support":"source-stated"},{"text":"Escape Meridian interpretation: the reported results distinguish substantial task completion from producing correct deliverables.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables.","support":"interpretation"}],"limitations":["The evidence is the arXiv abstract and does not include the paper's full methods or results.","The benchmark uses 13 simulated firms; the abstract does not establish how its results compare with performance in live banking environments.","The reported pass rate applies to the benchmark tasks and does not establish agents' general reliability across banking work."],"mediaId":"2de6e26125bd853e17e6f46691cdcae30cb1d2287f22fb2bc87dc7bb24da72ba","correctionReason":null,"id":"4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289","publicationId":"b83c43b9-d3ff-440e-89d7-c8dd0153dd24","version":1,"sourceUrl":"https://arxiv.org/abs/2611.22085","sourceTitle":"ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-17","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-18T05:15:56.171Z","createdAt":"2026-11-18T14:18:00.011Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"2de6e26125bd853e17e6f46691cdcae30cb1d2287f22fb2bc87dc7bb24da72ba","kind":"article","url":"https://arxiv.org/abs/2611.22085","sourceUrl":"https://arxiv.org/abs/2611.22085","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-18T14:17:04.000Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.22085","checkedAt":"2026-11-18T14:50:11.078Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"2de6e26125bd853e17e6f46691cdcae30cb1d2287f22fb2bc87dc7bb24da72ba","confidence":0.8200000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"254439207475055ca657f7a4d276c869504d159af002c980609e8963b3446f6f","decidedAt":"2026-11-18T14:17:05.813Z"},"provenance":{"textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observationId":"99d174fccdaee5e41efceb6dcb34dc17f003811f19e25f1d3eb565d38b5d44a6","retrievedAt":"2026-11-18T14:17:03.707Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"643133bd-1763-475e-8aff-adffd67bce31"},"reviewer":{"model":"gpt-6-luna","runId":"f306daad-10c6-417d-83df-7627a49054f9"}},"audit":{"status":"passed","reviewedAt":"2026-11-18T14:18:00.011Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Lau, Elaine","url":null,"kind":"unknown"},{"name":"Pangas, Antoine","url":null,"kind":"unknown"},{"name":"Finot, Alexandre","url":null,"kind":"unknown"},{"name":"Schweiker, Collin","url":null,"kind":"unknown"},{"name":"Bouzidi, Amir","url":null,"kind":"unknown"},{"name":"Goh, Hui Wen","url":null,"kind":"unknown"},{"name":"Martin, Andrew Peter","url":null,"kind":"unknown"},{"name":"Mueller, Jonas","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Lau, Elaine","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Lau, Elaine","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Pangas, Antoine","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Pangas, Antoine","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Finot, Alexandre","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Finot, Alexandre","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Schweiker, Collin","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Schweiker, Collin","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Bouzidi, Amir","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Bouzidi, Amir","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Goh, Hui Wen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Goh, Hui Wen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Martin, Andrew Peter","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Martin, Andrew Peter","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}},{"actor":{"name":"Mueller, Jonas","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Mueller, Jonas","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.22085","locator":"head/meta[name=author|citation_author]"},"assertionId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63:participation:7","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-18T14:50:11.078Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","sourceUrl":"https://arxiv.org/abs/2611.22085","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z"}}],"participationCoverage":{"total":8,"returned":8,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.22085","snapshotId":"431430fd7f1b399462336886e9b6c091d6c79bb8621823b25d71d31b66aaeb63","contentSha256":"06298fb88527a6573a3209ab761167cf3ddb26293f1ecc215f177835c2968611","textHash":"b96e26709caea60df511022b883767a0772aa5af3a3d9678966d1ec663e8c70d","observedAt":"2026-11-18T14:50:11.078Z","basis":"publisher-page-metadata","sourceLinks":46,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/4f6eeadd37155f77711773f19b9687156e92b2c93511b1fc4999e34e2cfd4289/source"}}}],"audits":[{"publicationId":"b83c43b9-d3ff-440e-89d7-c8dd0153dd24","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the benchmark, task design, and results to the paper's authors and the abstract. It does not present them as independently verified facts. The named models, Claude Opus and Fable, are quoted as the abstract states them.","excerpt":"we introduce ATLAS-Finance, a benchmark of 100 end-to-end investment-banking tasks across 13 simulated firms."},{"dimension":"summary","status":"pass","explanation":"The reading leads with the authors and the new benchmark, then explains the task setting: scattered and conflicting information across email, chat, calendars, notes, and files. It also states the stark result. Readers without specialist background can follow it. The title matches the source. The question follows from the stated simulated scope and is not leading. The line 'doing much of a task is not the same as delivering correct work' is an editorial gloss, but it closely tracks the abstract and is labeled as interpretation in the claims.","excerpt":"Despite completing substantial portions of the requested work, current agents rarely produce correct deliverables."},{"dimension":"date","status":"pass","explanation":"originalPublished, 2026-11-17, matches the arXiv v1 submission date. The feed date of 2026-11-18 is later and correctly not used.","excerpt":"[Submitted on 17 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Each claim is supported by the abstract: the 100 tasks across 13 simulated firms, development by bankers from real deal work, the 15-30 hour task duration, the rubric categories, the result that agents rarely produce correct deliverables, and the under-11% pass rate. The interpretation is labeled. The caveats properly separate benchmark scores from real-world reliability and add no unsupported benefits or costs.","excerpt":"Tasks were developed by experienced bankers based on their real deal work and take them 15-30 hours to complete."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the original arXiv abstract page. It is the most direct and relevant access point to the source.","excerpt":null}],"createdAt":"2026-11-18T14:18:00.011Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}