As of 13 November 2026, 11 of the 13 published benchmarks checked here show the best AI model still short of expert-level results on real professional work; 2 show models close to or at full marks. The widest gaps are in finance and legal work seeded from a real deal room (about 0), construction drawings where the defect is only visible in the drawing (5%), and long healthcare and finance tasks (under 25%). Generic office tasks in simulated worlds are close to solved: one benchmark's top score went from 24.0% in its January paper to 84.2% on its revised task set in November.
- Models fail most where the answer depends on messy evidence across many files (deal rooms, audits, claims) or on geometry and drawings (CAD, construction, chips).
- Several benchmark authors say real proprietary records were the missing ingredient: the mortgage benchmark used synthetic loans and the audit benchmark synthetic engagements.
- Benchmarks built from simulated office tasks are saturating quickly, so generic workflow data loses value faster than specialist records.
- A benchmark score is not a price. It shows where developers still need examples of real work, which is why the records behind these sectors are the ones they ask for.
Published benchmarks by sector
Compiled by Origin Data Partners research. Scores are the best reported in each paper or on the benchmark's own page, in the paper's own metric, so rows are not comparable with each other. The CSV holds the same rows.
| Sector | Benchmark | Published | What it measures | Real records used? | Best reported score | Reading | Source | Origin guide |
|---|---|---|---|---|---|---|---|---|
| Several (chip design, cyber, biology, finance, legal and more) | GDPx (SemiAnalysis) | 2026-11-02 | Expert tasks in 11 domains, built with 14 environment vendors | Partly: finance seeded from a real private-equity deal room; chip tasks from taped-out silicon | Best model 12% overall | Gap remains | SemiAnalysis | Linked company records |
| Finance and law | GDPx finance and legal tasks | 2026-11-02 | Binary-graded finance and legal tasks | Finance: about 4,000 anonymized files from a real private-equity transaction | About 0 for every model | Gap remains | SemiAnalysis | Deal data from investment firms |
| Healthcare and finance | DAYJOB | 2026-10-01 | 130 long professional tasks (50 healthcare, 80 finance), graded on every rubric item | Tasks written by working professionals; client records not stated | Best model passes 24.7% (healthcare) and 23.9% (finance); median setup 0.6% and 2.5% | Gap remains | arXiv 2610.01306 | Healthcare claims and prior authorizations |
| Insurance claims | InsClaimBench | 2026-10-07 | Claim adjudication from rules to payout decision and amount (3,780 cases) | Yes: real claim materials (auto, property, health) | Joint decision and amount accuracy 47.54% to 73.15% | Gap remains | arXiv 2610.09671 | Insurance claims files |
| Audit | FinancialAuditBench | 2026-09-26 | 90 workpaper completion and review tasks in six audit engagements | No: synthetic engagements; the authors cite limited access to proprietary data | Best model 69.17% on one attempt, 44.44% across eight | Gap remains | arXiv 2609.32835 | Linked company records |
| Operations research consulting | OR-License | 2026-10-16 | End-to-end solutions from original client materials, scored against the expert deliverable | Yes: 31 instances from paid consulting engagements worth $3.68M, de-identified | Best 0.87 of expert quality (expert = 1) | Gap remains | arXiv 2610.24087 | Linked company records |
| Mortgage lending | MortarBench | 2026-06-17 | Loan origination agent tasks | No: synthetic loan files, because no public benchmark existed | At most 77.1% exact match | Gap remains | arXiv 2606.19416 | Non-bank lender loan files |
| Mechanical CAD | RealCADBench | 2026-09-03 | Parametric CAD from real design intents (12,632 tasks) | Yes: real industrial design intents | Solid IoU 0.28 to 0.54 | Gap remains | arXiv 2609.03773 | Manufacturer CAD and quality records |
| Chip and hardware design | Phoenix-bench | 2026-05-13 | Agent bug-fixing on 511 instances from 114 hardware repositories | No: public GitHub repositories only | The same agent loses 37% to 58% against its software benchmark score | Gap remains | arXiv 2605.15226 | Chip design and verification data |
| Construction drawings | AEC-Bench | 2026-03-31 | Drawing understanding, cross-sheet reasoning and project coordination | Yes: real construction drawing sets | Defects that need the drawing itself: 5% mean reward (text-detectable ones: 100%) | Gap remains | arXiv 2603.29199 | Construction project records |
| Construction drawings | DrawingVQA | 2026-07-16 | 92 expert questions on 33 issued-for-construction drawings | Yes: real issued-for-construction drawings | Best 77.2%, still below graduate and young professional engineers | Gap remains | arXiv 2607.15418 | Construction project records |
| Aviation maintenance | CAMB | 2025-08-28 | Knowledge and reasoning for civil aviation maintenance | Industry maintenance material; work orders not stated | Best overall precision 68.87% | Gap remains | arXiv 2508.20420 | Aviation repair station records |
| Tax (VAT reconstruction) | kw-bench | 2026-10-26 | Five-year VAT reconstruction across three SAP generations | Built from real engagements, delivered as raw system extracts | 0.9986 and 1.000 for the two frontier setups | Solved at the frontier | arXiv 2610.36405 | None yet |
| Banking, consulting and corporate law | APEX-Agents (paper, v1.0) | 2026-01-20 | Long cross-application professional tasks | No: simulated worlds built by professionals | Best 24.0% at publication | Near saturation | arXiv 2601.14242 | Linked company records |
| Banking, consulting and corporate law | APEX-Agents (current leaderboard) | Read 2026-11-13 | Same benchmark, current 240-task version | No: simulated worlds | Top model 84.2% | Near saturation | Mercor APEX leaderboard | Linked company records |
What the failures have in common
The low scores cluster in two kinds of work. The first is work whose answer has to be defended against many messy, linked records: a deal room, an audit file, a claim with its payout, a loan file. The second is work where geometry or images carry the meaning: CAD models, construction drawings, chip designs and their bugs.
Where real records were used, the scores are lowest; where benchmark authors could not get them, they say so. That gap is what AI developers try to close by licensing real work product, which is why the linked records of a specialist firm and the sector record sets below are the ones they ask for.
Two benchmarks are included because they cut the other way: a VAT reconstruction task that frontier setups solve completely, and a simulated professional-task benchmark whose top score rose from 24.0% in its January paper to 84.2% on its revised 240-task version in November 2026. Clean, simulated tasks age fast.
Method and limits
Only published benchmarks and papers are included: arXiv papers, an analyst report and one benchmark leaderboard. Each score was read at its source on 13 November 2026 and recorded with a verbatim sentence. Benchmarks known only from podcasts, press releases or blocked pages were left out (Handshake AI's Atlas and Vials, IBM AssetOpsBench, Mercor's Physical-Engineering exam).
Scores use each paper's own metric and task set and are not comparable across rows. Most papers are preprints, not peer reviewed. Model names and versions change quickly; a newer model may already do better. The editorial and sourcing policy explains how Origin checks and corrects figures.
Real agent runs as evaluation data
Benchmarks built from real deployment traces are among the hardest published. Companies that run agents in production hold such runs; see licensing agent traces from production.
Use and cite this data
Free to reuse under CC BY 4.0 with a link to this page. Download the CSV. For what AI developers have paid for data, see the AI training data price index.
13 November 2026: first edition.
Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.
Common questions
Where do AI models still fail at professional work?
In 11 of 13 published benchmarks checked on 13 November 2026, the best model is still short of expert results: finance and legal tasks from a real deal room (about 0), construction drawing defects (5%), long healthcare and finance tasks (under 25%), insurance claim decisions with amounts (at most 73%), parametric CAD (solid IoU at most 0.54) and hardware bug-fixing.
Which AI benchmarks use real business records?
InsClaimBench (real claim materials), OR-License (paid consulting engagements), RealCADBench (real industrial design intents), AEC-Bench and DrawingVQA (real construction drawings), and the GDPx finance environment (a real private-equity deal room). The mortgage and audit benchmarks used synthetic data because real files were not available.
Are AI models close to replacing generic office work?
On simulated office tasks, scores are rising fast: one benchmark's top result went from 24.0% in its January paper to 84.2% on its revised task set in November 2026. Scores stay low where real, messy records or drawings are involved.
Cite this data
Origin Data Partners, “Where AI Still Fails at Professional Work: Benchmarks”, updated November 13, 2026. https://origindatapartners.com/guides/where-ai-fails-at-professional-work
Free to cite and reuse with a link to this page (CC BY 4.0). Download: Professional-work benchmarks (CSV)
Sources and further reading
- GDPx finance and legal tasks (SemiAnalysis)
- DAYJOB (arXiv 2610.01306)
- InsClaimBench (arXiv 2610.09671)
- FinancialAuditBench (arXiv 2609.32835)
- OR-License (arXiv 2610.24087)
- MortarBench (arXiv 2606.19416)
- RealCADBench (arXiv 2609.03773)
- Phoenix-bench (arXiv 2605.15226)
- AEC-Bench (arXiv 2603.29199)
- DrawingVQA (arXiv 2607.15418)
- CAMB (arXiv 2508.20420)
- kw-bench (arXiv 2610.36405)
- APEX-Agents (paper, v1.0) (arXiv 2601.14242)
- APEX-Agents (current leaderboard) (Mercor APEX leaderboard)