Accenture's EnterpriseVAL: LLMs Clear Only 35% of Finance Tasks Without a Critical Failure
Large language models complete only 35% of enterprise finance and accounting tasks with no critical failures, according to EnterpriseVAL, a model-readiness benchmark Accenture and Faculty premiered alongside Accenture’s new Data-as-a-Service offering. The first release covers order to cash, procure to pay, record to report, and financial planning and analysis.
The benchmark asks a different question from the leaderboards most buyers see. It scores whether a model can finish real enterprise work without a failure serious enough to matter — the question a finance leader faces before letting an agent touch the ledger. Accenture’s John Crosthwait frames the shift directly: the challenge is moving from access to powerful models to making them perform in real business environments.
Data-as-a-Service is built around that gap. It combines verified enterprise specialists who create, label and evaluate training data; pre-built, quality-checked datasets developed by practitioners; and reinforcement learning environments for testing models against realistic enterprise scenarios, including safety evaluations and red teaming. Accenture says its expert network spans 35 industries and 73 languages. In one client example it developed 500,000 prompts across 15 enterprise domains for supervised fine-tuning ahead of a model release. Its target customers include frontier AI labs, which puts a large consultancy into the expert data and evaluation market that supplies them.
Verification keeps showing up as the weak point. A community study on the Hugging Face blog found that 18,640 of 47,293 public SWE-agent submissions that changed product code — 39.4% — came after the agent ran no test or script following its last edit, and those unverified submissions had 0.68 times the odds of resolving the issue. For companies planning finance agents, the practical read is to demand task-level evidence of no critical failures on their own processes before rollout, and to keep a human checkpoint wherever the benchmark says models still miss. A 35% clean-completion rate is a scoping tool, not a verdict: it tells you which tasks to hand over first and which to hold back.