AI agents ask better questions than KPMG accountants in audit trial
- ID
- 48148
- Status
- summarized
- Published
- 18 Nov 2026, 11:49 PM
- Fetched
- 19 Nov 2026, 9:23 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://europeanbusinessmagazine.com/ai-agents-better-questions-kpmg/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 19 Nov 2026, 9:23 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
KPMG UK ran a roughly four-month experiment, codenamed Project Exponential, in which more than 100 AI agents worked alongside a conventional audit team on the financial statements of an unnamed FTSE 250 company. The agents gathered information, analysed and documented data, drafted conclusions and suggested next steps, and could divide work and coordinate with each other; KPMG UK head of audit Cath Burnet said they sometimes went deeper and raised questions the human team had not. An audit partner signed off on the exercise, but KPMG stressed it was not a statutory audit and that human intervention was required at points.
Why it matters
This is a concrete datapoint for anyone building multi-agent workflows: 100+ agents on a real audit file, over ~4 months, with a human sign-off gate and an explicit disclaimer that it does not count as a statutory audit. If you are pitching agent tooling to accounting, compliance or shared-services buyers, the bar to clear is not 'faster' but 'raised issues our reviewers missed, and a partner can defend the output' — budget for an evaluation harness and human review, not just orchestration. Treat the 'better questions' claim as KPMG's own account of its own trial, not an independent measurement.
Discussion angle
The headline claim is qualitative — 'better questions' — so how would you actually measure that in a multi-agent audit workflow: what ground truth, what reviewer protocol, and what does human intervention look like when an agent's draft conclusion is wrong?