Large Language Models
Relay is model-agnostic, selecting the best model for each task. The better the models get, the more ambitious the experiences we can build.
Audit lab / built in public
Relay.auditThe lab for agentic audit
Re-evaluating the audit lifecycle and execution model for the modern AI era. Relay.audit experiments with agentic workflows.
Latest from the buildThe audit industry is at a tipping pointOnly 28% of inspected audits met the relevant audit standards in 2025.OpenAIHermesModal
Relay Experiment
The foundation of the Relay Experiment.
Relay is built across the technical and audit layers required to make agentic execution useful, inspectable, and governed inside real engagements.
Agentic Harness
The agentic harness is the orchestration engine turning general-purpose LLMs into audit-specific agents. LLMs are necessary but insufficient for audit work.
Data and Integrations
Native integrations with your audit methodology, engagement software, email, Excel, and more. The aOS operates as part of the firm's ecosystem, not alongside it.
Context and Knowledge
The agent is constantly evolving, learning from current-year and prior-year workbooks in real time while referencing client emails and submitted source documents.
Audit Capabilities
Audit-specific domain knowledge with reference to local and international standards such as IFRS and GAAP, ISA methodology, and Office and Google suite workflows.
Product and Interface
Purpose-built surfaces for every audit workflow: drafting, research, preparation, client delivery, and workpaper finalisation. Built from the ground up for audit.
Security and Governance
The foundation that makes the aOS enterprise-ready: ethical walls, cross-matter isolation, audit trails, and secure data storage according to your organisation's requirements.
Experiment track
Audit execution, rebuilt as agent workflows.
We are breaking the audit lifecycle into small, testable agent runs: planning, evidence requests, population testing, exception analysis, and reviewer-ready write-ups.
Please prepare an analytical procedure on the client's expenses.
Inspecting client's financial statements
Gaining understanding from dynamic knowledge base on the client
Checking materiality of the client
Inspecting audit methodology for sample collection
Creating workbook and selecting sample from the population
Awaiting review before sending to client
Audit Bench
Benchmarking audit judgement, not chat.
Audit Bench is our AI evaluation framework for real-world audit work. It uses lifelike audit scenarios to benchmark how different models plan, inspect evidence, apply methodology, and produce reviewer-ready output.
Audit Bencheval-042
GPT-5ClaudeGeminiLlamaCustom agent
Real audit files, rebuilt as controlled cases
Synthetic-but-grounded workpapers, source docs, populations, exceptions, and reviewer prompts.
Designed with chartered accountants
Cases are shaped with practising audit judgement so models are tested against how audit work actually breaks.
Compare agents on audit execution
Benchmark planning quality, evidence use, sampling decisions, exception handling, and final write-up quality.