The 2026 Accounting AI Benchmark

The methodology used for this benchmark was to divide a set of specific, domain-related questions into different categories. These categories represent the core workflow of a general accounting system.

Questions were designed against a provisioned chart of accounts and a minimal context capable of providing the information required for the questions to function without loading too much information into the prompt.

Each benchmark runs in an isolated environment per organization, without any link to a real account in our system. Each run is agnostic to the others.

The grading is deterministic, meaning there is no “reasoning” behind the answers beyond a simple binary-logic decision.

Each benchmark is allowed to run multiple times, computing accuracy, standard deviation per category, and difficulty tier.

The whole benchmark is task-oriented, not trivia-based, giving us the flexibility to perform actions in our system such as delegate_to_record_draft or use other tooling systems expected for the agent.