Your Finance Agent Needs an Evaluation Harness, Not Just a Prompt

A finance agent may provide persuasive responses, yet still make decisions lacking sufficient evidence. Thus, merely improving the system prompt is insufficient for accounting automation. The correct approach is to implement an evaluation harness, a small, repeatable test system that runs realistic finance cases through the agent and scores the results against explicit expectations.

Begin by defining the decision-making scope rather than selecting an LLM or tool framework. For a transaction-review agent, an effective contract might specify the desired action, confidence level, supporting evidence IDs, whether review is needed, and the rationale. The evaluation harness should also verify if the suggested category points to actual evidence, if the evidence is relevant, if the agent flags missing or contradictory documentation, if it stays within the approved chart-of-accounts, if the output is deterministic for auditing and comparison, and if the output meets other operational requirements.

Create a dataset incorporating deliberate ambiguity by including various transaction types such as routine cases, missing-evidence cases, conflicting cases, and adversarial cases. Store the expected behavior as structured data, ensuring regression tests are possible. Measure the workflow in layers, including schema validity checks, evidence grounding checks, policy compliance checks, decision quality checks, and operational behavior checks (latency, tool-call count, duplicate handling, and retry safety).

Ensure the tools are tested, not just the text, by recording each tool call as an event, allowing you to examine side effects such as duplicate writes, failed document lookups, tool timeouts, and write operations with idempotency keys.

Keep the evaluation process continuous, running the curated case set on every prompt, model, tool-schema, and policy change. Maintain a versioned baseline for comparing releases and, in addition, perform production sampling, tracking failures by category and setting thresholds for passing releases. The aim is not to eliminate all reviews but to ensure the agent sends the right work to review with sufficient context for efficient resolution.