Blog

Guide · November 2026

How to evaluate AI agents on your business processes

Why a public benchmark is not enough

A public benchmark tests general skill on general tasks. Your process has its own rules, thresholds, systems and exceptions. An agent can do well on a benchmark and still give a wrong refund or miss a rule that your team applies every day.

The best test set is the work your team already did. Each past case has a known outcome: what the employee did. You can compare the agent’s outcome with it.

Build the test set from past cases

In machine learning, a test set is “a set of examples used only to assess the performance” of a model “on unseen data”. The same rule applies to an agent. Keep the test cases apart from everything you use to build it.

  • Collect cases with their inputs, the steps taken and the final outcome.
  • Include the exceptions and the hard cases, not only the common path.
  • Hold back a later period of time as the test set. Write the agent’s instructions only from the earlier period. Then nothing from a test case leaks into what the agent knows.

Run the agent as the employee did, away from live systems

For each test case, give the agent only what the employee could see when the case arrived. Remove anything that shows the employee’s answer. If the case had more messages later, give them to the agent in the same order.

Let the agent act through sandboxed copies of your tools, so it can do the real steps, such as a refund or an order change, and nothing touches your live systems.

Score the outcome, not the path

An agent rarely takes the same path as an employee. Compare where the two ended up: what money moved, what the customer was told, and which rules were followed. Rate each case: better, as good, worse or harmful.

A model can do this scoring as a judge. Research on LLM judges reports “position, verbosity, and self-enhancement biases”, so test the judge before you trust it:

  • Run it twice on the same cases. It must give the same verdict.
  • Remove a key action from a case that passed, such as a refund that was never issued. It must fail the case.
  • Change only the words of a correct reply. It must still pass the case.

Decide before it goes live

Set the pass rule before you see the results, and give harmful cases their own rule. When you change the agent’s instructions, measure each new version again on test cases it has never seen.

After the test, run the agent beside your team until its answers match theirs. Let it go live only when your team approves.

How Subliminal evaluates agents

Subliminal builds an automation from a mapped workflow and tests it on real past cases it has never seen. It runs beside your team until its answers match theirs, and goes live only on your team’s approval.

At Fella Health, agents were tested on real recorded tickets with a temporal holdout and sandboxed tools. Run twice on the same 74 cases, the judge gave the same verdict on 73. The full method: the Fella Health case study.

Sources