Skip to main content

NEPTUNE OPS / GUIDES

How to evaluate AI software for a public affairs team

Start with a real task, compare evidence and review effort, and define what a successful pilot must demonstrate.

By Neptune Ops · Updated

Choose a task before choosing a tool

Write down one recurring task, its current owner, the information it needs, and the output someone actually uses. A useful pilot has a narrow question: can this workflow produce a sufficiently accurate, reviewable result with less total effort?

A research brief, an internal summary, or a first draft can be easier to evaluate than a broad promise to transform a department.

Create a small evaluation set

Collect representative examples your team can assess. Include an ordinary task, an ambiguous request, a case with missing evidence, and a task where the correct response is to ask for more information. Keep a few examples out of the initial vendor demonstration.

  • Define the expected result and unacceptable errors.
  • Record the sources a reviewer should check.
  • Use the same task set when comparing products.
  • Include review and correction time in the comparison.

Ask for evidence and uncertainty

For a research task, inspect whether each material claim has a relevant source. Open the source and confirm that it supports the claim. For a drafting task, check the draft against the supplied facts and instructions. A polished answer can still be wrong.

Ask the vendor to show how the workflow handles conflicting information, stale records, missing context, and corrections from your team.

Assess the operating requirements

Document who can access information, how retention works, which inputs may be used for training, and what export or deletion options are available. Have the appropriate internal owner review the vendor's actual terms and configuration.

Confirm onboarding, support, integration, usage limits, and the complete pricing basis. Compare the proposed engagement with the capabilities your team already has.

Make the pilot decision explicit

Before the pilot starts, set a review date, a budget, and the criteria for continuing. Track accepted outputs, material errors, reviewer time, and task completion. These are proposed evaluation measures, not claims about a particular product's performance.