The weakest AI business cases begin with a dramatic percentage. The strongest begin with a measured workflow.

A credible return model for diligence should separate preparation, professional judgment, review, and rework. AI can reduce some of these components, but it rarely removes all of them.

Establish the current baseline

Measure one repeated workflow over several matters:

  • hours spent collecting and organizing sources;
  • hours spent extracting or calculating information;
  • hours spent drafting the first work product;
  • hours spent checking sources and correcting mistakes;
  • elapsed time from source arrival to reviewer-ready output;
  • number of review cycles;
  • number of late source changes that trigger rework.

Use actual time records or a short observational sample. Estimates from memory tend to undercount checking and coordination.

Model the assisted workflow

For a pilot, track the same components. Add the new work created by the system:

  • configuring the source scope;
  • resolving failed or incomplete runs;
  • reviewing AI-supported claims;
  • correcting workflow logic;
  • maintaining templates and permissions.

This prevents the common mistake of treating generation time as total delivery time.

Calculate value in three buckets

Capacity value is the reduction in hands-on preparation and rework.

Cycle-time value is the benefit of reaching a reviewable answer or deliverable earlier.

Control value is the reduction in effort required to reconstruct sources, review status, and versions.

A simple model is:

Annual value = capacity recovered + cycle-time benefit + avoided control work − platform and implementation cost

Keep the control value conservative unless the organization already measures audit preparation, error remediation, or repeated source checking.

Use a decision threshold

Define success before the pilot starts. For example:

  • a measurable reduction in first-draft preparation time;
  • no increase in reviewer time;
  • source support available for every material claim;
  • fewer review cycles;
  • acceptable failure and correction rates.

The purpose of an ROI model is not to prove that AI works. It is to decide whether one configured workflow creates enough repeatable value to justify deployment.

How we approach this topic

This field note is based on the workflow and product-design questions we encounter while building Underlying. It is educational, not legal, investment, or security advice. Product examples describe design patterns unless explicitly stated as generally available.

Reviewed by Underlying Product Team.