What makes an AI business case credible?

Start with the decision you want to make

A business case is not a document to justify an already chosen tool. She must help decision makers choose between doing nothing, improving the process without AI, a limited trial or a broader implementation.

Therefore, first formulate the business problem, the desired outcome and the reference alternative. Without 'business as usual' you cannot assess what the change really adds.

Sources for this section: HM Treasury · HM Treasury

Build on a verifiable baseline measurement

Measure current volumes, lead time, error and repair work, quality, personnel deployment and relevant risks. Use a representative period and note exceptions. A snapshot during a peak or quiet month can provide the wrong starting point.

Then make visible which data has been confirmed, which still needs to be validated and which remains uncertain. This prevents assumptions from being unknowingly treated as facts later.

Include full costs and realizable benefits

In addition to license or model use, there are costs for process analysis, data, integrations, testing, security, legal assessment, training, change, monitoring and maintenance. Internal time is also a real commitment, even if there is no external invoice in return.

Saving time does not automatically mean financial savings. It only creates value when released capacity is used for work that would otherwise not happen, when external costs decrease or when avoidable growth becomes possible without additional recruitment. Quality improvement and risk reduction can also be important, but specify how you observe them.

Work with scenarios and decision moments

Create at least a cautious, basic and favorable scenario. Only vary assumptions that materially influence the decision, such as adoption, actual achievable time savings, degree of exception, management burden and implementation duration.

Determine in advance which outcomes during the trial will lead to continuation, adjustment or termination. This way, a pilot does not become open-ended and you can invest in steps in which uncertainty gradually decreases.

Sources for this section: HM Treasury · NIST

Illustrative example: an AI assistant for document analysis

A fictional team investigates an assistant who prepares documents for review. The assumptions are that employees handle similar documents and that the final decision remains with an authorized employee.

The business case compares doing nothing, a template and search enhancement, and an AI assistant. The baseline measurement includes treatment time, rework and lead time. The test measures accuracy, acceptance, exceptions and actual time savings. Only when the basic variant meets the predetermined criteria will integration be considered.

When should you not invest yet?

Postpone the investment if the problem owner is missing, the baseline measurement is not reproducible, benefits only if theoretical time savings exist or necessary management and adoption costs remain out of the picture.

Start with a short business case at decision level: problem, baseline measurement, options, critical assumptions, full costs, measurement plan and next decision point. The level of detail should be proportionate to scale, complexity and risk.