How I examine AI value
An AI tool being used does not establish business value. The question is what changes in the work, whether the outcome meets the required standard and what remains after the full cost is counted.
I start with a specific job, an owner and a measurable outcome. This readable method summary accompanies the redacted visual sample. It describes the approach, not a client result or a promise of savings.
What are the eight steps?
Classify. Define the work, the unit of work, its owner and what an accepted outcome looks like.
Baseline. Record how the work runs today. Include time, volume, quality, cost and the evidence behind those measures.
Redesign. Describe how the work will change. Include human decisions, exceptions and the responsibilities that remain.
Trace. Connect the proposed change to an operational result and then to a business outcome. Each link needs evidence.
Value. Estimate the benefit and include the full cost of producing it. Keep assumptions and ranges visible.
Prove. Agree how the claim will be tested, what counts as a successful result and who checks the evidence.
Realize. Check whether the expected benefit actually appeared in the work or the business results.
Sustain. Review whether the result lasts as volumes, quality, costs and operating conditions change.
What counts as an accepted outcome?
An accepted outcome is a finished piece of work that meets the quality standard and is kept. A draft, a generated answer or a completed tool call may still need checking and correction. Measure the result that the business can use.
For an AI supported customer reply, this means examining the resolved ticket, including handling time, rework and quality. For a document review or a questionnaire, define the accepted review or answer set before estimating the value.
How is the value claim built?
Trace the chain from AI working as intended, to people using it, to better work, to a changed result. Only then ask what changes in cash, margin, time available or risk. Count each benefit once and keep the baseline beside the claim.
Time released can create capacity. It becomes a cash saving only when a corresponding cost actually changes. More output needs an owner and a credible use for that capacity. Revenue growth needs evidence that the changed work contributed to it.
Which costs belong in the calculation?
Include the AI service, surrounding systems and data, integration, people checking the work, failed attempts, rework, testing, controls, training and ongoing operation. A low model bill can sit inside an expensive workflow.
For an AI agent, compare the full cost with accepted outcomes over the same period. Token prices are one input. The useful comparison also includes repeated steps, failed attempts, tool use, testing effort and human review.
How strong is the evidence?
A hypothesis, an indicative estimate, a tested claim and a realised result are different stages. The sample distinguishes Hypothesis, Indicative, Validated, Realized and Sustained. The wording of a claim should reflect its evidence.
Use ranges while uncertainty remains. Identify the data still needed, the person who owns it and the test that could change the decision. Do not turn an estimate into a realised result by changing its label.
When does a use case need a full business case?
A fuller costing is useful when better evidence could change a decision to fund, expand, contract, staff or stop the work. A smaller, reversible change may need a lighter check. Show readiness, cost, value, risk and timing alongside each other rather than hiding them in one combined score.
Read the AI readiness method or review the AI Value Analysis scope.
Book a conversation about your AI investment decision.
Download the AI economics and value methodology sample (PDF, 18 pages)