Artificial Intelligence / EXPLAINER
The AI Demo Is Over. Now Measure the Work.
An impressive answer is a beginning. A useful AI system has to improve a real task across ordinary cases, difficult exceptions and the moments when it should ask for help.
Start with a decision someone already makes
An AI demonstration usually presents the cleanest moment in a workflow: a question goes in and a convincing answer comes out. Daily work has more texture. Inputs are incomplete, names are inconsistent and people change their minds. The system's value depends on how it behaves through that mess, including the work required to check and correct its output.
A useful evaluation begins by naming a specific task. Consider a fictional support team that wants help sorting equipment-maintenance requests. The initial goal might be to identify the relevant equipment and propose a routing queue. That is narrow enough to measure. Asking the same system to diagnose every fault, authorize repairs and close tickets at once would make a failure harder to understand.
Define what a good result looks like
NIST's AI Risk Management Framework emphasizes that validity and reliability depend on the intended use and conditions of operation. It recommends realistic testing and attention to errors with different consequences. For the fictional support team, an incorrect queue assignment and an invented equipment identifier are both mistakes, but they create different kinds of work.
Build the assessment around those differences. Record whether the identifier exists, whether the proposed queue is acceptable and whether a reviewer must open the original request to resolve uncertainty. Include requests with missing serial numbers, multiple devices and unfamiliar wording. A polished explanation should not compensate for sending the task to the wrong team. The measure should reflect the decision the organization actually needs to make.
Compare against the existing process
Suppose a team processes 200 requests in a controlled trial. A manual baseline requires an average of four minutes per request. An AI-assisted process takes one minute to draft and one minute to review, while 40 difficult requests each require an additional six minutes of correction. The illustrative total is 640 minutes: 400 for drafting and review, plus 240 for correction. The baseline is 800 minutes.
That example produces a 160-minute saving, or 20 percent, under its stated assumptions. Reporting only the one-minute draft would exaggerate the improvement. The trial should also compare the quality of final assignments and the effort required to maintain the tool. A faster process that leaves more unresolved requests may simply move work to another queue.
Make uncertainty an ordinary outcome
A system can be helpful when it says that an input is insufficient. In the support example, it could return a proposed queue with a missing-information flag, or decline to assign a nonexistent identifier. The interface should make that status clear to the reviewer. Hiding uncertainty inside a confident paragraph increases the chance that someone will overlook it.
NIST's Generative AI Profile discusses confabulation and the tendency for people to over-rely on automated outputs. One practical response is to show the evidence needed to check a claim. For an equipment identifier, that might mean the matching inventory record. For a summary, it might mean the relevant passage in the original request. Verification becomes easier when the tool places the supporting material beside the proposed result.
Keep the evaluation alive
The first test set will eventually become incomplete. New equipment, new request patterns and revised routing rules can change the task even if the model itself stays fixed. A team should preserve representative cases, review errors and repeat the evaluation when the workflow changes. NASA's published AI use-case inventory illustrates how varied applications can be, from navigation to scientific analysis; a single generic score cannot describe success across them.
For the fictional team, the next release should answer a concrete question: did it reduce the recurring identifier errors without making routing worse? Keep a small group of untouched cases for that comparison and inspect the difficult examples directly. AI adoption becomes more useful when progress is expressed in completed work, fewer avoidable corrections and clearer limits. The most convincing evidence arrives after the demonstration, when the system has faced the same awkward cases as the people it is intended to help.
Sources & further reading
Primary sources behind this article. Illustrative scenarios and design judgments are identified in the text.
