# AI use case assessment record Rafał Mokrowiecki · mokrowiecki.pl v1.0 · 2026-11-01 · Hypothetical example A working decision record for a single use case. Complete it with the process owner, IT and relevant control functions. No automatic score or approval is generated. All figures, roles and decisions are fictional. Not client results or a benchmark. ## Context and accountability ### Company, use case, record version Enter your company name. Assign a use case ID and document version. Illustrative organisation. DEMO-AI-01, draft version 1. All data are hypothetical. ### Accountable people Process and decision owners; quality reviewer; IT; security and other required functions. Name who accepts risk. Process and decision owner: service manager. Quality reviewer: senior specialist. IT: tool administrator. Security: environment approver. Illustrative roles without personal names. ### Problem and boundaries Which activity should change, for whom, at what volume and with what effect? What is out of scope? Also consider a non-AI alternative. Drafting responses to recurring enquiries using approved guidance only. Out of scope: complaints, financial decisions and autonomous sending. Alternative: a response template and improved guidance search. ## Baseline before the trial ### Baseline Data source, dates, case count, sample selection, task difficulty, current end-to-end time, quality and spread. State gaps explicitly. Hypothetical example: 12 manually prepared responses; mean end-to-end time 20 min, range 16–25 min. Simple and medium cases. No evidence for difficult cases. Working reference BASE-01; no actual trial dates exist. ### Criteria agreed before testing Measure, unit, denominator, target and stopping threshold; who assesses? Define a serious error. Missing evidence leaves a decision open, rather than implying zero errors. Illustrative target: mean end-to-end time below 20 min while meeting quality requirements. Serious error: incorrect information about a company obligation or data disclosure. Any such error pauses the trial for investigation. Expansion also needs difficult-case assessment. ## Data and controlled workflow ### Permitted data and tool Data categories, sources, permissions, approved environment and version, retention, supplier access and required approvals. Do not enter secrets or customer data in this record. Synthetic enquiries and demonstration guidance without personal data only. Environment and model to be agreed before a real trial; this record does not authorise arbitrary tool use. ### Verification, exceptions and fallback Who checks outputs before use, and how? When does a person take over? How are errors reported, use stopped and the previous process restored? A person checks the entire response against the guidance. Uncertainty or missing support leads to manual drafting. No autonomous sending. A real deployment must name the reporting channel and person responsible for stopping the tool. ## Trial plan and staff preparation ### Plan and reproducibility Trial period and size, representation of difficult cases, group comparability, instruction and model/source versions. Limits to generalisation. Assumed comparison of 12 task pairs of similar difficulty. No difficult cases or evidence that the effect persists. No actual model version or logs: this example illustrates documentation structure, not a completed experiment. ### Time, skills and support Planned trial time, task-based training, reviewer competence, support channel, reported concerns and manager actions. Illustratively, the manager reserves time to learn the instructions and review responses together. Before a real trial, reviewer competence and a route for raising concerns must be confirmed. ## Economics and results ### Cash expenditure and staff capacity Separate one-off and recurring costs, period and currency, licences, implementation, data and support. Distinguish staff time from cash. Count verification included in task time only once. Illustrative assumption: licence PLN 300/month excluding VAT; preparation uses 6 hours of existing staff time, not valued and not assumed to be additional expenditure. Other costs unestimated. Verification is already included in the 15 min end-to-end time. ### Results and evidence Before/after figures, units, numerator/denominator, spread, serious errors, evidence source and deviations from plan. Separate observations, assumptions and gaps. Assumed example results: 12 AI-assisted responses, mean 15 min from starting to acceptance, range 11–22 min; 2/12 required correction, 0/12 serious errors under the illustrative definition. Correction time is included in 15 min. Zero in a small example does not establish safety. ### How will the benefit be realised? Will time released increase capacity, improve timeliness or reduce expenditure? Name an owner and mechanism; do not equate hours with cash. Difference in means: 20 − 15 = 5 min per task (25% relative to 20 min). In this example, 12 × 5 = 60 min difference in total time. No monthly extrapolation. No demonstrated cash saving or basis for ROI. ## Decision and ongoing accountability ### Decision and rationale Expand / improve and retest / stop / insufficient evidence. Compare evidence with criteria; document accepted risks, conditions and scope. Improve the plan and retest. The illustrative result is promising but excludes difficult cases and full costs. It does not authorise expansion. Agree the environment, dates and evidence before a real test. ### Approval and next review Who decided and when? Who monitors outcomes? Review date or trigger, retest conditions and person responsible for stopping use. A file entry is not an electronic signature. Example with no actual approval or decision date. Review after a representative trial and following a significant tool, data or process change. The decision owner remains accountable for the scope of continued use. --- NIST AI RMF 1.0 (2023): https://airc.nist.gov/airmf-resources/airmf/5-sec-core/ Independent template; no certification or full compliance claim. Use internally and adapt. Retain attribution when redistributing the original.