ARKHOS | BOARD DECISION BRIEFS
Are data and operations ready for live use?
Decision requested: Fund a shadow test alongside the existing workflow, with no influence on operational decisions. Do not approve live use until management shows that the data is fit for purpose, outputs remain useful after review, and staff can restore the existing workflow if the tool fails.
Illustrative example: A Saudi distributor wants AI to forecast spare-parts demand. Rather than connect forecasts directly to purchase orders, the planning team compares the tool's recommendations with its existing process at one warehouse. Staff retain approval authority. This is not an actual result or a promise of savings.
Three decision gates
Data | Does it represent what we will encounter?
Identify the sources of item, inventory and sales data, and name an owner for corrections. Test the impact of missing records and changing item identifiers. Do not interpret low sales during stockouts as low demand. Test on a later period excluded from training and report results for critical items separately from the average.
Operations | Who acts when it fails?
Measure missed stockout warnings, suggestions that could lead to excess inventory and time spent on review. Name who reviews and can override recommendations, and rehearse returning to the existing workflow when the data feed fails.
Go live | What evidence is enough?
Set thresholds for errors, cost and review workload before testing. Average improvement is not sufficient if critical items fail. If thresholds and the fallback drill pass, consider limited use with human review. Otherwise fix the gap or stop the test.
Board question: Which test result would change our decision, and who can suspend use if the criterion is not met?
Primary sources
NIST AI RMF 1.0, January 2023: Map 2.3, Measure 2.3–2.5, Manage 2.4. Voluntary guidance. The inventory scenario and decision thresholds are editorial examples, not NIST requirements.