Readiness depends on the decision
A warehouse can hold millions of records and still be unable to answer whether an order can ship tomorrow. Stock may include quality holds; an inbound date may be a supplier estimate; the promised delivery date may have been overwritten. Volume alone does not make those records suitable for a decision.
NIST’s AI RMF includes attention to data availability, representativeness and suitability within evaluation. Its trustworthiness guidance also links reliability to intended operating conditions and continuing monitoring. [1] [2]
Begin with a decision-level data contract
SCP would map the minimum information needed for one use case and agree what each field means. This is a proposed working approach, rather than a universal readiness score.
| Information | Question to resolve |
|---|---|
| Customer promise | Is the original commitment retained, with agreed time zone and calendar? |
| Usable stock | Are quality holds, allocations and units of measure handled consistently? |
| Inbound supply | What does the date represent, and when was it last confirmed? |
| Order and item IDs | Can records be joined without duplicates or accidental aggregation? |
| Events and ownership | Who corrects missing, conflicting or delayed information? |
Test the awkward records
A clean demonstration file is not enough. Include duplicates, missing values, unit conversions, partial receipts, cancelled orders and events arriving out of sequence. Agree how the workflow should respond when confidence is insufficient.
Illustrative example: an inbound receipt with no confirmed date should remain an explicit information gap. Treating a blank date as immediate availability would turn missing evidence into an optimistic promise. A stale timestamp should trigger validation rather than silently refresh itself.
Use checks that lead to action
Measure completeness and freshness for the fields that affect the decision, not just a single average across a database. Report failures by owner, source and affected order population. Sample apparently valid records against operating reality.
Client-specific thresholds belong in acceptance criteria. Tests should verify both valid records and failure handling; UAT should establish whether users can recognise a problem and follow the agreed correction route. A passed field check alone does not establish that an AI recommendation is useful.
Improve a bounded flow first
Establish source authority, correction ownership and a monitoring routine for the pilot boundary. Expand only when the team understands the limitations and the additional cost of supporting more products, sites or systems.
A Control Tower discovery workshop can identify the minimum connected records and priority gaps. The result should be a usable data plan and decision boundary, with any necessary specialist system work scoped separately.
Sources and evidence notes
Sources checked on 17 October 2026. Research findings and company reports are attributed above; practical recommendations are SCP’s interpretation. The examples do not imply endorsement, partnership or SCP client results.
- NIST: AI Risk Management Framework Core ↗
AI RMF 1.0, 2023; online guidance reviewed 17 October 2026. Voluntary public-sector risk framework; a revision is in progress, and the framework is not a certification or a forecast of commercial benefits.
- NIST: AI risks and trustworthiness ↗
AI RMF 1.0, 2023; reviewed 17 October 2026. Public-sector framework guidance on validation and reliability; not a guarantee of system performance.
Explore the industries and services
Explore the next step
Discuss a defined supply-chain challenge and agree the evidence needed for a useful decision.
Workshop format and any fees are agreed before booking. Follow-on work requires a separate scope and business agreement.