What will you learn?
- Start with one measurable task, such as extracting electricity consumption, billing period and unit.
- Test against human-reviewed records and preserve original documents and change history.
- A model prediction or alert is not a field measurement or a verified emissions report.
1. Define the task and responsibility
Ask one question: can an assistant extract electricity consumption and billing dates while flagging uncertainty? Assign a data owner, reviewer and deployment decision-maker. Define stop conditions and decisions reserved for specialists. Use AI RMF to consider governance, context, measurement and risk treatment, not as a quality certificate.
2. Prepare traceable data
Use an authorised test set containing clear, difficult and incomplete invoices; remove unnecessary personal information. Link fields to documents, pages, periods and units. Record model version, instructions and sources. The system must not invent emission factors or fill missing values without evidence and review.
3. Compare with a human reference before scaling
Reserve test records not used to tune settings, and compare outputs with a specialist-reviewed reference. Check values, units, periods and missing-data alerts. Define accuracy by field type and sample size, and record errors. Correct number extraction does not establish correct classification or carbon calculation.
- Compare with a manual process or simple rules to establish the additional value.
- Test overlapping periods, changed units and duplicate invoices.
- Record reviewer decisions and reasons for corrections while retaining original values.
4. Control the move into operation
Document an accept, revise or reject decision, with a review owner and reassessment schedule. People retain approval of consequential data and reports. Retest after model or document changes. The first deliverable is a trial record, use boundaries and training plan, not proof of compliance.
Assumptions and scope
- The facility is authorised to use the test data and selects a processing environment appropriate to its confidentiality.
- A specialist-reviewed reference, acceptable-error definition and human escalation rules exist.
Limits of interpretation and use
- This educational material has not yet received independent technical review. NIST AI RMF is voluntary and does not certify this trial.
- Probabilistic outputs do not replace measurement, calibration or independent verification required for a report.
- This guide uses AI RMF 1.0 (2023). NIST states that a revision is in progress; check updates before implementation.
References & editions
- Artificial Intelligence Risk Management Framework (AI RMF 1.0) ↗
National Institute of Standards and Technology (NIST) · NIST AI 100-1, 2023-01-26; DOI: 10.6028/NIST.AI.100-1
Accessed: - AI RMF Playbook — Map ↗
NIST AI Resource Center · Companion to AI RMF 1.0; web guidance consulted 2026-11-01
Accessed: - AI RMF Playbook — Measure ↗
NIST AI Resource Center · Companion to AI RMF 1.0; web guidance consulted 2026-11-01
Accessed: - AI RMF Playbook — Manage ↗
NIST AI Resource Center · Companion to AI RMF 1.0; web guidance consulted 2026-11-01
Accessed:
These guides were prepared with AI assistance and the listed references. They support reading and learning; no approval or review by Dr. Sherif is claimed before it is documented.
Editorial & review policy ↗Suggest a correction