Choose and implement an AI solution

First describe what should change in the work. Compare options, test them on a specific task and use the result to decide the scale of deployment.

1. Brief, options and budget

Define the required outcome, data, integrations and constraints. Include improvements to the existing process without adding an AI tool.

• Options, benefits and costs — a decision brief and comparison worksheet.
• AI solution architecture — a ready-made tool, integration or custom development.
• Tools for AI records and governance — requirements, licences and feature verification.

2. Pilot and decision to continue

Agree test tasks, acceptance criteria, costs and conditions for stopping the pilot. Record errors and the effort required to check outputs.

• Measurement and evaluation — quality, reliability and limitations.
• Generative AI risks — what to include in tests and operating rules.
• AI Verify — uses and limits of technical tests.

3. Handover and operation

Before wider use, address access, integrations, responsibilities, training, support and failures. Base acceptance on met criteria and known limitations.

• Operating AI systems — monitoring, changes and traceability.
• Preparing for review — evidence for internal checks or independent assessment.
• Managing risks in operation — measures, residual risks and changes.

A more detailed method

The voluntary NIST AI RMF provides a framework for risk management. Its functions complement each other and are revisited throughout the work. They are neither certification nor mandatory purchasing stages.

Introduction to NIST AI RMF · GOVERN — responsibilities · MAP — context · MEASURE — evaluation · MANAGE — measures

For help with a brief, supplier comparison or acceptance, see AI project support.