How to choose an AI pilot with measurable value

Start with a repeated task, its cost and quality requirements. Compare AI with the current process on equivalent examples. Include review time, correction costs and data constraints. Speed without acceptable quality is not value.

RA LAB practical guide · published 16 October 2026 · based on the library’s GRI / GTM sources.

GOAL

Choose one process and test AI value against the current way of working.

VALUABLE FINAL DELIVERABLE

An AI pilot brief: baseline, quality metric, cost, risks, human oversight and decision.

From problem to decision

1. Measure the current process

Choose a repeated task and comparable examples. Measure current time and errors. Do not enter client or sensitive data in this worksheet.

Task, frequency, time, quality and baseline evidence.

2. Set guardrails

Decide which data is allowed, who reviews outputs and how to return to the manual process. Track errors of different severity separately.

Allowed data, risks, human review and stopping rule.

3. Design the comparison

Compare current and AI processes on comparable tasks. Include tool cost and review time. Set the quality threshold before the test.

Task set, metric, quality threshold, duration and owner.

4. Evaluate actual value

Record results, errors and limitations. Scale only after checking quality, economics and team adoption.

Time before/after, errors, total cost and decision.

How to judge whether it helped

You should leave with a document that supports a decision and a next action. Completing steps does not guarantee a commercial result. Measure the chosen change separately over an agreed period. If evidence is missing, state that and plan how to collect it.

How this differs from S01–S13

This is a user route for a specific problem. The GRI AI research stages S01–S13 and the 13-week programme remain in the library as separate methodological work. This route is not a validated GRI diagnostic scale.

Your route0 / 4 · 0%

Completed 0 · remaining 4

Next action: Measure the current processThe percentage measures completed steps, not proven business growth.

Measure the current process

Choose a repeated task and comparable examples. Measure current time and errors. Do not enter client or sensitive data in this worksheet.

No names, contacts, client secrets or sensitive details. Saving happens only when you press the button.

Discuss this problem ↗

Discuss opens WhatsApp. Your answers are not sent automatically.

Data & privacy

Matching works without saving your query. Practice answers remain in this tab until you enable saving. Do not enter names, contacts, client secrets or sensitive information.

Server saving: answers and steps are linked to a random key in a secure browser cookie for up to 90 days. This is pseudonymous, not anonymous data. It does not sync across devices.

Server saving is not enabled yet: the data controller and hosting configuration are being confirmed. You can work in this tab and download your results.

How data is used and how to request deletion ↗

Sources and supporting materials

Methodological references. Links lead to original materials; worksheet examples are not research findings.

AI Risk Management Framework ↗NIST · A use-case-aware operating model organized around Govern, Map, Measure and Manage.Guidelines for Human–AI Interaction ↗Amershi et al. / Microsoft Research · 18 interaction guidelines covering expectations, explanations, correction, feedback and control.Trustworthy Online Controlled Experiments ↗Kohavi, Tang & Xu · A rigorous foundation for experiment design, metrics, statistics and organizational culture.Identifying and Scaling AI Use Cases ↗OpenAI · A field guide for finding, prioritizing and expanding AI use cases from tasks into end-to-end workflows.Open the full source catalog →