AI Pilot Protocol

How to use this document

• What it is for. This is a base protocol to design, run and evaluate AI pilots and proofs of concept before moving them into production: hypothesis, scope, data, controls, success criteria, stop criteria and the decision to go live, adjust or reject. Only cases already prioritized in the "AI Use Case Portfolio" enter a pilot.
• Who should own it. The AI lead, with the system owner of each pilot. Enter the supporting function in [RESPONSIBLE FUNCTION] and the person in [OWNER].
• Decisions to take first. Decide the maximum length of a pilot (Section 4.3), whether real data is allowed and with what authorization (Section 5), which controls are mandatory (Section 6) and confirm that who decides on go-live follows the lanes in the "AI Governance Operating Model" (Section 11.3).
• Fields to complete. Replace the bracketed fields and use the design record (aid B) and the evaluation report (aid C) for each pilot. Keep [OPTIONAL] clauses only if they apply.
• Country variants. This protocol has no country variants. If the pilot uses personal data or involves customers, workers or patients, check the notice and impact assessment obligations against the laws listed in "Related regulatory references."
• Review. Have privacy and security validate the rules on data and environments, and have legal review the notice to participants and the test terms with the vendor.
• Approval and publication. Have it approved by the AI lead (at the standard or reinforced level, [AI COMMITTEE]), record the approval in "Version history," and share it with the product, data and business teams.

AI Pilot Protocol

Field	Value
Organization	[LEGAL ENTITY NAME]
Owning function	[RESPONSIBLE FUNCTION]
Protocol owner	[OWNER]
Approved by	[APPROVED BY]
Effective date	[DATE]
Review frequency	[REVIEW FREQUENCY]
Contact	[AI GOVERNANCE EMAIL]

1. Purpose

1.1 This protocol sets out how [ORGANIZATION NAME] tests AI solutions under controlled conditions in order to decide, based on evidence, whether they deliver value with risk within the risk appetite (see "AI Risk Appetite Statement") before they are used at scale.

1.2 Its objectives are to limit exposure of people, data and processes during testing; measure results against criteria set in advance; detect quality, bias, security and adoption issues; and document the final decision.

2. Scope

2.1 It applies to any pilot, proof of concept, A/B test or limited rollout of an AI system, in-house or from a vendor, that uses the organization's data or interacts with real people, internal or external.

2.2 A pilot under this protocol is mandatory before go-live of:

• systems at the Medium or High risk level under the "AI Risk Classification Matrix";
• systems that make or support decisions about people;
• systems that interact with customers or the public;
• AI agents that take actions in internal or external systems.

2.3 Low-level systems may be piloted with the simplified design record (hypothesis, scope, success criteria and dates only). Exploratory tests in isolated environments, with fictitious or public data and no real users, do not require this protocol, but they are recorded in the "AI Use Case Portfolio."

2.4 The depth of the pilot depends on the risk level on the common scale:

Risk level	Pilot depth	Who approves the start and go-live
Low	Simplified record; no fairness report unless it affects people	System owner (green lane)
Medium	Full protocol; short impact assessment before starting	AI lead (yellow lane)
High	Full protocol with shadow mode or 100% human review, fairness analysis and a full impact assessment before starting	AI committee, or management at the basic level (red lane)
Unacceptable	Not piloted	—