How to use this fieldbook
This fieldbook is designed for teams moving from an AI-agent pilot toward dependable business use. It provides six worksheets that connect an agent's business purpose to its scope of authority, evidence, intervention controls, and change approvals. Work through the worksheets for one representative workflow rather than attempting to certify every agent on the first day. Each worksheet identifies a purpose, fields to capture, and an acceptance test that a different reviewer can reproduce.
The worksheets are original editorial tools, not a proprietary platform feature, a standards certification, or a declaration of regulatory compliance. The purpose is to help decision-makers ask better operational questions and preserve the results. NIST's AI Agent Standards Initiative illustrates that secure identities and interoperability are subjects of active standards work in 2026 [1]. It does not by itself create a universal legal obligation for agent deployment.
Worksheet 1: Define the business job before the technology
Purpose: Establish the concrete task the agent is intended to perform and the business owner who accepts its consequences. Avoid vague ambitions such as “automate operations” or “increase productivity.” Identify the specific process, input, authorized output, affected parties, and business system of record. A draft recommendation and an executed change are separate tasks, even when the same model can perform both.
Fields to complete: Agent ID ____; business outcome ____; process owner ____; approved environment ____; current lifecycle stage ____; users served ____; systems touched ____; excluded tasks ____; review date ____. The source of each entry should be an accountable team record, not a guess based on a product demonstration.
Acceptance test: Give two reviewers the brief and ask them whether an example action belongs to the approved task. Both should reach the same decision for an ordinary case and a deliberately ambiguous case. Record disagreements and refine the scope before adding new permissions. The NIST AI RMF Map function describes documenting intended purpose and context as part of risk management [2].
Worksheet 2: Write an action-permission agreement
Purpose: Separate what an agent can technically perform from what the organization permits. Document each connected application, data category, identity, credential scope, allowed action, approval condition, and prohibited action. Write the rule in language the business sponsor understands, then confirm that the technical system can enforce it. Do not treat a natural-language instruction as proof that a database write or external message will be denied.
Fields to complete: Tool ____; connection owner ____; identity ____; read permission ____; write permission ____; external recipient limits ____; approval threshold ____; expiration ____; suspension method ____; owner sign-off ____. Where an agent acts on behalf of a user, define which authority is inherited, delegated, or deliberately restricted.
Acceptance test: In a controlled environment, attempt an in-scope read and an out-of-scope write. Preserve the actual tool response, not merely the agent's explanation of why it declined. Microsoft documents dedicated agent identity constructs and access-governance concepts in Entra [3]. Such features are product-specific; the worksheet remains applicable without them.
Worksheet 3: Describe the business rules behind decisions
Purpose: Prevent an agent from applying technically accessible data to the wrong business question. Identify the authoritative definition of a customer, policy, supplier, account, threshold, or contractual condition. If different systems disagree, specify which one prevails and who may resolve the conflict. An agent that acts quickly on ambiguous context can create an error that is difficult to explain later.
Fields to complete: Rule name ____; authoritative source ____; approved version ____; effective date ____; business meaning ____; exception ____; conflict owner ____; evidence location ____; example valid decision ____; example invalid decision ____. Do not assign an evidence label to an invented record.
Acceptance test: Present the same fictional transaction with one important field changed. Check whether the approved business rule changes the allowed outcome. Keep the test record and expected outcome together. ISO/IEC 42001 describes governance through policies, objectives, and management processes for AI [4]. The worksheet is a possible operational application of that principle, not proof of conformity with the standard.
Worksheet 4: Build the event and accountability trail
Purpose: Allow authorized reviewers to reconstruct material actions. Define the events needed to demonstrate the request, data source, selected tool, permission decision, execution result, human approval, and any reversal. The detail should reflect business consequence and applicable privacy requirements. A system that stores only a final answer may not retain enough information to resolve a disputed business action.
Fields to complete: Event source ____; event type ____; agent run ID ____; action time ____; related business record ____; decision or approval ID ____; error outcome ____; retention class ____; authorized reviewer ____; exception route ____. If separate systems hold the evidence, specify how the references can be joined without inventing a complete history.
Acceptance test: Ask a reviewer to locate one permitted action and one denied action from the recorded source. The reviewer should be able to distinguish an observed event from an inferred intention. NIST SP 800-53 includes access-control, audit, and incident-response families that provide useful control vocabulary [5]. Applying similar concepts does not mean an agent has been assessed against a particular federal baseline.
Worksheet 5: Set human intervention and recovery thresholds
Purpose: Make human accountability practical at the point of consequence. Define which actions may run unattended, which require review before execution, and which need immediate interruption or escalation. Include the real destination of alerts and the authority to restart a suspended workflow. If a completed action cannot be reversed, record the compensation or notification procedure.
Fields to complete: Consequence category ____; approval point ____; reviewer ____; escalation trigger ____; stop control ____; stop-test date ____; rollback or correction method ____; affected-party communication ____; incident owner ____. Use a realistic scenario but avoid testing destructive actions on production data.
Acceptance test: Simulate an unacceptable proposed action and verify the approval path. Then simulate an already completed but incorrect change and verify the recovery responsibility. The EU AI Act's Article 14 requires certain oversight capabilities for covered high-risk AI systems [6]. Its legal scope is specific; a broader internal control can be a sound design choice without being a universal statutory requirement.
Worksheet 6: Establish the change and legal-applicability record
Purpose: Preserve the authority granted at a particular time and determine what would invalidate that approval. A changed tool, model, dataset, business rule, integration, or deployment geography may alter the workflow's risk. Record a change proposal, a risk comparison, the approver, and the effective date. Separately identify any applicable legal obligations through a qualified assessment of the actual use.
Fields to complete: Previous version ____; proposed change ____; newly accessible systems ____; changed data category ____; impact ____; jurisdiction ____; system role ____; applicability source ____; reviewer ____; approved effective date ____. Avoid claiming that all agentic systems meet a specific AI Act category.
Acceptance test: Have a different reviewer explain which changes require new business approval and which require additional Security or Legal review. The European Commission's AI Act implementation FAQ distinguishes applicability and enforcement milestones by category [7]. It is a starting point for jurisdiction-specific legal review, not a substitute for it.
Worked example: A fictional supplier research agent
A procurement team pilots an agent that compares approved supplier documents. It can read a controlled document library, generate a summary, and create a draft comparison. It cannot contact suppliers, modify payment details, accept bids, or issue purchase orders. A Procurement Operations Lead sponsors it. A platform administrator controls the agent's account and can suspend access. Each draft has a document reference and review record.
A project sponsor proposes allowing the agent to send clarification messages to suppliers. The original approval does not cover this new action. The team therefore updates the scope, tests the permission boundary, reviews recipient restrictions, adds a human approval point, and records the new effective date. The example is fictional and does not represent a verified implementation or customer result.
A practical 30-day implementation sequence
Days 1–5: Select one representative agent, assign the business sponsor, and complete Worksheet 1. Output: Versioned use-case record with boundaries. Days 6–10: Map tool access and run limited permission tests. Output: Authorized-action matrix. Days 11–15: Document business rules and trace an action to the source system. Output: Business-rule register and event references.
Days 16–20: Test intervention and correction in a controlled environment. Output: Approval, suspension, and recovery results. Days 21–25: Conduct a cross-functional review of one ambiguous case and one prohibited action. Output: Dated decision record. Days 26–30: Reassess one proposed scope change and document the next review date. Output: Revised operating specification. The sequence is illustrative and does not assert that a customer deployment has been completed.
A transparent readiness rubric
For each worksheet, use four states: Not described; documented; evidence-tested; and used in an approved workflow. Report the number of evidence-tested worksheets out of six, with the review date and sample scope. Do not describe documentation alone as proof of implementation. Also report whether each legal applicability conclusion has been evaluated for the specific workflow. A single total score should never erase a failed high-consequence control.
Reviewers can add an independent evidence note: What was observed; who tested it; where the record is held; and which limitations apply. If the agent changes tools or business purpose, reopen affected worksheets rather than preserving a green status automatically. This discipline makes the fieldbook maintainable rather than a one-time exercise.
The Intent Amplify perspective
Enterprise technology vendors can use a fieldbook to help buyers clarify a business problem before comparing software features. Intent Amplify Buyer Intelligence supports defining a relevant account and role universe [8]. Managed Demand Generation describes coordinating approved content, audience criteria, and validated responses [9]. A worksheet request should remain an educational interaction until separately supported evidence justifies another qualification status. Neither service implies that any particular company is noncompliant.
Pilot the fieldbook with one business team
Complete the six worksheets for a small, authorized workflow and share the result with its sponsor, Security, and Risk teams. Use their feedback to decide whether a broader governance discussion is useful. For a guided approach to reaching relevant enterprise-technology audiences with practical resources, see Intent Amplify Managed Demand Generation. https://www.intentamplify.com/solutions/managed-demand-generation.
Evidence notes and references
Bracketed citations identify the public source supporting adjoining factual claims. Frameworks, example scenarios, worksheets, and recommendations are editorial synthesis, not measured customer results. Standards and legal obligations apply only within their stated scope. Service delivery depends on approved program terms. Sources reviewed October 29, 2026.
- NIST, Announcing the AI Agent Standards Initiative (February 17, 2026). https://www.nist.gov/news-events/news/2026/02/announcing-ai-agent-standards-initiative-interoperable-and-secure
- NIST AI RMF Playbook, Map (Official voluntary guidance). https://airc.nist.gov/airmf-resources/playbook/map/
- Microsoft Learn, Governing Agent Identities (Official platform documentation). https://learn.microsoft.com/en-us/entra/id-governance/agent-id-governance-overview
- ISO, ISO/IEC 42001:2023 (Official standard description). https://www.iso.org/standard/42001
- NIST, SP 800-53 Rev. 5: Security and Privacy Controls. https://csrc.nist.gov/pubs/sp/800/53/r5/final
- European Commission, AI Act Article 14: Human Oversight. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14
- European Commission, Navigating the AI Act (Current implementation FAQ). https://digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act
- Intent Amplify, Buyer Intelligence. https://www.intentamplify.com/solutions/buyer-intelligence
- Intent Amplify, Managed Demand Generation. https://www.intentamplify.com/solutions/managed-demand-generation
