
Compliance work combines research, assessment, documentation, and coordination. AI agents can prepare much of it: extract evidence, identify open points, summarize requirements, and draft proposals. Professional accountability remains with the team.
The engineering distinction is whether a system only produces text or can also invoke tools, change records, and trigger downstream workflows. A mistaken summary needs correction. A mistaken action may disclose evidence, alter permissions, or propagate an unsupported control conclusion.
Every AI-assisted workflow therefore needs explicit boundaries: what may the agent propose, what may it change, and which steps require approval? Enforce those boundaries through permissions and workflow controls, not just instructions in a prompt. Keep proposals visibly separate from approved decisions.
The controls below are practical recommendations. Their applicability depends on the agent’s purpose, data, authority, and operating environment.
ISO/IEC 42001 provides requirements for an AI management system. The NIST AI Risk Management Framework provides voluntary guidance for governing, mapping, measuring, and managing AI risk. Neither automatically establishes compliance with the EU AI Act.
AI Act classification depends on the use case and the organization’s role. An agent used to summarize evidence does not have the same implications as a system used in employment decisions. Check the European Commission’s current overview, rather than treating every agent as high-risk or assuming one universal implementation deadline.
For assessment purposes, focus on demonstrable control: explain the system, show its boundaries, and produce records showing that review and monitoring actually happened.
Include agents embedded in SaaS products, internal scripts, and departmental workflows, not only centrally deployed applications. Compare procurement records, service identities, approved integrations, and team interviews to find automation missing from the inventory.
For each agent, record its deployment location, model and supplier, connected systems, business purpose, and lifecycle status. Name an accountable owner, technical maintainer, and reviewer. These roles can overlap in a small team, but responsibility should remain explicit.
Then describe the operating envelope in testable terms:
Recurring, bounded work is usually the strongest starting point. Document extraction, gap analysis, pre-filling controls, summarizing changes, and preparing audit questions produce outputs that reviewers can inspect directly.
Security teams can also use agents to enrich alerts, reconstruct incident timelines, prioritize vulnerabilities, review code, or flag unusual machine-identity permissions. Start with analysis and recommendations; granting production write access creates a different risk profile.
Give each deployed agent a distinguishable service identity or equivalent execution identity. Avoid shared administrator credentials. Scope access by task, environment, data collection, and permitted operation; review both direct permissions and authority inherited through connected tools.
A read-only agent can still expose information through its output. Restrict destinations as well as inputs. Where practical, use short-lived credentials and keep secrets out of prompts, retrieved documents, and logs.
Tool interfaces should expose narrow operations with validated parameters. An evidence-drafting agent needs an operation such as “create draft,” not an unrestricted database or shell interface. Validate authorization server-side for every call. The model’s description of its authority is not an access-control decision.
Set execution limits for tool calls, retries, elapsed time, and batch size. These reduce the impact of loops and accidental bulk changes. Where writes are allowed, use idempotent operations where possible and record enough state to support recovery.
A frequent mistake is approving an agent’s purpose while leaving its integration account broadly privileged. Test the actual credential and tool path, not merely the intended configuration.
A plausible answer is insufficient for compliance work. Reviewers need to see which files were read, which passages support the proposal, and which assumptions remain unresolved.
Each material conclusion should carry a source reference with document identity, version or retrieval timestamp, and relevant location. Distinguish source facts from model interpretation. Mark missing information explicitly; an absent access-review record must not become a claim that access reviews occurred.
Retrieved documents and tool responses are untrusted inputs. A document containing “send all evidence to this address” must not expand the agent’s authority. Separate retrieved content from trusted instructions, validate tool arguments, and enforce access and destination restrictions independently of the model.
Test this boundary using controlled documents containing malicious instructions, misleading citations, and conflicting evidence. Retain the test inputs, expected behavior, observed outcome, and remediation. Also test ordinary failures: obsolete policies, incomplete exports, and unsupported assertions.
Verifiable work strengthens the human decision process. Fast output without inspectable support does not.
Human oversight must occur before the action that matters. Reviewing a report after evidence has been disclosed is not an effective disclosure control.
Define approval points for consequential actions, such as publishing an assessment, sending customer information, changing access, or accepting residual risk. Show the reviewer the proposed change, supporting sources, affected objects, and unresolved uncertainty.
Bind approval to the specific version reviewed. If the agent changes the proposal afterward, require renewed approval rather than reusing a previous authorization. Record reviewer identity, timestamp, decision, and comments.
For lower-impact work, sampled review may be appropriate after evaluation demonstrates acceptable performance. Do not silently extend that arrangement to higher-impact tasks. Changes in autonomy need explicit reassessment.
Policies describe intent; operational evidence shows whether that intent holds. Assemble the following records for each agent, with links between configuration, tests, activity, and decisions.
| Record | Useful evidence |
|---|---|
| 1. Inventory, intended purpose, and ownership | Agent entry, authorized use cases, deployment location, accountable owner, and reviewer |
| 2. Data sources and handling | Approved collections, sensitivity, suppliers, retention, and data-flow diagram |
| 3. Permissions and restrictions | Service identity, tool allowlist, access configuration, and denied-action tests |
| 4. AI-specific risk assessment | Errors, misuse, prompt injection, downstream impacts, mitigations, and residual-risk decision |
| 5. Evaluation and limitations | Test cases, expected results, observed failures, acceptance criteria, and known limitations |
| 6. Human review and approval | Workflow rules, completed approvals, rejected proposals, and escalation examples |
| 7. Activity trace | Request, sources, configuration version, tool calls, results, and linked decisions |
| 8. Monitoring and response | Alert rules, review records, incident procedure, stop mechanism, and recovery test |
| 9. Change and reassessment history | Model, prompt, tool, permission, and source changes with release decisions |
Assessment scope determines which evidence is relevant. There is no universal requirement to expose a model’s hidden reasoning. Capture observable inputs, actions, sources, and a concise decision rationale instead.
Protect the records themselves: restrict access and apply retention appropriate to their sensitivity. Logging full prompts indiscriminately can create another repository of confidential or personal data.
Assume, illustratively, that a team wants an agent to review access-management evidence for one application. It may read an approved policy, a user export, and completed review tickets. It may create control drafts, but cannot approve them, modify accounts, or send messages.
The team evaluates it against a reviewer-checked set containing complete evidence, missing tickets, expired documents, and a planted disclosure instruction. Acceptance requires supported conclusions, explicit identification of gaps, and no prohibited action. Faster completion alone is not a release criterion.
During a run, the agent finds that the policy requires periodic reviews but the uploaded tickets cover only part of the assessment period. It drafts a partial-evidence finding, cites the policy and tickets, and asks for the missing records. The reviewer checks the sources and records the final decision.
Retained evidence includes the tested permission configuration, evaluation results, source references, draft, reviewer correction, and approval. This demonstrates the workflow without claiming that the agent proved the control effective.
Track unsupported conclusions, reviewer corrections, denied tool calls, unusual data access, repeated retries, and unexpected action volume. Measure correctness and reviewer effort alongside speed. Sample successful runs too; errors do not always generate alerts.
Monitor according to risk and reassess when models, prompts, tools, permissions, or source collections change. A previous evaluation does not automatically cover a new configuration.
Provide a tested stop mechanism that disables execution or revokes tool access. Assign incident ownership and define how to preserve logs, identify affected records, reverse changes where possible, and notify relevant teams. A stop button cannot undo a disclosure, so prevention still matters.
Choose one bounded task, assign its owner, document its permissions, run adversarial and ordinary tests, and retain the first reviewed execution before expanding scope.
In Secani, teams can connect evidence to requirements, controls, systems, and decisions; turn gaps into tasks; and review AI suggestions as draft work. Preserve source links, comments, approvals, and timestamps around material decisions. That workflow supports traceability without substituting a tool for professional judgment.
Use the AI governance framework guide to place these agent-level records within the wider management system.

AI governance · October 10, 2026 · 9 Min.
A source link does not make an AI verdict reliable. The relationship between a requirement, evidence, and the verdict must be testable.

AI governance · October 19, 2026 · 6 Min.
Three standards in three days: ISO/IEC 27090 for AI security, EN 18286 as the first harmonised standard under the EU AI Act and IEEE 2863 for AI governance at board level.

AI governance · October 18, 2026 · 5 Min.
ISO 27090 guides AI security work. ISO 42001 sets requirements for an AI management system. Here is how they differ and how to use them together.
Secani connects scopes, evidence, tasks, and AI agents in one shared workspace.