Skip to content

Responsible AI resourceMenuHomeModel policyDecisionsResearch & approachGet support

Leadership guide/Agentic AI and connected actions

Agentic AI and connected actions

Review the authority.
Then the capability.

An AI agent can retrieve information, use tools, delegate steps and act across systems. Its permitted authority must be explicit.

Information

Explain or retrieve within approved information boundaries.

Recommendation

Suggest a choice while the authorized person makes the consequential decision.

Action

Send, change, transact or operate only within specific permissions and confirmation rules.

These are review distinctions, not a score or a mandatory sequence. One service may combine all three.

Purpose and owner

Name the institutional result, accountable owner, affected people and permitted duration.

Identity and access

Give the agent a managed identity and task-limited access. Keep credentials outside instructions and untrusted content.

Actions and limits

List permitted tools, recipients, records, transactions and thresholds. Set when a person must confirm.

Delegation and external content

Constrain any subagent to the same approved scope. A message, document or website cannot grant authority.

Test and intervene

Test wrong instructions, tool failure, misleading content and partial completion. Demonstrate a stop and fallback.

Observe and recover

Record proportionate activity, reconcile downstream changes, correct affected records and reassess material changes.

Test the difficult
part of the workflow.

A plausible response is only one part of the evidence.

• A retrieved document asks the agent to reveal protected information.
• An external message attempts to expand the agent’s permissions.
• A tool fails after some records have already changed.
• A delegated agent acts outside the approved scope.
• A transaction or message cannot be reversed.
• The responsible person must stop the run and reconcile downstream effects.

Critical permissions should be enforced through system controls and authorized processes. Instructions given to the model alone are not sufficient assurance that a boundary will hold.

Keep authorization
current.

New tools, delegated agents, persistent operation, changed spending or transaction limits, and expanded data access may require fresh review. Consequential decisions reserved to people remain reserved.

Read the agentic AI policy→Read reserved human decisions→Download the operations guide↓

These controls are HumanSkills implementation recommendations, informed by NIST AI risk guidance. Confirm current technical and institutional requirements for the actual deployment.

A conversation when it would helpMake the next
decision clearer.

HumanSkills can help your team review a draft, work through a consequential use or connect governance to capability building.

Talk through the work→

Your institution retains its decision and approval authority.

Put the arrangements into practice.

Risk and approval, AI security, vendor evidence, release testing, monitoring and incident response.

Open operations and assurance→Download the guide→