Agentic AI
AI systems that plan and execute multi-step tasks using tools, with approval gates and observability.
How a question reaches a governed answer
What it is
Understanding Agentic AI
An AI agent is a system where a model decides which actions to take, invokes tools or APIs to take them, observes the results and continues until a goal is met or a limit is reached. The difference from a chatbot is consequence: an agent changes state in real systems.
Why it matters
Automation of multi-step tasks that involve judgement and therefore resisted conventional rules-based automation. The value is genuine; the governance requirement is correspondingly higher because the system acts rather than advises.
Problems it addresses
- “Multi-step processes require judgement at each step and cannot be automated with rules.”
- “Research and information-gathering tasks consume professional time.”
- “Case handling requires reading context across several systems.”
Capabilities
Agent design and scoping
Narrow, well-defined tasks with explicit success criteria and stopping conditions.
Tool definition
Deliberately constrained capabilities with authorization checks at the tool boundary.
Orchestration and memory
State management across steps with bounded context.
Approval gates
Human authorization required before consequential actions.
Observability and evaluation
Complete execution traces and outcome measurement.
Use cases
Case triage and preparation
Gathering and summarizing context before a human decision.
Research assistance
Multi-source information gathering with cited output.
Operational runbooks
Diagnostic sequences with human approval before remediation.
Before you commit
Prerequisites, governance and risks
The questions that determine whether this will work in your environment, rather than in a demonstration.
What has to be true first
- Mature evaluation practice from prior generative AI work — agents are harder to evaluate, not easier
- Complete observability before any autonomous action is permitted
- Explicit authorization model at the tool boundary
Governance
Agents must be governed like any actor that can change enterprise state: a bounded set of tools, approval gates on consequential actions, permission-awareness inherited from the requesting user, complete execution logging, and a defined accountability point. An agent without an audit trail is not deployable in a regulated process.
Risks
- Compounding error across steps — a wrong early step propagates
- Unbounded action without hard limits on cost, iterations and scope
- Difficult evaluation — outcomes are path-dependent and non-deterministic
- Attribution of responsibility when an agent acts incorrectly
Where Agentic AI leads
Every connection below is recorded in the content itself, and each one says why it is here.
Capabilities
Technologies
Sectors
Explore solutions using Agentic AI
Open the guided finder with this technology already selected, and see which capabilities apply to your goal, sector and environment — with the reasoning shown.
Start from a goal