Definition
Agentic AI describes systems given a goal rather than a single instruction, which then handle the chain of steps required to reach it: read the current state, form a plan, call tools to execute, inspect what came back, and decide the next step. "Agentic" describes the architecture and operating model, not the model family — you can build an agentic system on the same model you already use for ordinary generation.
Versus generative AI
| Dimension | Generative AI | Agentic AI |
|---|---|---|
| Input | A prompt or question | A goal or task |
| Output | Text, image, code | A system change plus a summary |
| Steps | Usually one | Several, chained |
| Tools | Usually none | Essential |
| Evaluation | Done by the human | Done by the system, then a human at the boundary |
| Failure | An unsuitable output | A retry or a logged escalation |
A concrete contrast: ask a generative model to draft a reply to a customer complaint and you get good text you paste yourself. An agentic system reads the ticket, extracts the order number, checks shipping status in the system, writes the reply, sends it, and updates the ticket — and if the complaint requires financial compensation, it stops and waits for approval.
The four building blocks
- The model: understanding, phrasing, and proposing the next step.
- Knowledge: company documents, policies and data, retrievable at the moment of need.
- Tools: real interfaces into systems (read/write) with an explicit scope per tool.
- Permission policy: what runs automatically, what requires approval, what is forbidden outright.
The operating loop
The strength of the agentic pattern is disciplined iteration. The system does not assume a step succeeded; it checks the result. Did the external system return an error? Is data missing? It then retries with different inputs or escalates to a human instead of continuing on a false premise. Retry counts and stop conditions must be defined explicitly, otherwise the loop becomes an endless spin.
What makes it enterprise-ready
- Clear permission scope per tool, following least privilege.
- Human approval points on every irreversible or financially material action.
- An audit log of each tool call and its result, reviewable after the fact.
- Current, company-owned knowledge with a single source of truth.
- Measurement: completion without intervention, escalation rate, and failure reasons.
Common misconceptions
- "Agentic means no humans" — no; it means splitting decisions between system and human along declared boundaries.
- "A stronger model is enough" — a stronger model does not compensate for missing tools, knowledge or permissions.
- "Agentic fixes undefined processes" — a vague process yields vague agent outcomes; a clear definition comes first.