Agentic describes an AI system that pursues a goal by choosing its own next steps: it plans, calls tools or APIs, checks what happened, and adapts until the task is done or it hands over to a person. Agentic AI is software built this way, as opposed to a model that answers one prompt at a time and leaves every action to a human.
The word marks a shift in where decisions sit. A chatbot decides what to say. An agentic system also decides what to do next, inside limits a company sets: which systems it may read, which actions it may take, and when it must stop and ask.1
One request, handled two ways
A customer writes to a retailer's support chat. The model behind both systems below could be the same. The difference is what the system is allowed and designed to do with it.
Chatbot
Answers- RetrieveFinds the help-center article on delivery delays.
- ReplyExplains that parcels can take up to 10 working days and links the refund request form.
- WaitThe customer fills in the form. A support agent checks the carrier and issues the refund one or two days later.
Answered, not resolved. Work moves to a person and a second contact is likely.
Agentic agent
Resolves- PlanGoal: resolve a missing delivery. Needs the order, its tracking status, and the refund policy.
- Toolorders.get("48213")→ paid €84.90 · shipped 11 Nov
- Toolcarrier.track(shipment_id)→ no scan for 6 days · status: lost
- CheckPolicy: refund a lost parcel automatically up to €150. Above that, a person approves.
- Actpayments.refund("48213", 84.90)→ refund confirmed
- ReplyConfirms the refund, gives the 3 to 5 day timeline, and offers to reorder with express delivery.
- LogCloses the ticket with a summary and tags the carrier issue for the logistics team.
Resolved in one conversation. A person sees only refunds above the limit.
Illustrative example. Tool names, amounts, and the €150 threshold are placeholders a company would set for itself.
What makes a system agentic
Classic AI textbooks define an agent as anything that perceives its environment and acts on it.3 Language models added a practical way to build one: the model reasons about the task, chooses an action such as an API call, reads the result, and reasons again. Research on this pattern, known as ReAct, showed that interleaving reasoning and actions beats either one alone on multi-step tasks.2
A goal, not a prompt
The system receives an outcome to reach ("resolve this delivery problem"), and the number of steps is not fixed in advance.
Planning
The model breaks the goal into steps and revises the plan when a result surprises it.
Tool use
It can call functions: read an order, search a knowledge base, update a CRM, issue a refund. Models can learn when and how to call such tools.4
State and memory
It keeps track of what it has done and learned within the task, and sometimes across sessions, so each step builds on the last.
Agentic is a spectrum
"Agentic" is a property a system has more or less of, so it is more useful to ask how much a system decides on its own than whether it is agentic at all. The scale below is a working model, not an industry standard.
Scripted bot
Next step decided by: the flow designerDecision trees and keyword matching. "Press 1 for billing."
Answering assistant
Next step decided by: the userA language model answers from a knowledge base. Every action stays with a person.
Tool-using assistant
Next step decided by: the user, per requestCalls one tool when asked, such as checking an order status, then stops.
Agentic workflow
Next step decided by: a predefined pathModels handle steps inside a fixed sequence: classify, retrieve, draft, check, send.
AgentExample above
Next step decided by: the model, within permissionsChooses its own steps toward a goal, with limits on what it may touch and when to escalate.
Multi-agent system
Next step decided by: an orchestrating agentSpecialized agents for triage, returns, and billing hand work to each other.
Agentic AI, generative AI, and AI agents
The three terms overlap and are often used interchangeably. They describe different things.
| Term | What it names | Typical output | In a support team |
|---|---|---|---|
| Generative AI | Models that produce new content | Text, images, code | Drafts a reply for a human to send |
| AI agent | A specific system that acts on a goal | Completed tasks, changed records | The refund agent in the example above |
| Agentic AI | The approach, and the category of systems built with it | Outcomes, with a log of actions | A strategy to let agents resolve contacts, not only answer them |
Where companies use agentic AI
The pattern works best where a task has many steps, reads several systems, and follows a clear policy.
Resolve, not deflect
Order status, returns, rebooking, billing changes. The agent completes the change the customer asked for.
read order → check policy → act → confirm → logReach out before the complaint
A delayed shipment or failed payment triggers the agent to message the customer with a fix already in hand.
event → assess impact → offer remedy → notifyAdvise and build the basket
Asks about needs, searches the catalog and stock, compares options, and adds the chosen items to the cart.
clarify → search → compare → add to cartProcess documents end to end
Reads invoices, matches them to orders and receipts, flags mismatches, and posts clean ones to the ERP.
extract → match → post or escalateAnswer employees with sources
Searches policies, contracts, and tickets, answers with citations, and opens a request when a person must act.
search → cite → open requestQualify inbound leads
Asks qualifying questions, enriches the record, scores fit, and books a meeting with the right rep.
qualify → enrich → book → update CRMWhere agentic systems fail, and the usual controls
An agent that can act can also act wrongly. The risks are manageable, but they are design work, not settings.
Errors compound
ControlA wrong first step carries through every later step. Test on real transcripts before launch and keep scoring live conversations afterwards.
Prompt injection
ControlText in an email, web page, or document can try to redirect the agent.5 Treat all retrieved content as untrusted and give each tool the narrowest permission it needs.
Irreversible actions
ControlRefunds, cancellations, and data changes need thresholds. Above them, a person approves before the agent acts.
Cost and latency
ControlEach loop is another model call. Cap the number of steps and route simple requests to a cheaper path.
Accountability
ControlLog every tool call and decision. Under the EU AI Act, people must be told when they are interacting with an AI system.6
Where the word comes from
"Agentic" came to AI from psychology. Stanley Milgram used "agentic state" in 1974 for a person who stops acting on their own judgment and becomes an instrument of someone else's will.7 Albert Bandura later used "agentic" in the opposite sense: people as agents who act with intention, forethought, and self-regulation.9
The AI industry took the second meaning, and the term became common in 2024 as model providers shipped tool use and agent frameworks. Milgram's sense still describes a well-run agent: it acts on a principal's intent, within that principal's rules.
Common questions
Is ChatGPT agentic?
In a plain chat it answers and you act. When it browses, runs code, or operates a browser on your behalf, it behaves agentically. "Agentic" describes how a system is set up to work, not a product.
Is agentic AI the same as AGI?
No. Artificial general intelligence refers to human-level ability across most tasks. Agentic systems today are narrow: they pursue specific goals with specific tools and permissions.
Does agentic AI replace support teams?
It takes over repetitive, policy-driven contacts and hands the rest to people with the context already gathered. Teams shift toward exceptions, complex cases, and improving the agent.
How should a company start?
Pick one high-volume request type with a clear policy, such as order status or rebooking. Give the agent read-only access first, measure it on real conversations, then allow actions one at a time.
References
- Anthropic. Building effective agents. December 2024.
- Yao, S. et al. ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.
- Russell, S. and Norvig, P. Artificial Intelligence: A Modern Approach, 4th edition. Pearson, 2020. Chapter 2.
- Schick, T. et al. Toolformer: Language Models Can Teach Themselves to Use Tools. NeurIPS 2023.
- OWASP GenAI Security Project. LLM01: Prompt Injection. OWASP Top 10 for LLM Applications, 2025.
- European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50.
- Milgram, S. Obedience to Authority: An Experimental View. Harper & Row, 1974.
- Model Context Protocol. Specification and documentation.
- Bandura, A. Social Cognitive Theory: An Agentic Perspective. Annual Review of Psychology 52 (2001): 1-26.