What agentic AI is, and what an AI agent is

Agentic AI is the family. An AI agent is one member of it.

Agentic AI means AI that does a task in steps. It reads what it needs, uses a tool, checks the result and goes again until the job is done. A tool is any function the AI is allowed to call, such as “find this customer” in your ERP. An ERP is the software a company runs its orders, stock and accounts in.

Anthropic, the company that makes Claude, draws the line most clearly. Its engineering guide of 19 December 2024 calls any system where an AI model works with tools an “agentic system”. The AI model is the engine that reads and writes the words, like the one behind Claude or ChatGPT. The guide then splits these systems in two:

  1. Workflows. Code fixes each step in advance. The AI model does its part at each step.
  2. Agents. The AI model decides its own steps, and which tools to use, as it goes.

So every AI agent is agentic AI. A fixed workflow that uses an AI model counts as agentic AI too, even with no agent in it. The glossary has the one-line versions of both terms.

Four vendors' definitions in plain words

We read each vendor's own page on 15 November 2026 and put its definitions in plain words below. The source list at the foot links every page.

Scroll the table sideways to see every column.

SourceWhat agentic AI meansWhat an AI agent is
Anthropic, Building effective agentsEvery system where an AI model and tools work through a task. It comes in two kinds: workflows and agents.A system where the model directs its own process and its use of tools.
OpenAI, A practical guide to building agents (34-page PDF)The guide does not define the term. It calls the series of steps toward a goal a workflow.A system that does tasks for you on its own. The model runs the workflow and picks tools to act. A plain chatbot is not one.
Google Cloud, What is agentic AI?A branch of generative AI, the AI that makes new text or code. It coordinates agents, each using a language model to act through tools, and uses several agents for complex work.A building block of agentic AI that focuses on one task.
IBM Think, What is agentic AI? and What are AI agents?A system that reaches a set goal with limited supervision. It is made of AI agents, coordinated when there are several.A system that does tasks on its own by planning its steps with the tools it has.

The one point where they disagree

All four treat agentic AI as the wider term. They split on what makes a system agentic.

For Anthropic, one model running fixed steps with tools already counts. Google Cloud and IBM describe agentic AI as agents working together toward one goal. On Google's reading, one agent doing one task is an AI agent, and the group working together is the agentic AI.

Both readings agree on the part that matters when you buy or build: an AI agent chooses its own next step, and a workflow does not.

Agentic AI, AI agents and chatbots side by side

Three shapes of AI software, compared on what changes for a business. The first column, a plain chatbot, sits outside agentic AI.

Scroll the table sideways to see every column.

Chatbot or other generative AIWorkflow (agentic AI)AI agent (agentic AI)
Who picks the next stepYou, one prompt at a timeCode written in advanceThe AI model, as it works
Uses tools such as an ERP's APINot in a plain chatbot, by OpenAI's definitionYes, at points the code setsYes, and it chooses which tool and when
Best fit, in the vendors' adviceAnswering and draftingTasks that split cleanly into fixed steps (Anthropic)Open-ended problems where nobody can predict the number of steps (Anthropic)
Cost and riskLower: Anthropic says agentic systems often trade speed and cost for better resultsPredictable, since the path is fixedHigher cost, and one wrong step can feed the next (Anthropic)
ERP example from vendor documentsSAP Business One's Ask AIAn Odoo 20 AI server actionAn Odoo 20 agent with skills, or NetSuite's agents

What agentic AI looks like inside an ERP

Three ERPs our team has worked in now ship AI of their own. Read against the definitions above, each one lands in a different column.

SAP Business One: generative AI only

SAP's guide to the Business One web client, release 10.0 FP 2608, describes Ask AI. You ask a question about the screen you are on. SAP says the AI answers from its general knowledge and the public internet. It cannot read, query, create, update or delete any Business One data. By OpenAI's test, that makes it a chatbot feature, not an agent. Our note on SAP Business One and Joule covers what else SAP has and has not shipped.

Odoo 20: both shapes

Odoo builds its agents from four parts:

  1. Instructions set the agent's job and how it behaves.
  2. Knowledge is information the agent can use while it works.
  3. Skills and tools set what the agent can do.
  4. Triggers decide when the agent starts, such as when a user opens a chat with it.

Odoo says an agent with no skills can only give information. Once it has skills, its tools can create or update records. Odoo also warns that an agent started by a trigger acts directly. It does not ask a user to approve the records it creates or changes.

Odoo's AI server actions have the other shape. An AI server action is one step inside an Odoo automated workflow. The AI reads a record and picks one tool from a set list, plus the tool's inputs. A standard server action then does the work. Odoo says the AI part does not enforce business rules, so the tool's own code must. In our reading this matches what Anthropic calls routing: a step that sorts each input and sends it down one of a few set paths.

NetSuite: agents inside the user's role

Oracle's agents help users find data and finish work inside supported NetSuite workflows. You start one from Ask Oracle, where you ask a question in everyday words, or from a NetSuite page. Oracle says they cannot reach any record, report, page or action that the user's role cannot. A role is the set of rights a NetSuite login has.

Oracle also says the agents do not replace NetSuite's approval workflows or your review of results. Some of their features need NetSuite Next, which Oracle is releasing to customers in phases.

All three ERPs also open a door for an agent built outside them: Oracle's AI Connector Service, Odoo's MCP server and SAP's MCP server for the Service Layer, Business One's web API. MCP, the Model Context Protocol, is a shared standard that lets AI apps call a system's tools. Our guide on how to build an AI agent for an ERP walks through each route.

Which one an ERP job needs

Start with the simplest build that does the job. Anthropic's guide says exactly this, and adds that the answer is sometimes no agentic system at all.

Much ERP paperwork has known steps. Take an emailed purchase order:

  1. Read the order in the email.
  2. Match each line to the item master, the list of every product you sell.
  3. Apply the customer's price list.
  4. Save a draft sales order for a person to check.

That is a workflow. The AI model does the reading, and code keeps the steps in order.

An agent earns its higher cost where the steps change from case to case. One example is an exception queue, where each stuck order needs a different lookup before anyone can fix it.

Four questions that show which shape a job needs

  1. Can you write the steps down in order before the job starts? Then build a workflow. Anthropic says workflows give predictable, consistent results on well-defined tasks.
  2. Does each case need a different path? Then an agent may be worth its cost. Anthropic suggests agents where nobody can predict how many steps a problem will take.
  3. Is the job hard to write as rules? OpenAI's guide names three signs: judgment calls and exceptions, rule sets too big to keep up to date, and messy documents or plain language to read. If none fits, it says, a fixed program may be enough.
  4. Does it write to your ERP? Then plan where a person approves, whichever shape you build. The OWASP Gen AI Security Project lists high-impact actions taken without a check as a risk.

This split is our reading of the two guides, applied to ERP work.

Our team's ERP work covers RFQ automation, daily bookkeeping, and purchase and sales order generation, in NetSuite, SAP Business One, Odoo, ERPNext and custom ERPs. An RFQ is a request for a quote. Whichever shape a job takes, we build it to stop at a draft that a person approves. See AI agent development for how a project runs.

What four AI assistants said about agentic AI

On 14 November 2026 we asked four AI assistants the same question. They were the ChatGPT app, Claude, the Gemini app and Perplexity. The question was: “What is agentic AI and how is it different from a chatbot?”

All four gave the same core answer. Agentic AI has a goal, works in steps and uses tools. Perplexity added that a chatbot can be the front end of an agentic system.

None of the four cited Anthropic's or OpenAI's guides. The ChatGPT app and Gemini both cited IBM, and the ChatGPT app also cited Google Cloud.

Each assistant answered once, so treat this as a sample of one day, not a rate.

Sources

Every source behind a figure or a vendor fact in this article, with the day we read it.