Agentic AI vs AI agents vs automation: what is the difference?
What each term means, how they compare on autonomy, tools, memory, failure and cost, and how to tell when a plain workflow is the better build.
Automation follows rules a person wrote down. An AI agent is software that uses a language model to choose its own next step and then take it in your systems. Agentic AI is the wider idea: giving a model a goal, tools and room to decide. Most Australian businesses need all three, with rules doing most of the work.
The three terms get used as if they mean the same thing, and that confusion costs money in both directions. Businesses pay per run for an agent to do a job a free workflow would do, or they force a workflow to cope with work that needs judgement and end up with forty branches that still miss the odd case. This post sets out what each term means, how they compare, and how to tell which one a job needs. For the wider picture of agents in an Australian business, start with our guide to AI agents for business.
What is agentic AI?
Agentic AI describes a way of building with language models: the model is given a goal and some tools, and it decides which steps to take to reach the goal. It is a property of a system, so a system can be more or less agentic depending on how much of the path the model chooses.
Anthropic, the company behind Claude, uses "agentic systems" as the umbrella term and splits it in two. In its engineering guide Building effective agents it says:
Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
That split is the useful one for a business. In an agentic workflow, your process sets the order and the model does the steps that need reading or writing. In an agent, the model sets the order as well.
What is an AI agent?
An AI agent is the software itself: a language model with three things added. It has a goal, it has tools it can call (search, read a CRM record, send an email, update a field), and it is allowed to take several steps in a row without a person clicking through each one. Take away the tools and you have a chatbot, which can answer but cannot act.
The agents we run on Claude each do one job a team can name. For Plungie, an agent writes a short brief on each buyer from HubSpot and sales data before a salesperson's call. For Flint, an agent reads the documents a borrower uploads, fills the loan record and flags mismatches for a broker to check. Both decide their own steps inside a narrow job, and both hand anything uncertain to a person.
We run one on our own sales team too. Our Sales Briefer reads the CRM, researches the company and emails the salesperson a pre-call brief, with nobody clicking through the steps. When we filmed it in June 2026 it had built 255 briefs and sent 121 of them straight to an inbox.
What is automation, and is a HubSpot workflow AI?
Automation is any process a computer runs from rules a person wrote: when this happens, do that. A HubSpot workflow is the everyday example. A form is submitted, the contact is assigned to an owner, a task is created and an email goes out, in that order, every time. There is no model involved and nothing is decided at run time beyond the if-then branches you built.
That predictability is the point. A workflow does the same thing on the thousandth run as on the first, it can be read and checked by anyone with access to the portal, and inside a HubSpot plan you already pay for it costs nothing extra each time it fires. HubSpot's guide to creating workflows covers the mechanics.
The lines are now blurring inside HubSpot itself. A workflow can hand one step to an agent and take the result back, which is an agentic workflow in Anthropic's sense. We walk through that pattern in running the Data Agent inside your HubSpot workflows.
How do agentic AI, AI agents and automation compare?
The clearest way to separate them is to ask the same questions of each.
| Automation (rules workflow) | Agentic workflow | AI agent | |
|---|---|---|---|
| Who decides the next step | The rules you wrote | The rules, with a model doing one or more steps inside them | The model, working towards a goal you set |
| Tools | The actions you built in | The actions you built in, plus whatever the model step is allowed to call | Any tool you give it access to, chosen by the model at run time |
| Memory | Only what is stored on the record | The record, plus what the model step is given for that run | Everything it has read during the run. Anything between runs has to be built in, usually as notes written back to the CRM |
| Copes with variation | Badly. Anything outside the rules falls through | Within the step the model handles | Well, which is also why it needs checking |
| How it fails | Loudly and the same way each time: a branch nobody built, a property left blank | Mostly loudly, with a model step that can be wrong in a plausible way | Plausibly. A fluent, confident answer or action that is wrong, which is harder to spot |
| Cost per run | Nothing extra inside a HubSpot plan you already pay for | The model usage for the AI steps only | Model usage on every step, so it rises with how much it reads and how many steps it takes |
| Testing | Run it with a test record and check the result | Test the rules as normal, then test the model step on real cases | A fixed set of real cases with known answers, scored before launch and checked again after |
| Good first job | Routing, reminders, field updates, handoffs | Summarising a call into a record, classifying an inbound email, then routing by rules | Research, triage and drafting from several sources |
Most real systems sit somewhere along that line. The question is how much of the path you let the model choose for a given job.
The four boxes are a way of describing how much of the path the model chooses. They are not product names.
When is a plain workflow the better build?
More often than the talk about agentic AI suggests. Anthropic's own advice in the same guide is to find the simplest solution possible, and only increasing complexity when needed
. In our experience a workflow is the better build when any of these is true:
- You can write the rule down. If the job is "when X happens, do Y" with no "it depends", a model adds cost and a way to be wrong.
- The same input must always give the same output. Commission rules, invoice timing, lifecycle stage changes and anything an auditor will ask about belong in rules you can show someone.
- It runs at high volume for low value. A reminder sent ten thousand times a month costs nothing as a workflow and real money as an agent.
- Your data is not ready. An agent believes whatever your CRM tells it. If there are four records for one company, fix that first. Our five data checks are the place to start.
- Nobody will own it. An agent needs a named person who reads its output and is told when it fails or goes quiet. A workflow needs far less looking after.
The fuller test for HubSpot, with the costs side by side, is in do you need an AI agent, or will a HubSpot workflow do it?
When does it make sense to go agentic?
When a person currently reads something before deciding what to do, and every case is a little different. Support questions written a hundred ways, researching a prospect before a call, and pulling an answer out of several documents are the usual examples. Rules handle those badly, and the per-run cost of a model buys something a workflow cannot do.
Even then, start with as little autonomy as the job allows. Let the rules run the process and give the model only the step that needs judgement. If that works, widen it. When we build an agent, the prototype runs in shadow mode for 4 to 8 weeks, doing the work alongside your team and changing nothing until its output has been checked. Anything risky or irreversible waits for a person, under the action policy in our AI Security and Governance Overview.
Before anything goes near a customer, read how we test an AI agent and whether you can undo what an agent just did. Australian guidance on agents, including the Australian Cyber Security Centre's paper on careful adoption of agentic AI, is set out in our AI governance guide.
What does each cost to run in Australia?
A HubSpot workflow costs nothing extra per run inside a plan you already pay for. An agent spends money every time it runs. On HubSpot, a resolved Customer Agent conversation is 50 US cents, about 77 cents AUD before GST. On your own build, you pay the model provider's usage price for every step the agent takes, so an agent that reads more or takes more steps costs more each run.
The full numbers are in what HubSpot's AI agents cost to run in AUD and what a custom AI build costs an Australian business. Our AI consulting page sets out the build and running ranges for our own work.
Frequently asked questions
Is agentic AI the same as an AI agent? They overlap, but they are different things. Agentic AI is the approach of giving a model a goal, tools and room to decide its steps. An AI agent is a piece of software built that way. A system can be partly agentic, such as a workflow with one model step, without being a full agent.
Is a HubSpot workflow an AI agent? A standard HubSpot workflow follows if-then rules you set and involves no model. It becomes an agentic workflow when one of its steps hands work to an AI agent, such as the Data Agent, and takes the result back.
Is ChatGPT or Claude an AI agent? On its own, the chat app is an assistant that answers. It acts as an agent when it is given tools and allowed to take steps in your systems, such as reading your CRM, searching the web or sending an email on your behalf.
Is agentic AI more expensive than automation? Per run, it usually is. A HubSpot workflow costs nothing extra inside your plan, while an agent pays for model usage on every step. It is worth it when the job needs judgement a rule cannot give, and money wasted when the rule could have been written down.
Should we replace our workflows with agents? Rarely all of them. Keep the rules that work and hand only the steps that need reading or judgement to a model. Most of what a business wants automated turns out to be rules.
Read next
- AI agents for business in Australia: a practical guide
- Do you need an AI agent, or will a workflow do?
- What HubSpot's AI agents cost to run in AUD
- Running AI agents inside your HubSpot workflows
- AI governance in Australia: the 2026 guide
- Claude for business in Australia
Sources
- Building effective agents, Anthropic, for the definitions of agentic systems, workflows and agents.
- Create workflows, HubSpot Knowledge Base.
- HubSpot pricing, for Breeze credits and agent pricing. Prices change, so check on the day you budget.
- AI Security and Governance Overview, Neighbourhood, for the action policy our agents run under.
Neighbourhood builds AI agents on Claude for Australian businesses, and builds HubSpot workflows where rules will do the job. Our AI engineers work out which one a job needs before anything is built.
Agent, agentic workflow or plain automation?
Book 15 minutes and bring one job your team does every week. We will tell you which of the three it needs, roughly what it would cost to run, and what your data needs first.