Everyone is selling AI agents. Very few people selling them can explain what one is, how it differs from the automations you may already have, or what it will cost to run once the demo is over. This guide is AI agents explained in plain English, from someone who builds them for businesses in Wollongong and the Illawarra.
Short version: most businesses need fewer agents and more boring automation than they think. But where an agent is the right tool, it can take on work that used to need a person reading, deciding and typing all day.
Automation vs workflow vs agent
These three words get used interchangeably. They are not the same thing, and the difference changes the cost, the risk and how you build it.
Automation
An automation is a fixed rule: when this happens, do that. When a form is submitted, create a contact. When a deal is won, create a project. There is no judgement involved. It does the same thing every time, which is exactly what you want for predictable work.
Workflow (with AI steps)
A workflow is a sequence of steps that you design, where one or more steps uses AI to do something a rule cannot. For example: when an enquiry comes in, AI reads the message and classifies it as sales, support or junk, then fixed rules route it. You still decide the path. AI fills in specific gaps along the way.
Agent
An agent is given a goal, a set of tools and some rules, and decides for itself which steps to take to reach the goal. It might look up a contact in the CRM, read past emails, check a price list, draft a reply and create a task, in whatever order makes sense for that case. It loops: act, check the result, decide what to do next.
| Automation | AI workflow | AI agent | |
|---|---|---|---|
| Who decides the steps | You, in advance | You, in advance | The agent, at run time |
| Handles messy input | No | Yes, at specific steps | Yes, throughout |
| Predictability | Very high | High | Lower, needs guardrails |
| Best for | Repetitive, rule-based tasks | Known process with a reading or writing step | Varied cases needing research and judgement |
| Example | Won deal creates an Asana project | AI classifies enquiries, rules route them | Agent researches a lead and prepares a tailored first reply |
If you can draw the process as a flowchart with no "it depends" boxes, you need an automation, not an agent.
How an agent actually works
Under the hood, an agent has four parts:
- A model. The AI doing the thinking, such as Claude or an OpenAI model, accessed through an API.
- Instructions. A detailed brief: the goal, the rules, the tone, what it must never do, and when to stop and ask a human.
- Tools. Specific actions it is allowed to take, such as search the CRM, read a document, create a draft, update a field. Each tool is a defined connection to one of your systems, usually through an API.
- Context and memory. The information it can see for this case: the email thread, the client record, your policies, past notes.
The quality of an agent depends far more on the instructions, the tools and the data than on which model you pick. A great model with access to a messy CRM and vague instructions will make confident mistakes.
What each one costs to build and run
Here are my prices for builds (AUD) and an honest view of running costs. Running costs depend heavily on volume, so treat those as ranges to estimate against, not quotes.
| Type | Typical build effort | Running cost drivers | Build price from |
|---|---|---|---|
| Single automation | Days | Automation platform fees or hosting, minimal | $2,500 |
| AI workflow or set of automations | One to a few weeks | Platform fees plus AI usage per item processed | $8,500 |
| AI agent | Several weeks, including testing | AI usage per task (higher, as agents take multiple steps), hosting, monitoring | $12,000 |
Estimating running costs
AI usage is billed by the amount of text processed (tokens). An agent uses more than a simple workflow because each case involves several steps, each reading and writing text. A rough way to estimate:
- Count how many cases per month the agent will handle.
- Estimate how many steps each case takes and how much text each step reads (an email thread, a client record, a policy document).
- Run a test batch of 20 to 50 real cases and look at the actual usage. That beats any estimate.
- Multiply out, then add a buffer of 30 to 50 percent for edge cases and retries.
Example: an agent that handles 400 inbound enquiries a month, with a test batch showing an average AI cost of 15 cents per case, would cost about $60 a month in model usage, or $90 with a buffer. Add hosting, monitoring and a few hours a quarter of maintenance. The usage cost is often the smallest line. Maintenance is the one people forget.
Illustrative figures. Model pricing varies by provider and model, and changes over time.
Is it worth it? The payback test
Same example. If each of those 400 enquiries took a person 8 minutes to read, look up, categorise and draft a first reply, that is about 53 hours a month. If the agent cuts the human part to 2 minutes of review, you get back about 40 hours a month. At $65 an hour loaded, that is around $2,600 a month, or about $31,000 a year, against a build from $12,000 and modest running costs.
Now run the same maths on 40 enquiries a month instead of 400. You save about 4 hours a month. That is a nice-to-have, not a project. Volume is what makes an agent pay.
How to pick your first use case
Score each candidate out of 5 on these six questions. Anything scoring under 20 out of 30 is probably not your first project.
| Question | Scores high when |
|---|---|
| Volume | It happens dozens or hundreds of times a month |
| Time per case | Each case takes a person several minutes or more |
| Messy input | Input is free text: emails, documents, forms with comments |
| Low stakes per mistake | A wrong answer is caught in review and is cheap to fix |
| Data available | The information needed lives in systems the agent can reach |
| Clear definition of good | You can show examples of a good and bad outcome |
Good first use cases
- Enquiry triage: read each enquiry, classify it, pull out the key details, update the CRM and draft a reply for review.
- Document intake: read incoming documents such as supplier quotes or application forms, extract the fields and populate the right records.
- Meeting to CRM: turn a call transcript into notes, next steps, deal updates and tasks.
- Internal knowledge assistant: answer staff questions from your policies and procedures, with a link to the source.
- Lead research: before a sales call, gather what is in your CRM and the company's website into a one-page brief.
Poor first use cases
- Anything that sends messages to clients without a human seeing them first.
- Decisions about money, credit, employment, health or legal matters.
- Processes that happen a handful of times a month.
- Processes nobody has written down or agreed on.
- Anything that depends on data you know is a mess.
Worked example: an enquiry triage agent
Here is what a typical first agent looks like end to end, for a service business getting a few hundred enquiries a month through its website, email and phone messages.
- Trigger. A new enquiry arrives: a form submission, an email to the info inbox, or a voicemail transcript.
- Look up. The agent searches HubSpot by email and company domain. Existing client, open deal, or new?
- Classify. It reads the message and decides: new work, existing job question, complaint, supplier, job applicant or spam.
- Extract. It pulls out the details that matter: service needed, location, timing, budget if mentioned, phone number.
- Update. It creates or updates the contact, sets the lead source and service interest, and creates a deal if it is new work.
- Route. It assigns an owner using your rules (by service and region) and creates a follow-up task due today.
- Draft. It writes a first reply in your tone, using your service descriptions and FAQs, and saves it as a draft for the owner.
- Escalate. Complaints, anything mentioning legal action, and anything it is unsure about go straight to a named person with a summary.
Notice how much of that is plain automation with a few AI decisions in the middle. Steps 2, 5 and 6 are rules. Steps 3, 4 and 7 are where the AI earns its keep. That mix is normal, and it is why I always design the process first and decide where the AI sits second.
What the person gets: a CRM record already filled in, a task, a summary and a draft reply. Their job shrinks from reading, searching, typing and writing to checking and sending.
Guardrails
Guardrails are the limits that stop an agent doing something stupid, expensive or embarrassing. A well-built agent has several layers:
- Least access. The agent gets only the tools and permissions it needs. Read access to the CRM does not mean delete access.
- Draft, do not send. For anything client facing, the agent prepares and a person approves. You can relax this later for low-risk categories once you trust it.
- Hard rules in the instructions. Never quote prices, never promise dates, never give legal or medical advice, always escalate complaints.
- Hard rules in code. Some limits should not rely on the AI following instructions. Cap how many records it can update per run. Block certain fields from being changed at all.
- Confidence and escalation. When the case is unusual or information is missing, the agent hands it to a person instead of guessing.
- Logging. Every action is recorded: what it saw, what it decided, what it did. You need this to debug, improve and answer "why did it do that?".
- Spending limits. Usage caps with your AI provider so a loop or bug cannot run up a surprise bill.
- A kill switch. One simple way to pause the agent immediately.
Human in the loop
Human in the loop means a person reviews or approves the agent's work at defined points. It is not a sign the agent is bad. It is how you get value safely while you learn what it gets right.
I usually stage it like this:
Shadow mode
The agent runs alongside the team. It produces outputs but nothing is used. You compare its work with what people actually did.
Approve everything
The agent's drafts and updates go into a queue. A person approves, edits or rejects each one. Track the edit rate.
Approve by exception
Low-risk categories with a low edit rate go through automatically. Everything else still gets reviewed.
Spot checks
A person reviews a random sample each week and all escalations. Rules get tightened when issues appear.
Move between stages based on evidence, not enthusiasm. If people are editing more than one in five drafts significantly, you are not ready for the next stage. Fix the instructions, the tools or the data first.
For anything that makes or significantly affects decisions about people, keep a human in the loop permanently. It is good practice and, under Australian privacy law, automated decisions about individuals are getting more attention, not less. Get legal advice if you are unsure.
Data readiness: the bit everyone skips
An agent is only as good as what it can see. If your CRM has three records for the same client, half the phone numbers in the notes field and deal stages nobody updates, the agent will work from that mess.
Before building, check:
- Is the data in a system, or in people's heads and inboxes? An agent cannot read what was never written down.
- Is there one source of truth for clients, jobs and pricing, or several competing copies?
- Are key fields filled in consistently? Lifecycle stage, owner, industry, status.
- Are duplicates under control?
- Are your policies and procedures written and current? If the agent is meant to follow your refund policy, the refund policy has to exist.
- Can the systems be reached by API? Most modern tools, including HubSpot, Asana, PandaDoc and WordPress, can. Some older or locked-down systems cannot.
If several of those are a no, start with a data clean-up and some basic systems integration. It is less exciting than an agent, and it is what makes the agent work. The CRM data clean-up guide walks through it.
FAQ
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a conversation. An agent pursues a goal by taking actions in your systems, such as looking up records, updating fields and creating drafts. Some chatbots have agent features, but answering questions alone does not make something an agent.
How much does it cost to build an AI agent?
My AI agent builds start from $12,000. Simpler AI workflows start from $8,500 and single automations from $2,500. Running costs are usually modest at small volumes, but maintenance should be budgeted too.
How long does it take to build an agent?
Several weeks for a focused first agent, including connecting the tools, writing and testing the instructions, and a period in shadow mode. Most of the time goes into testing on real cases, not writing code.
Which AI model should an agent use?
It depends on the task, cost and data terms. I build with Claude and OpenAI models and pick per use case. The instructions, tools and data quality matter more than the model choice.
Can an agent work with HubSpot?
Yes. HubSpot has a solid API, so an agent can read and update contacts, companies, deals, tasks and notes within the permissions you give it. It is one of the most common setups I build.
Next steps
List three processes that eat your team's time. For each, ask: could I draw this as a flowchart with no "it depends" boxes? If yes, automate it. If no, score it with the table above. If it scores well and your data is in decent shape, it might be your first agent. If you want a second opinion, get in touch.