Everyone is suddenly selling "AI agents". Your software vendors, your bank, the bloke at the networking breakfast. If you've nodded along without being sure what one actually is, you're not alone, and you're not behind. Here's the plain-English version, with no jargon and no hype, written for people who run a business rather than a lab.
An AI agent is software that is given a job, looks at the information it needs, decides the next step, takes that step using the tools you've allowed it to use, and checks in with a person when something matters.
That's it. The important words are job, information, tools and checks in. A chatbot answers a question. An automation follows a fixed recipe. An agent works towards an outcome, and it can change course along the way.
A simple example. A new enquiry lands on your website at 9pm. A basic automation sends "Thanks, we'll be in touch." An agent reads the enquiry, checks your CRM to see whether this person is already a customer, works out what they're asking for, drafts a proper reply with the right information attached, books a callback in the right person's calendar and leaves a note on the record so your team knows exactly what happened. Then it waits for someone to approve the reply before it goes out.
Most businesses already have some automation: an email that fires when a form is submitted, a task created when a deal moves stage. Those are rules. "If this happens, do that." They're reliable and cheap, and they're still the right answer for plenty of jobs.
Rules break down when the input is messy. An email that says "Hi, following up on the quote from last month, can we push the start date and add the second bathroom?" doesn't fit a rule. Someone has to read it, find the quote, understand the change and decide what happens next. That reading and deciding is the part an agent can take on.
In practice the best setups use all three. Workflows for the predictable parts, an agent for the judgement in the middle, a person for anything that carries real risk.
This is the part the sales pitches skip. An agent is only as good as what it's given. Think of it like a capable new staff member on their first day. Smart, quick, keen. But if you hand them a messy shared inbox, three spreadsheets that disagree and no instructions, they'll make a mess too.
The agent needs to find the right customer, the right job, the right invoice and see how they relate. That's why we build agents on top of HubSpot. Contacts, companies, deals, tickets and custom objects all sit in one CRM, joined together by associations. The agent has a map to follow instead of a pile of paper. We wrote about getting systems ready for this in preparing your systems for AI, and the advice still holds.
"Help with sales" is not a job. "Reply to every new website enquiry within ten minutes with a draft a person can approve" is a job. Pick one, agree what done looks like, and measure what it costs you to do it by hand today.
Can it send emails or only draft them? Can it change a deal stage? Can it talk about price? Write these down before you switch anything on. We start every agent in draft mode, where a person approves its work, and it only earns more freedom once its work is consistently approved without changes.
Somebody in your business should look at what the agent did each week, fix what it got wrong and decide whether it's worth what it costs.
The best first agents are boring. They take a repetitive admin job that eats hours and does nothing for your reputation when it's done late. A few we see again and again:
Notice what's not on the list: signing contracts, approving refunds, giving legal, financial or health advice. Those stay with a qualified person. An agent can prepare the paperwork and route it to the right person, but the decision is theirs.
Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027. In our experience the reasons are rarely the AI itself. They're the same reasons any project fails: no clear job, messy data, nobody owning the result, and no way to tell whether it's paying for itself.
The fix is unglamorous. We use a simple four-step method:
If that sounds a lot like hiring and managing a person, that's the point. The businesses getting value from AI treat agents like team members with a job description, not like magic. If you want the bigger picture on closing that gap, our post on the AI implementation gap for Australian businesses is a good next read.
Maybe not yet. If your CRM is half empty, your team keeps customer details in their heads and nobody can tell you how long enquiries take to answer, start there. An agent will expose those gaps fast. If your data is in decent shape and you've got a repetitive job that keeps falling through the cracks, an agent is probably worth a proper look. You can see the kinds of agents we build on our agents page.
Not sure which side of the line you're on? Book an AI Readiness Audit and we'll look at your HubSpot data, pick the one job worth automating first and tell you honestly whether an agent will pay for itself. If you already know the job, go straight to Build my solution.