What is an AI agent?
An AI agent is software that works towards a goal on its own. You tell it what you want done. It plans the steps, uses tools such as your inbox, calendar or database, looks at what happened, and adjusts. A person sets the goal and the limits, and checks the result.
The key difference is choice. Something has to decide what happens next. In an agent, an AI model makes that call. In plain automation, a person decided in advance.
What is the difference between an AI agent, a chatbot and automation?
| Automation | Chatbot | AI agent | |
|---|---|---|---|
| What it does | Runs fixed steps when something triggers it | Holds a conversation and answers | Pursues a goal across several steps |
| Who decides the steps | A person, in advance | Mostly the conversation | The AI, within limits you set |
| Uses your systems? | Yes, in a set order | Sometimes, often just reads | Yes, and picks which ones |
| Handles surprises? | No. It stops or fails | Somewhat, in words only | Often, by trying another route |
| Predictable? | Very | Fairly | Least. Needs testing and guardrails |
Automation
Think of a rule that says: when a form is submitted, add a row to the sheet and send a thank-you email. It never varies and it never gets tired. It is also cheap to run and easy to trust.
Chatbot
A chatbot talks. A visitor asks about opening hours or returns, and it answers from what it has been given. A basic one can only point to pages. A good one is wired to your documents so its answers are right. Either way, it talks and rarely acts.
AI agent
An agent acts. Ask it to follow up every quote that has gone quiet for a week and it can find those quotes, read the history, draft a message in the right tone, send it for approval and log what it did.
When should you use each one?
Start from the job, not the buzzword.
- Use automation when the steps are always the same and the inputs are tidy. Moving data between tools. Sending reminders. Filing a document under the right name.
- Use a chatbot when people mostly need answers. Staff asking where a policy is. Customers asking the same ten questions.
- Use an AI agent when the work has several steps, the path varies, and judgement helps. Researching a property, chasing a document across email threads, preparing a first draft of a report from several sources.
Often the answer is a mix: automation for the fixed parts, an agent for the parts that vary, a chat window as the front door. And where the answer must be identical every time, a rules engine may be the safer part to put in the middle.
What are some examples of AI agents in a business?
- Property research. Gathers listings, past sales and notes on a property into a single brief for the team.
- Admin triage. Reads incoming email, sorts it, drafts replies and puts the tricky ones in front of a person.
- Document analysis. Reads a stack of contracts or reports, pulls out the key terms and flags what doesn't match.
- Customer follow-up. Notices when a lead has gone quiet and prepares a message for someone to approve.
- Reporting. Collects numbers from several systems and writes the weekly summary.
For real estate teams, this is the idea behind Agentsy for Real Estate: AI assistants for admin, property research and day-to-day communication. Our AI agents page explains what we build and how we test it.
What do AI agents cost and what can go wrong?
An agent costs more to run than a chatbot, because it makes more model calls to finish one job. That AI usage is billed by the provider each time. Ask for it to be estimated before you commit. We bill AI usage at cost, in your name, so you can see it.
The risks rise with what the agent is allowed to do.
- Wrong action. An agent that can send, delete or pay can get it wrong. Start with read-only, then drafts, then actions with approval.
- Prompt injection. Text it reads, in an email or on a web page, may try to give it orders. Limit what it can do after reading outside text.
- Silent drift. It works well on day one and then something changes. Keep a set of test cases and run them often.
- No one in charge. Name a person who owns what the agent does.
The sensible way in is a ladder. Give the agent a small job with a person approving each result. Watch it for a few weeks. Widen its limits only as it earns trust. For the security side, see vibe coding security risks.
What should you ask before building an agent?
- What is the goal, in one sentence, and how will we know it was met?
- Which tools and data can it touch, and which are off limits?
- What can it do alone, and what needs a person to approve?
- What happens when it is unsure or fails?
- How will we see what it did, and undo it if we need to?
- Who owns it once it is running?
If the answers come easily, you're ready to try one. If they don't, you've found the work to do first, and it is cheaper to find it now.
Do you need an AI agent at all?
Maybe not. If a simple automation does the job, use it. It is cheaper, steadier and easier to explain to your team. If people only need answers, a well-fed chatbot is enough. Reach for an agent when the work really does branch and a person is currently spending their day doing the branching.
Our own line is "AI implementation. Not AI theatre." An agent that looks clever in a demo and does nothing on Monday is theatre. One that takes the repeat work off your team is the point: keep the people, lose the busywork.