An AI agent is software that uses an AI model to work toward a goal in steps, with tools: it reads an email, looks up a customer record, checks a calendar, books the appointment, updates the file and drafts the reply, deciding the next step from what it just found. A chatbot answers a question in a chat window. An agent goes and does part of the work, which makes it more useful and also means it needs tighter rules.
For a small business, that difference decides almost everything: what an agent can take off your plate, what it can break, and how it should be set up. Wiring AI into the email, files and customer records you already use is our AI integration work, and agents are the part of it that takes action. Below is what they are, where they fit by role, where they go wrong, and what a well-built one looks like.
Chatbot, automation or agent: what’s the difference?
All three get called “AI” in sales pitches. They behave very differently.
| What it does | Good for | Watch out for | |
|---|---|---|---|
| Chatbot | Answers questions in a chat window | Hours, services, policies, “do you do this?” | Confident wrong answers |
| Automation | Follows fixed rules you wrote: when X happens, do Y | Steps that run the same way every time | Breaks quietly when something changes |
| AI agent | Reads the situation, picks the next step and uses tools to take it | Work where the steps change from case to case | Wrong actions, loops, too much access, cost |
Plain automation is the dependable workhorse. A form comes in, a row is added to the spreadsheet, a text goes out. No judgment involved, and that’s the point. An agent earns its place when the steps can’t all be written down ahead of time: a caller wants a different service than the last one, a vendor’s invoice comes in a new layout, a customer’s email asks three things at once.
Anthropic, the company behind Claude, draws the same line in its engineering guide Building effective agents (published December 2024, checked October 7, 2026). It describes workflows as AI and tools “orchestrated through predefined code paths,” and agents as systems where AI models “dynamically direct their own processes and tool usage.” Its advice is to “find the simplest solution possible” and add complexity only when it’s needed. We agree. If a fixed rule does the job, you don’t need an agent, and we’ll tell you so.
What an AI agent could do in a small business, by role
These are the kinds of jobs we look for. Each one maps to problems we lay out on our industry pages, and in every one a person stays in charge of what leaves the building.
The front desk
The phone is where most owners feel it first. Brittany was losing calls she couldn’t get to, so we built her an AI voice agent that answers the phone. It picks up every time, holds a real conversation, knows her services, qualifies the caller, books straight into her calendar and texts a confirmation. That’s an agent in the plain sense of the word: it listens, decides what the caller needs, then uses tools (her calendar and a text message) to get it done.
The same pattern fits a property manager whose urgent leak calls need to reach whoever is on call, or a clinic where a new patient gets voicemail at 8:41 in the morning.
The office
Office work is mostly reading, finding and retyping, which is exactly what agents are good at. At an accounting or law office, an agent could:
- ask clients for missing documents, send reminders, and file what comes back
- read statements and pull the figures out for a person to check
- draft the client email from the meeting notes, ready for a person to edit and send
At an insurance agency, it could read dec pages and flag renewals early. At a trucking company, it could read a rate con into dispatch for the dispatcher to check.
The field
Crews don’t want to type at 9 at night. For trades and field service, a tech speaks the field notes on the way to the truck, and an agent turns them into the job ticket, adds the parts used and flags anything that needs a quote. On an oilfield crew, the same idea turns a spoken daily report into a written one, and pre-fills the JSA for the crew to review.
The owner
Owners mostly need to know what’s slipping. An agent with read-only access could go through yesterday’s open quotes, jobs finished but not invoiced, and follow-ups nobody made, then hand you a short list each morning. It changes nothing and only points you to where to look, which makes it a good first agent for most businesses.
If you haven’t used AI much yet, start with ChatGPT for small business before jumping to agents.
Where AI agents go wrong
The same thing that makes an agent useful, that it acts, is what makes its mistakes cost more. These are the failures we plan for.
- The wrong action, done with confidence. AI models can misread a document or misunderstand a request and never flag the doubt. A chatbot that misreads gives you a bad answer. An agent that misreads books the wrong day, emails the wrong customer or updates the wrong record.
- Runaway loops. An agent that keeps retrying a step can send the same email twice, or ten times, or keep calling a system that’s down. Anthropic’s guide recommends “stopping conditions (such as a maximum number of iterations) to maintain control.”
- Access that’s too broad. OWASP, a nonprofit security community, lists “Excessive Agency” among the top risks for AI applications in its 2025 list (checked October 7, 2026). It names three causes: “excessive functionality,” “excessive permissions” and “excessive autonomy.” An agent that can read your whole inbox can also be fooled by a customer email with instructions hidden in it.
- Cost creep. Every step is a paid call to an AI model. Loops and long tasks multiply that quietly, and Anthropic’s guide warns that agents trade autonomy for “higher costs, and the potential for compounding errors.”
None of these mean agents are a bad idea. They mean the limits have to be decided before the agent goes live.
What a well-built AI agent looks like
Whoever builds your agent, these are the four things to ask for. They are the rules we hold every agent to, and they matter more than which AI model sits underneath.
- Least-privilege access. The agent gets only the logins and permissions its one job needs, and nothing else. A booking agent can see the calendar. It can’t see payroll. This matches OWASP’s guidance to “limit the permissions” an AI tool has “to the minimum necessary.”
- A person approves anything it sends, spends or deletes. Drafts cost nothing to throw away. Outgoing emails, payments, refunds, deleted records and anything a customer reads go past a person first. The only exceptions are narrow, routine steps a person signed off on in advance, like a booking confirmation the caller just asked for. OWASP calls this “human-in-the-loop control,” with a person approving “high-impact actions before they are taken.”
- A log of every action. Every lookup, draft, booking and update is recorded, so you can always see what it did and when.
- Start with one narrow job. One task, a short list of tools, and a clear handoff to a person when it isn’t sure. Prove that job works before asking it to do a second one.
When we build one, everything runs in accounts you own, logins are role-based, and we set it up so your data isn’t used to train public AI models. Our ChatGPT safety guide covers what happens to the information your team types in. How we build to these rules for a particular business, from what it connects to the limits it runs under, is what our AI integration work covers.
Do you need an AI agent?
Probably not for everything, and maybe not yet. A good first agent has three things going for it: the job eats real hours every week, the steps vary enough that a fixed rule keeps breaking, and a mistake can be caught before it reaches a customer. If the steps are the same every time, plain automation is cheaper and more predictable. If you just need answers, an assistant that answers from your own documents may be all it takes.
If you’d like a second opinion on which job in your business is worth handing to an agent, and which ones aren’t, that’s what the call below is for.