What is an AI agent, and what can one do for a small business?

Tools and permissions

On its own, a language model can only read and write text. What makes an agent useful is the set of tools it is given. A tool is one specific action in one specific system: search the CRM for a contact, create a draft in Gmail, read the attachments on an email, add a row to the job tracker.

Each tool comes with permissions, and this is where the most important design decisions get made. An agent that can read your calendar but not change it is very different from one that can book, move and cancel appointments. A sensible build gives each agent the smallest set of tools the job needs:

• Read-only access where it only needs to look things up
• Permission to draft rather than send, until you trust the output
• Write access limited to specific fields, folders or record types
• No access at all to systems unrelated to the job

Think of it like onboarding a new hire. You wouldn't hand a first-week employee the bank login and the admin password for every system. You give them access to what the job needs and widen it as they prove themselves. Agents deserve the same treatment.

How AI agents differ from chatbots, Zapier-style automation and RPA

These terms get mixed up, and sales pages don't always help. Here is the plain difference.

Chatbots

A chatbot answers questions in a conversation. Many website chatbots only answer from a script or an FAQ. Even a capable AI chatbot usually stops at words: it tells the customer something, but it doesn't go and do anything in your systems. An agent may have a chat window as its front door, but behind that door it can look up an order, book an appointment or open a ticket.

Rules-based automation

Tools like Zapier and Make connect apps with fixed rules: when this happens, do that. "When a form is submitted, add a row to this spreadsheet." They are reliable, inexpensive to run and easy to check, and for predictable tasks they are often the right answer. Their weakness is anything messy. If the information arrives as free text, a scanned PDF or an email that could mean three different things, a fixed rule can't read it.

Agents fill that gap. In practice the best systems mix the two: fixed rules for the predictable steps and an agent for the steps that need reading or judgment. Much of our workflow automation work looks exactly like that.

RPA (robotic process automation)

RPA software mimics a person clicking through screens: open this window, click this button, copy this field. It is useful for older systems with no other way in. It is also brittle, because when a screen layout changes the bot can break, and traditional RPA follows recorded steps without understanding what it is looking at. Agents can read and interpret content. When an agent has to work through a website's screens it can share some of that fragility, so a good build uses a proper connection wherever one exists.

AI agents for small business: examples by department

These are illustrative examples of work agents handle well. Every business is different, and the right first project comes from looking at your actual workflows, not from a list.

Front office and sales

• Replying to new inquiries within minutes, with a draft that uses your pricing rules and real calendar openings
• Qualifying leads by asking the questions your best salesperson would ask, then routing promising ones to a person
• Logging every conversation against the right contact so nobody has to update the CRM by hand

Operations

• Triaging a shared inbox: service requests to dispatch, billing questions to billing, complaints flagged to the owner
• Turning job notes and photos into a tidy summary for the customer file
• Checking that each new job has what it needs (signed estimate, deposit, address) before it gets scheduled

Finance and admin

• Reading supplier invoices, matching them to purchase orders and drafting the bill in your accounting software for approval
• Sending polite, escalating reminders on unpaid invoices and stopping the moment a payment lands
• Pulling numbers from several systems into the weekly report

Customer service

• Answering common questions (hours, service area, order status) from your own documents, and handing anything else to a person with the conversation summarized
• Drafting replies to online reviews for the owner to approve

What AI agents are bad at, and the real risks

Agents are useful, and they are also new technology with real failure modes. Anyone selling you an agent should be able to talk about these plainly.

Hallucination

Language models can produce confident answers that are wrong. An agent asked about a policy it was never given may invent one that sounds reasonable. The fix is not hope. It is giving the agent your actual documents, instructing it to say it doesn't know when the answer isn't there, and keeping a person in the loop for anything customer-facing until the output has earned trust.

Acting on bad data

An agent is only as good as what it reads. If your CRM is full of duplicate contacts, or the price list in the shared drive is out of date, the agent will act on those errors faster and more consistently than a person would. Cleaning up the source data is often the first job.

Over-permissioning

An agent with more access than it needs can do more damage when something goes wrong. That might be a bug, a misread email or a message written specifically to trick it. That last one is called prompt injection: text hidden in an email or document that tries to give the agent new instructions. Narrow permissions limit how far any mistake can spread.

Judgment calls

Agents are weak at decisions with no clear rules, high stakes and lots of unwritten context: whether to offer a discount to an upset long-time client, how to answer a legal threat, whether to take on a risky job. Those decisions should stay with people.

How a good build manages the risk

The risks above are manageable. They just need to be designed for from the first day, not patched in after something goes wrong.

• Human approval where it matters. The agent drafts; a person approves anything that goes to a customer, moves money or can't be undone. As trust builds, you can relax this for low-risk steps.
• Narrow scope. One agent, one clear job. An agent that triages the inbox doesn't also need to issue refunds.
• Exceptions flagged, not guessed. When something doesn't fit the rules (a missing field, an unusual amount, an angry customer), the agent stops and routes it to a person with the context attached.
• Logging. Every action is recorded: what the agent read, what it decided and what it did. When something looks off, you can see exactly why.
• Evaluation on real examples. Before launch, the agent is tested on a set of your own past emails, invoices or requests, and its output is compared with what your team actually did. Then it runs alongside your team before it runs on its own.

The internet is the agent's infrastructure

Here is the idea behind how we build. Traditional software runs on infrastructure: servers, databases and APIs. Your business runs on infrastructure too. It just looks different. It is the inbox, the CRM, the accounting system, the shared drive, the scheduling tool and the supplier portal your team logs into every day.

AI agents use the internet the way software uses infrastructure. APIs, inboxes and web apps are their servers. That means you usually don't need new software for an agent to be useful. It works inside the systems you already pay for, through official integrations where they exist and through email, exports or the web interface where they don't.

It also means the agent's work shows up where you already look. The draft is in your Gmail. The new contact is in your CRM. The bill is in QuickBooks, waiting for your approval.

How to get started with AI agents

You don't need an AI strategy document to start. You need one well-chosen workflow.

1. List the repetitive work. Write down the tasks your team does over and over, especially ones that involve reading something and then updating another system. Our guide to what to automate first walks through a simple scoring method.
2. Pick one task with clear rules and real volume. Inquiry replies, invoice entry and inbox triage are common starting points.
3. Decide where a person approves. Write down which steps the agent may do alone and which need a yes.
4. Check the data. Make sure the documents, price lists and records the agent will rely on are current.
5. Test on past examples, then run alongside your team. Switch over only when the output is consistently right.
6. Measure, then expand. Track time saved and errors caught, then choose the next workflow.

If your work involves confidential client files or regulated records, also think about where the AI runs. Some agents can use a model that runs on hardware in your office instead of a cloud service. Our private and local AI page explains how that works.

When you're ready for a second opinion, see how we design and build AI agents for small businesses, or book a free workflow audit and we'll look at your workflows with you.