Guide · AI agents
What is an AI agent, and what can it really do for a small business?
By Dev Patel, founder of AgeDash. Published . 6 minute read.
An AI agent is a program that uses a language model to understand a request, decide on the next step, and carry it out with tools it has been given, such as a calendar, a CRM or a spreadsheet. For a small business, the useful agents today are narrow: they do one job, such as booking visits or reading invoices, with checks around them and a person to hand over to.
The three parts of an agent
- A model that reads and writes language. It turns "can I come Saturday morning" into a service, a day and a time, and writes the reply.
- Tools it is allowed to use: check the calendar, create a contact, read a file, send a message. Each tool is a function with clear inputs and outputs.
- A loop with rules. The agent looks at the request, picks a tool, looks at the result, and decides whether it is done or needs another step. The rules say what it may never do and when to stop and ask a person.
Take away the tools and you have a chatbot. Take away the model and you have a normal automation with fixed steps. The agent sits in between: fixed tools, flexible order.
What agents do well today
| Job | Why it works | Example |
|---|---|---|
| Booking in chat | Clear goal, few tools, easy to check | WhatsApp booking |
| Reading documents | Models read messy layouts; code checks the numbers | Invoice processing |
| Answering from your documents | Answers come from your text, with the source attached | AI WhatsApp assistant |
| First drafts | A person reviews before anything is sent | Replies, summaries, CRM notes |
| Sorting and routing | Classifying a message is easier than solving it | Which team, how urgent, which customer |
Where agents fail
- They can be confidently wrong. A model may misread a number or invent a detail that sounds right. Anything that matters must be checked by code or a person.
- They follow instructions hidden in text. A customer message or a document can contain words like "ignore your rules". Agents that read outside text must have limited tools, so even a fooled agent cannot do harm.
- Open-ended jobs drift. "Handle customer service" is too broad. "Answer order status questions from the order system and hand over everything else" works.
- Long chains of steps multiply errors. An agent that is right most of the time at each of ten steps is wrong at the end far more often. Keep chains short, and check between steps.
- Costs scale with text. Reading long documents and long histories on every message gets expensive. Good design sends the model only what it needs.
How to tell a real agent from a demo
A demo shows the happy path. Before you buy or build, ask these questions:
- What happens when it is unsure? A good answer names a handover, with context, to a specific person.
- What can it never do? There should be a written list, enforced by which tools it has, not just by its instructions.
- How are its outputs checked? Look for checks in code, such as totals, dates and IDs.
- Where are the logs? You should be able to see every decision and tool call afterwards.
- What does it cost per conversation or per document, at your volume?
- How was it tested? Ask for results on your own real examples, including the awkward ones.
Agent or simple automation?
Use a simple automation when the input is already structured: a form, an order, a row in a sheet. It is cheaper, faster and more predictable. Use an agent when the input is free text, a document or speech, or when the right next step depends on what the person said. Many good systems are mostly simple automation with one agent step in the middle, where the reading happens.
How to start
- Pick one repeated job where the input is messy but the right result is clear.
- Collect twenty real examples, including bad ones.
- Write the tools the agent needs and the rules it must follow.
- Build it as a workflow with checks after the model, and test it on your examples.
- Launch with a person reviewing results, and reduce review as it proves itself.
Our guide to choosing what to automate first helps with the first step, and our AI agents service covers how we build them.
Common questions
Is ChatGPT an AI agent?
ChatGPT on its own is a chat assistant. It becomes agent-like when it is given tools, such as browsing or running code. A business agent is the same idea, with tools limited to your systems and rules for your job.
Can an AI agent run my whole business?
No. Today’s agents are reliable for narrow, well-defined jobs with checks around them. Decisions with real risk, such as payments, pricing and hiring, should stay with people.
Do I need my own AI model?
Almost never. Small businesses use models from providers such as OpenAI, Anthropic and Google through their APIs. Your own data is given to the model at the moment it is needed, not trained into it.
Is it safe to give an agent access to my CRM?
Give it only the actions it needs, such as create a contact or add a note, and not delete or export. Log every action. Then even a mistake or a hostile message cannot do much harm.