01What can an AI agent actually do for a small business?
The agents that earn their keep sit on a short, repeated task with clear sources. Examples that fit most small businesses:
Each of these has a source the agent can read and a result someone can check. That is what makes them safe to start with.
- Answer questions on your website or WhatsApp from your own pages and policies: delivery areas, opening hours, what a service includes.
- Qualify a lead. The agent asks three or four questions, writes the answers into your CRM and flags the ones worth a call today.
- Look up an order. A teammate or customer asks for a status, and the agent reads it from Shopify or your order system instead of guessing.
- Draft replies to common emails, for a person to check and send.
- Summarise a long thread, a call transcript or a set of reviews into the points someone needs to act on.
02AI agent vs chatbot vs automation workflow
These three get mixed up, and the difference decides what you should build.
A simple rule helps. If the task is the same fixed sequence every time, build a workflow. If the task involves reading language and choosing between a few actions, an agent can help. Many projects need both: the workflow moves the record, the agent writes the reply. Our n8n lead automation guide shows a workflow with one AI step in it.
| Rule-based chatbot | Automation workflow | AI agent |
|---|
| How it decides | Fixed menu or keyword rules | Fixed steps you designed, such as "when a form arrives, add a row" | A language model reads the request and chooses from the actions you allow |
| Handles messy questions | Poorly | It does not read language unless you add an AI step | Yes, within its sources |
| Takes actions in your tools | Rarely | Yes, every time, the same way | Yes, the ones you scope |
| Predictability | High | High | Lower. It needs testing and limits |
| Typical tools | Widget builders | n8n, Zapier, Make | A model plus your data and tools |
03Where do AI agents go wrong?
Most failures come from giving an agent too much freedom or too little information.
These are the same rules we apply to the agents we build, set out on our AI agents page.
- Inventing answers. If the source does not say it, a badly set up agent may make something up: a price, a delivery date, a refund rule. The agent should say it does not know and hand over to a person.
- Acting without a check. An agent that can issue refunds or change orders needs a confirmation step. By default, an agent should read and draft, and a person should approve anything that moves money.
- Stale sources. If your policy page changed last month and the agent still reads the old one, it gives last month's answer.
- No way out. A customer who wants a person should reach one quickly, with the conversation attached.
- Unclear data rules. You need to know which model provider is used, what is stored, and for how long. That has to match your privacy policy.
04How do you scope a first agent?
Pick one task, then write it down before anyone writes code.
Then build, and test with questions your team did not write. Expect to adjust the instructions and sources after the first week of real use.
- Name the task in one sentence. For example: "Answer delivery and returns questions on WhatsApp for the India store."
- List the sources the agent may read: specific pages, a policy document, an order lookup.
- List the actions it may take, and the ones it must never take.
- Write ten real questions, with good and bad answers. Use your actual customer messages.
- Decide the handover: who gets the chat, through which channel, and how fast.
- Decide what gets logged and how long it is kept.
05What does an agent need from your business?
- Accurate source content. If your website does not answer a question, the agent cannot either. Fixing the page often helps customers more than the agent does.
- Access to the tools it needs, with the narrowest permissions that work.
- An owner on your side who reviews conversations and updates sources.
- A privacy notice that covers the chat, and a decision on where transcripts are stored. UK clients usually ask about storage first. Indian clients usually start with WhatsApp as the channel.
06Should a small business build or buy?
Off-the-shelf tools are a good start for a website FAQ bot, especially if your content is in good shape. A custom build makes sense when the agent must read your own systems, such as Shopify orders, a CRM or a booking tool, or must take actions inside them with your rules. A custom agent also lets you change the model provider later instead of being locked to one.
Either way, start small. One agent on one task, measured against the time it saves or the leads it qualifies, tells you more than a plan for ten agents.
Decide the measure before launch. For a support agent, count the questions it answered without a handover and read a sample of those answers every week. For a lead agent, compare how many qualified leads reach a person, and how fast, with the month before. If the numbers do not move after a few weeks of tuning, change the task or switch the agent off.
07Plan your first agent with us
If you have a task in mind, book a project call and bring ten real messages your team answers every week. We will tell you whether it needs an agent, a workflow or both. See AI agents and business automation, or the UK automation page.
Questions
What is an AI agent in simple words?
An AI agent is software that understands a request in plain language, looks up information in the tools you connect, and takes an action you have allowed, such as replying or updating a record.
What is the difference between an AI agent and a chatbot?
A rule-based chatbot follows fixed menus or keywords. An AI agent reads free-form questions, uses your own sources, and can take scoped actions in your tools.
Is an AI agent the same as automation?
No. An automation workflow runs the same fixed steps every time. An agent decides between actions based on language. Many projects use both together.
Can an AI agent handle refunds or orders?
Only if that action is scoped and has a confirmation step. By default, an agent should read and draft, and a person should approve anything that moves money.
How do I stop an AI agent from making things up?
Limit it to named sources, tell it to say when it does not know, test it with real questions, and give customers a quick route to a person.
Talk through the project
Bring the current site, the deadline and the parts that have to work on day one. The call is for scope, not a scripted pitch.
Book a call · Pricing · Process · See all our projects