AI Agents for Small Businesses: Where to Keep a Person

An AI agent is software that gets a goal instead of a fixed list of steps, decides for itself which steps to take, and uses the tools you connect to it (your inbox, your CRM, your calendar, a search engine) to get there. For a small business that is useful for work that is repetitive but not identical every time: answering routine enquiries, preparing follow-ups, collecting information for a report. It is risky wherever a wrong step costs money, damages a client relationship or cannot be undone, and that is exactly where a person should keep the final yes.

AI is powerful, but not all-powerful, and the word "agent" is used so loosely in marketing right now that it is worth slowing down before buying one.

What is the difference between an AI agent and an automation?

Anthropic, the company behind Claude, draws the line in its guide on building agents. A workflow is a system where AI and tools are orchestrated through predefined code paths: the steps are fixed, and AI does one part, for example writing a draft. An agent is a system where the AI dynamically directs its own process and tool use: it decides what to do next.

Some products marketed as "agents" use predefined workflows. Ask which steps are fixed, what the AI can decide and what requires approval. For a well-defined task, a workflow can make the steps easier to inspect and limit unnecessary model calls: an enquiry arrives, AI sorts it and writes a draft, the CRM gets a record, a person gets a task. The same guide advises finding the simplest solution possible and only adding complexity when it is needed. I agree, and I would add a business reason: you can explain a workflow to your team in two minutes, and people trust what they understand.

Where do AI agents help a small business?

The good candidates share three traits: they happen often, they follow a pattern with some variation, and a mistake can be caught before it hurts.

• New enquiries. Reading an enquiry from a form, email or messenger, pulling out what the customer needs, creating the CRM record and drafting the first reply. A person checks anything unusual.
• Follow-ups. Noticing that a quote has been open for a week without an answer and preparing a polite follow-up for the salesperson to send.
• Research and preparation. Collecting public information about a prospect before a call, or summarising a long email thread before someone picks it up.
• Internal reporting. Pulling numbers from the tools you already use into one weekly page, so you decide on data rather than on the feeling of the week.

What these have in common is that the agent prepares and a person decides. The time saved comes from the preparation, which is where most of the hours go anyway.

Where should a person stay in the loop?

Ask yourself what happens if the agent gets it wrong, and whether you would notice. If the answer is "the customer would notice first", a person approves before anything goes out. That covers:

• anything sent to customers, suppliers or the public;
• anything involving money: discounts, refunds, payments, quotes;
• deleting or changing records;
• decisions about people, such as hiring or performance.

There is a technical reason too. Agents read text from outside your company, and that text can contain instructions. OWASP, the security community behind the best-known list of risks for AI applications, puts prompt injection first in its 2025 list and names two of the defences directly: give the model only the access it needs for its job, and require a person's approval for privileged operations. An agent that can read your inbox should not also be able to send money.

You would not be unusual in keeping that control. In the UK government's 2025 survey of 3,500 businesses, 84% of those using AI kept at least some human input or checking on its results, and agentic AI was the least adopted technology among them, at 7%.

What goes wrong with AI agents in a small business?

Three problems are worth planning for from the start.

The first is an agent placed on top of a process nobody has written down. The agent then copies the chaos faster. Before you add AI, write down what should happen today when an enquiry arrives, and who does what.

The second is speeding up one step and choking the next. If an agent doubles the number of qualified enquiries and your sales team still works the same way, you have moved the queue from the inbox to the salespeople. Plan the step after the agent at the same time.

The third is cost nobody controls. Agents use AI services that are billed by usage, and an agent stuck in a loop can run up a bill quietly. Set spending limits with the provider, give one person ownership of the account and look at usage weekly at first. Controlling AI costs is necessary, and it is entirely possible.

From our own system I run my own work on a system of AI agents built on Claude, MCP, n8n and my own servers: research, documents, reports and email drafts. Each agent has only the access its task needs, keys are kept in a separate vault, and a person approves the templates, rules and company knowledge they work from and checks samples of what goes out. On this site I am setting up the assistant the same way: it will say that it is an AI, answer only from an approved knowledge base and the public price list, take down enquiries, and pass anything non-standard to me as a draft that I send myself. It is being set up now; I describe the design, not results.

How to start

Pick one process from the list above, write down how it works today, and build the simplest version: AI prepares, a person approves. Measure for a few weeks, then decide whether to give it more freedom. The 30-day plan for implementing AI walks through this step by step, and the AI acceptable use policy template helps you set the rules first. If your first candidate is customer email, read AI email replies for customer service. More on first steps is in the getting started with AI guide, and if you are unsure whether you need outside help, see what an AI consultant does, and when you do not need one.

A sensible first project for most small businesses is new enquiries: getting every one into the CRM with an instant acknowledgement and a task for the right person. That is what my lead system service sets up.

Talk it through

If you are not sure whether you need an agent, a simple automation or nothing at all, we can sort that out in a free 30-minute call. Afterwards you get a short written outline with up to three tasks worth starting with.

Calls are in English.

Free 30-min call with Lidia

Sources

• Building effective agents - Anthropic (19 December 2024), accessed 18 Oct 2026.
• LLM01:2025 Prompt Injection - OWASP Top 10 for LLM Applications, accessed 18 Oct 2026.
• AI adoption research - UK Department for Science, Innovation and Technology (GOV.UK, updated 13 February 2026), accessed 18 Oct 2026.