"AI agents" are everywhere in the news: software that reads, decides and acts on its own. For a small business, the promise is appealing. Fewer repetitive tasks, faster replies to customers, less time spent copying information between systems.
The reality is more modest, and still useful. Agents today are good at some things and unreliable at others. The businesses that get value from them pick narrow tasks, keep a human in the loop and measure the results. Here is how I think about it.
What an agent actually is
An agent is a language model connected to tools. Instead of only answering a question, it can take steps: look up an order in your system, read an email, fill in a form, draft a reply, create a task. It decides which tool to use next based on what it has seen so far.
That flexibility is what makes agents powerful and also what makes them risky. A chatbot that gives a wrong answer is a problem. An agent that takes a wrong action, such as cancelling the wrong booking or emailing the wrong customer, is a bigger one.
Where agents work well today
The best first tasks share a few traits: they are repetitive, the inputs are text or documents, mistakes are easy to spot, and a human can review the output before anything irreversible happens.
Triage and routing. Reading incoming emails or messages, classifying them (booking request, complaint, quote request, spam) and routing them to the right person or queue, with a suggested reply.
Drafting. Preparing replies, quotes, summaries or follow-ups for a human to review and send. The human stays the sender; the agent saves the typing.
Extracting data from documents. Reading invoices, forms or delivery notes and filling structured fields, with a person checking anything uncertain.
Answering questions from your own information. Answering staff or customer questions using your policies, price lists and FAQs, with links to the source, and handing over to a person when it is not sure.
Internal lookups. "What is the status of job 1042?" or "Which bookings tomorrow have no assigned technician?" answered from your own system.
Where agents are still unreliable
- Long, open-ended tasks with many steps and no checkpoints. Errors compound.
- Anything involving money without human approval: refunds, payments, purchases.
- Actions that cannot be undone: deleting data, sending messages to customers automatically, changing contracts.
- Tasks that need judgment about people: hiring decisions, disputes, sensitive complaints.
- Situations where wrong but confident answers are costly: legal, medical or safety advice.
None of this means "never". It means these areas need tight limits, strong review and a lot of testing before automation.
Keeping a human in control
The design pattern that works best for small businesses is simple: the agent proposes, a person approves.
- The agent prepares a draft reply; a staff member reads it and clicks send.
- The agent suggests a booking change; the manager approves it.
- The agent extracts invoice data; an accountant reviews flagged fields.
Over time, as you measure accuracy, you can let the agent act alone on the safest, most predictable cases, and keep review for the rest.
Some other safeguards:
- Least privilege. Give the agent access only to the tools and data it needs. Read access before write access.
- Logs. Record every action the agent takes and why, so you can review and correct.
- Limits. Rate limits, spending limits and clear lists of allowed actions.
- Protection against manipulation. Any text the agent reads, such as a customer email, can contain instructions meant to trick it. See prompt injection risks in LLM apps.
Measuring whether it works
Before building anything, decide what success looks like:
- Time saved per week on the task.
- Percentage of drafts approved without changes.
- Errors caught in review.
- Response time to customers.
Then test with real examples from your business before rolling it out. A small set of past emails or documents, with the correct outcome noted, is enough to start. I cover this in evaluating LLM output.
How to start without a large project
- Pick one task. The one that is most repetitive and least risky. Email triage is a common choice.
- Collect examples. Twenty to fifty real cases with the right outcome.
- Build the simplest version. Often a model with a good prompt and one or two tools, not a complex multi-agent system.
- Run it in shadow mode. The agent suggests; people work as usual and compare.
- Measure, then decide. If the suggestions are good enough, move to "agent proposes, person approves". If not, adjust or stop.
This approach costs far less than a big AI project and gives you real evidence about whether the task is a good fit.
A worked example
Imagine a service company that receives quote requests by email and through its website. Today, someone reads each request, asks for missing details, estimates a price range from a price list and replies. An agent could read the request, check which details are missing (address, size, preferred date), look up the relevant services in the price list, and prepare a draft reply asking for the missing information or proposing a range. A staff member opens the queue, reads each draft, edits where needed and sends. Nothing reaches a customer without a person's approval, and the team can see within a few weeks how many drafts needed changes.
Off-the-shelf or custom?
Many tools now offer built-in AI features: email clients that draft replies, helpdesks that classify tickets, accounting software that reads invoices. If one of them covers your task well, use it.
A custom agent makes sense when:
- The task depends on your own system or data that off-the-shelf tools cannot reach.
- You need it inside your own app for customers or staff.
- You need control over where data goes and how it is stored.
If you already have a product, adding AI features to it is often the most valuable route. See adding LLM features to your product.
Privacy and data
Before sending customer data to any AI provider, check:
- What data is sent, and whether all of it is needed.
- Whether the provider uses it for training, and under which settings.
- Where it is stored and for how long.
- What your customers have been told.
Remove personal details that the task does not need. It is easier to keep data private than to explain later why it was not.
Realistic expectations
A well-chosen agent can take hours of repetitive work off your team every week and make responses faster and more consistent. It will not run your business, and it will make mistakes. Plan for review, measure honestly, and grow the scope only when the results earn it.
If you have a task in mind and want to know whether it is a good fit for an agent, read about me, see my work, and get in touch to set up a call.