AI agents for business, explained for an owner

The office manager forwards a supplier's email to the bookkeeper with the note "can you check this against the order?" That sentence is a task: read the mail, find the order, compare, report back. An AI agent is a program that takes a sentence like that and does the work, using the same systems the bookkeeper would open.

Chatbot, assistant, agent

A chatbot answers a question from what you type into it. An assistant does the same with access to your documents. An agent goes one step further: it takes a task, plans the steps, uses tools along the way (your mail, your CRM, your accounting program, a browser), checks its own result, and writes the outcome back where it belongs. You hand it work the way you hand work to a new colleague, with a brief and a check at the end.

The technical ingredients are a language model (the part that reads and writes), a set of tools the model may call, a memory it can read from (in our builds, the knowledge vault), and rules about what it may do alone and what needs a person. The last part is the one that matters most to an owner.

What agents do reliably today

The work that pays off first has a pattern: it repeats, it follows rules a person can write down, and a mistake is easy to spot before it reaches a customer. In firms with 5 to 50 people that usually means:

• Reading incoming mail, sorting it, and drafting the reply from your documents for a person to send.
• Drafting quotes from a price list and a short brief.
• Reading supplier invoices, matching them to orders, and booking the ones that agree.
• Answering a missed call by text and booking the callback.
• Keeping the CRM current from call notes and mail threads.
• Pulling the weekly numbers into a report that a person then edits.

Each of these is one agent with one job. The most common mistake we see is one agent asked to do everything, which produces a colleague who does many things badly.

Where they still fail

Tasks with no fixed shape, where every case is a judgment call, are poor first projects. Long chains of steps, where an early misread compounds. Anything where the agent has to recognise that it is out of its depth: models are getting better at saying "I am not sure", and they are not there yet. And any workflow whose underlying data is a mess: an agent matching invoices to purchase orders that do not exist will fail loudly, which is at least honest.

The gate

Every step that touches a customer or money waits for a person at first. The agent drafts, a person approves; the agent proposes a booking, a person releases it. Each decision is logged. After a few weeks the log shows which steps the agent gets right every time, and you loosen the gate there, one step at a time. Owners tend to loosen it faster than they expected, because the log gives them something they never had with a human hire: every decision, written down, with the reason.

Where the agent's knowledge comes from

A general model knows the world and nothing about your firm. What makes an agent useful is the material it works from: your price list, your past quotes, your house rules, the way you write to clients. In our builds that material lives in one indexed place, the knowledge vault, and every agent starts there. An agent without it is a stranger every morning.

How to start

Pick the task that repeats most and hurts most. Write down, on one page, how a careful new employee would do it and where they would ask for help. That page is the agent's brief and the shape of the gate. Build that one agent, run it on real work for a month with a person checking, then decide about the second. Firms that start with ten agents at once rarely get one of them into daily use.

Joinery builds agents this way: one workflow first, in your tools, with the gate where money or customers are touched, handed over as yours. Workflow automation is the usual first step; a company of agents is where it can lead.