AI agents: from chatbot to a colleague that finishes the job

Two years ago every AI conversation was about chatting. In 2026 it is about doing. The term you hear everywhere is agentic AI, or simply AI agents. The difference with a chatbot is bigger than it looks, and for smaller companies that difference is the interesting part.

The difference in one paragraph

Classic automation follows rules you defined in advance: if this happens, do that. A chatbot holds a conversation: question in, answer out. An AI agent sits between and above those. It is given a goal, looks at the situation, works out which steps are needed and carries them out until the task is done.

A practical example. A chatbot says: your quote has been sent. An AI agent notices the quote has been open for five days, sends a friendly reminder, processes the reply, books a call in the calendar and updates the status in your system to followed up.

Why this only works now

The models are finally good enough to run several steps in a row reliably and to handle exceptions. Gartner expects that by the end of 2026 around 40 percent of business applications will include a task-specific agent, up from less than 5 percent in 2025. By 2028 a third of enterprise software would include agentic AI, up from under 1 percent in 2024.

For smaller businesses the maths is appealing. McKinsey estimates that current agent technology can automate roughly 45 percent of activities in smaller companies, against roughly 30 percent with traditional automation. That gap sits mostly in work that was too messy for fixed rules.

Which processes suit this?

In smaller businesses we keep seeing the same five.

• Chasing quotes: reminding, answering questions, booking the call, updating the status.
• Sorting inbound requests: pulling them from the inbox, categorising, routing to the right person.
• Processing invoices and receipts: reading them, booking them, flagging anomalies.
• Handling complaints: acknowledging, gathering information, proposing a fix, escalating when needed.
• Recurring reports: pulling data, summarising, in your inbox every Monday.

The common thread: work with volume, with a recognisable pattern, where the cost of a small mistake is limited. That is the zone where an agent is already reliable.

Where do you start?

Not with the most complicated process, and not with everything at once. Pick one process that meets three conditions: it happens often, it follows a pattern, and you can measure whether it improves. Then you will know within a month whether it works.

A good test: which chore costs you or a colleague hours every week and gives nobody any satisfaction? That is almost always the best first candidate.

The biggest gain is rarely in the most exciting process. It is in the boring work that comes back every week and that nobody will miss.

Where it goes wrong

There are three ways to get this wrong, and we see all three regularly.

• No boundaries. An agent that spends money, sends contracts or promises things to customers should have a human approval step.
• Automating what should be abolished. Fix the messy process first, automate it second.
• Measuring nothing. Without a baseline you still will not know after three months whether it delivered anything.

Does this replace jobs?

In smaller companies it usually plays out differently than the headlines suggest. In a small business the owner or a small team does everything, from quotes to admin. What an agent takes over is usually not somebody's job, but the part of the week nobody got round to. In practice we more often see work being followed up properly than someone becoming redundant.

What does it cost?

That depends heavily on how many systems need to be connected. A single process with one or two integrations is well within reach for most small businesses. The honest advice: first work out how many hours the process costs today, because that determines whether it is worth doing. Without that number any price is a guess.