AI agents for business: what they do and which tasks you can give them

A supplier’s email lands at eight in the morning with a delivery note attached as a PDF. Someone opens it, copies four fields into the management software and files the message. Again at quarter past eight, and so on for half the morning. When an SME talks about AI agents for business, it’s almost always talking about tasks like this, and it’s worth being clear about what an agent is before checking whether any of your work fits.

How an agent differs from a chatbot or an n8n workflow

A chatbot answers what you ask it and then waits for the next question. A classic automation, the kind you build with Zapier or an n8n workflow, runs a sequence someone mapped out beforehand: this comes in, do that, then the other thing. An AI agent works differently. You give it a goal and a handful of tools it can call, and it decides which steps to take and in what order, when to look something up and when to stop because it thinks it’s done.

For whoever is paying, the difference that matters shows up when the input changes. A fixed workflow breaks if the supplier sends the delivery note in a different format or the customer words their query back to front; someone has to go in and fix it. An agent copes with that kind of variation fairly well, but the trade-off is that it can get things wrong in a different way every time. That’s why we don’t sign off on any agent until it has run for a few weeks with written limits, a log of what it does and someone keeping an eye on it.

And do you need to know how to code to build one? Mostly, no. n8n, Make and the tools that now come built into Claude or ChatGPT let you put together a simple agent without writing code, and some people do it on their own. What none of them spare you is the slow part: deciding what the agent can and can’t touch, giving it access to the data it needs and checking its decisions before letting it loose.

Email, invoices, leads: the tasks where they hold up today

The agents that hold up in production at a mid-sized company share an unglamorous trait: they handle admin work that involves text, high volume and reversible consequences if something goes wrong. On that ground, they work.

• Sorting support email and drafting replies to first-line queries, leaving a person to approve them.
• Reading invoices and delivery notes that arrive as PDFs, extracting the fields and entering them into the management software.
• Qualifying the leads that come in through the form, looking up public information about the company and flagging the promising ones to sales.
• Drafting product copy or replies to quote requests based on technical data sheets.
• Reviewing long documents and pulling out what matters: the terms of a supplier contract, the changes between two versions of a tender specification.

Beyond that, things get complicated. Anything that reaches the customer unfiltered, pricing, anything with legal or accounting consequences and, in general, decisions where a costly mistake can go unnoticed for weeks: there, we still put a person in front, and not out of token caution. The cost isn’t where you’d expect either. The tool subscription is usually the small part, and what eats up the budget is defining the process properly and keeping it alive as it changes.

Four filters to tell whether one of your tasks is a candidate

Go back to the delivery note from the start. It arrives every morning, always looking the same, and that’s the first condition: the task has to repeat genuinely and in the same shape. Being tedious isn’t enough. The agent relies on the pattern, not on the heroic one-off that happened once back in March.

The data has to be somewhere the agent can reach. If the information lives in email, a shared spreadsheet or a CRM with an API, we’re in good shape. If it lives in the warehouse manager’s head, in a filing cabinet or in a crooked scanned PDF, the project doesn’t start with the agent; it starts with putting that data somewhere readable.

Try explaining the task to someone who joined the company tomorrow. If in half an hour you can tell them what they look at, what they decide and what they do in the three odd cases, you have a rule you can explain, and the agent can follow it. If you end up saying it depends, you’ll see, you’ll get the hang of it over time, the task isn’t ready to automate yet, and this is the filter that knocks out the most candidates for us.

An agent has to be built, tested and adjusted again every time the supplier changes the delivery note format. A task that comes up twice a month doesn’t pay for that maintenance; the maintenance swallows it whole, and below a certain volume it’s cheaper to keep doing it by hand.

If you have a task that passes all four and want an opinion before putting money into it, tell us about it and we’ll look at it with you.