AI agents for business: what they really do (and don't)

An AI agent for business is not a virtual employee you plug in to replace a role. It is a sequence of specific tasks, guided by rules, that reads, sorts and drafts for you within a narrow scope. Mixing the two up is expensive: either you expect an agent to do what it can’t, or you never automate anything because the picture of a total replacement scares you for no reason.

This guide starts from what an AI agent actually does in a small business, then covers what we put in place in practice and where to start if you want to try it yourself.

What the problem looks like

Take a business with ten to thirty employees and no AI agent. Every morning, someone opens the shared inbox and sorts the requests: a quote request, a document to match up, an overdue client follow-up, a question that has already been asked ten times.

That person reads every message, looks for the information across several different tools, copies what is needed into a spreadsheet or a management system, then replies or passes it on. On a repetitive task like this one, it often adds up to several hours a week per person, and it shows up nowhere because it gets done in between other things.

The problem is not the task itself, which is often simple. The problem is how often it comes back, and the constant interruption it causes to more useful work. A salesperson who has to sort their emails before they can sell, someone in accounts who has to sort documents before they can close the month: both lose the same kind of time. Not the five minutes of the task, but the twenty minutes of broken concentration around it.

That repetition is what an AI agent looks at, not the whole job. It replaces neither the salesperson nor the person in accounts. It does the sorting and the first read, and leaves the decision and the exceptions to the person.

This scenario repeats itself, in slightly different forms, in just about every department. A customer service team rereads and sorts requests that come back almost word for word from one week to the next. A recruitment team re-enters the same information from one application to the next before passing it up to the right manager. Stock management copies product references from one supplier into another system because the two tools don’t talk to each other.

None of these tasks is complex on its own. It is their number and how often they recur that weigh on a working week, far more than how hard they are.

What doing nothing costs

Before talking about tools, measure the weekly time spent on this task and the number of people involved. To illustrate the calculation, I use two fixed assumptions: €35 per hour of fully loaded cost and 47 working weeks a year. These figures are not a verified industry average; adjust them to your real costs for an estimate that fits your business.

I then apply a theoretically recoverable share of 60 to 75% of the time, and a cap of 15% of revenue. These are modelling choices, not measured gains, and not a promise of return. Checks, exceptions, rework and set-up costs still have to be priced in.

This is exactly the method I use when I size a project with you, starting from your own figures rather than an industry average. The result is not a promised gain, it is an estimate, and the same inputs always give the same amount.

I also look at a second indicator: the share of your tasks that is already automated, and the time your team spends on high-value work that still lacks a tool. The lower that share, the more room there is to improve, and that is often where the real opportunity lies, not just in the time lost on a single isolated task.

What we put in place

In practice, an AI agent for business is built as a small chain, not as a piece of magic software. First a trigger: an incoming email, a file dropped into a folder, a box ticked in a CRM. Then a rule: what the agent must read, which information to extract, and in what format to output it.

A human check comes in where it matters: before an email goes out to a client, before an amount is approved, before a sensitive piece of data is changed. Finally there is an output: a draft ready for sign-off, a row added to a spreadsheet, a notification sent to the right person.

In practice, these steps are often assembled with an automation tool such as n8n: I walk through installing n8n step by step in a separate guide. It orchestrates the trigger and the rule, and builds the human check in as a step of the flow, not as a fix after the fact. The AI agent itself only does one thing at a time: read a document, classify a request, draft a reply. It is the chaining of several simple, checkable steps that, from a distance, gives the impression of an intelligent system.

What breaks, and how to avoid it

Three pitfalls come up on almost every automation project, and all three can be seen coming.

Where to start if you do it yourself

If you want to try this before calling anyone, here is how to move forward without spending a month on it.

1. Pick a single task. Not “automate customer service”, but “sort the quote request emails that land in one specific inbox”.
2. Measure the real time for a week. Note who does what, how many times, and how long it really takes, not a rough guess.
3. Write the rule in plain English before you open a tool. If you can’t explain in five sentences what the agent has to do, it won’t be able to do it either.
4. Add the human check from the very first test. A sign-off before sending, even a manual one, while you check that the agent understands the task properly.
5. Look at custom software if off-the-shelf tools don’t fit. Custom software takes longer to build, but it adapts to the way you work rather than the other way round.

If you get to the end of this list and decide not to go ahead, that is not a failure: it is already a clear diagnosis of what you are missing to move forward on your own.

Frequently asked questions

Can an AI agent replace a role in my business?

No, not as things stand. An AI agent automates a sequence of specific, documented tasks, not the whole of a role that mixes judgement, client relationships and exceptions. It frees up time on the repetitive part, which can change what a role involves, but rarely removes it entirely.

How long before I see a result?

Allow a few weeks to scope a single task, document it and connect a first agent to it. The time saved becomes visible as soon as the task runs without routine rework, but the initial scoping cannot be improvised in a day.

Do I need an IT department to get started?

No, not for a first, clearly defined task. Most of today’s automation tools can be run without knowing how to code. What takes experience is the scoping: choosing the right task, writing the rule and putting the human check in the right place.

Can an AI agent make mistakes?

Yes, like any automated system. That is precisely why the human check stays a step of the flow and not an option. An error caught before sending costs one review. An error sent straight to the client costs a lot more.

How is it different from an existing off-the-shelf tool?

An off-the-shelf tool applies general logic to everyone. An AI agent built for your business applies your own rules, to your own documents, with your own exceptions. The difference shows most on the special cases a standard tool ignores.