← Journal· AI · Aug 4, 2026 · 4 min read

Practical AI agents for business: what actually works

Most companies asking us for an AI agent want something simpler and more reliable. Here is how we tell the two apart, and what it costs to run the real thing.

Every second enquiry we get now contains the word agent. Most of the time, what the person actually wants is a script with good judgment attached, and it does not need to be an agent at all. That distinction is worth getting right before you spend anything, because the two things fail in completely different ways and only one of them fails quietly.

What an agent is genuinely good at

An agent is a language model that decides what to do next and calls tools to do it. The valuable part is not the model, it is the loop: try something, look at the result, adjust. You are paying for that loop, so it only makes sense when the path through a task changes every time. Handling an inbound support ticket that might touch billing, shipping or account settings is a good fit, because you cannot write the branches in advance. Producing a weekly report from three fixed queries is not, and an agent will do it slower, more expensively and less reliably than fifty lines of ordinary code.

The agent projects we have seen survive all share a shape. The task is narrow, the tool list is short, and a person sees the output before a customer does. One client runs an agent over inbound quote requests. It reads the email, pulls matching items from their catalogue, drafts a quote, and then stops. A salesperson approves or edits it. That final stop is the main reason the thing is still in use a year later rather than switched off after the novelty wore through.

The part nobody budgets for

Getting the first working version of an agent takes days now. Making it trustworthy takes months, and hardly any of that time goes into prompts. It goes into unglamorous scaffolding: what the agent is permitted to touch, what happens when a tool call fails halfway, how you find out that it did something strange at three in the morning on a Tuesday. Teams that skip this ship a demo that impresses everyone in the room, then watch it quietly fall out of use six weeks later because nobody trusts it unsupervised and supervising it costs more than doing the work.

Cost is the other thing people misjudge. An agent that reasons over a long document on every run can be twenty times the price per task of a single model call, and you will not notice until the volume arrives. We now estimate cost per task before writing any code, and more than once that number has ended the conversation early, which is the correct outcome.

If you are considering one, here is what we look for before agreeing it is the right shape of solution:

The steps genuinely vary between runs. If you can draw the flowchart, build the flowchart instead.
A small set of tools, ideally under ten. Every extra tool makes the agent's choices measurably worse.
A human checkpoint in front of anything that reaches a customer, a contract or a bank account.
Full traces of every step, stored and searchable. When it goes wrong you need the path it took, not just the answer.
A set of real past cases collected before you build, which you rerun after every change to see what you broke.

That last point is the one clients push back on most, and it is the one that decides whether the project works. Without a fixed set of examples to test against, every change to an agent is a guess. You improve one behaviour, break two others, and have no way of knowing until a customer tells you. Thirty real cases with expected outcomes, written down before a line of code, is a cheap thing that pays for itself in the first month of changes.

If you can draw the flowchart, build the flowchart. Agents are for the work you cannot draw.

We build agents and we still talk clients out of them regularly, usually into a smaller system that does one useful thing well: classify the message, extract the fields, draft the reply. Those are unfashionable and they run for years without anyone thinking about them. Start there, measure what it saves, and add autonomy only where the variety of the work actually demands it. The companies getting real value from this are not the ones who built the most ambitious agent. They are the ones who picked a task small enough to verify.