"We tried AI, and nothing came of it." This sentence comes up so often that it deserves a closer look. In almost every case, the test consisted of opening a chat interface, asking a few questions, noticing that the answers were generic, and concluding that the tool wasn't suited to the business.

The conclusion is right, but the cause is misidentified. A chat assistant is built to answer a question. It has neither your data, nor your context, nor a mission to accomplish. Comparing it to a colleague is like judging an employee on their first day, without having given them access to anything.

What defines an agent

  • A defined mission. Not "help the team", but "qualify every incoming request and suggest a reply in under ten minutes".
  • Access to the material. Your documents, your catalog, your rules, your history. Without material, all you get is style.
  • Tools. Reading a database, writing to a file, sending a message, creating a task. An agent that can't do anything does nothing.
  • A success criterion. What you expect, and how you check that it's right. Without a criterion, there's no way to know whether the agent is useful.
An agent with no mission and no data produces plausible text. That is exactly what nobody needs.

Why specialization changes the result

A generalist agent is mediocre at everything. An agent specialized in one precise task (reviewing a quote, preparing a tender response, sorting requests, checking the consistency of a catalog) is judged on a single thing, and so it can be tuned.

Specialization brings an additional, less obvious benefit: it makes errors visible. When the mission is narrow, you spot right away what's wrong and fix the instruction. When it's vague, all you notice is a general dissatisfaction, with no idea what to adjust.

What an agent must not do

An agent does not decide for you on anything that commits the company. It prepares, suggests, sorts, writes a first draft. Validation stays human wherever a mistake is costly: a price, a contractual commitment, a message to an unhappy customer.

This boundary is not a moral precaution, it is a rule of efficiency. The setups that last are those where the human keeps the decision and delegates the preparation. The ones that fail are those that tried to skip the review to save five minutes.

The real cost, and where it lies

The cost of an agent is not in the model used. It is in the setup: describing the mission, giving access to the right material, defining what a good result looks like. This is organizational work, not IT work. The companies that succeed with AI are the ones that agreed to do that work.

The five-line test

Pick a repetitive task, write down in five lines what a good result means, provide the necessary material, and run it on real cases for two weeks. You will then know whether AI serves your business, instead of drawing conclusions from a ten-minute conversation.

AI doesn't replace a profession. It replaces the part of the job that nobody ever wanted to do.