Last week the owner of a manufacturing company sent me a PDF she paid 1 200 EUR for. Thirty pages on which agents they should deploy.
Last week the owner of a manufacturing company sent me a PDF she paid 1 200 EUR for. Thirty pages on which agents they should deploy.
Messy formatting. Text that was obviously written by a model. Headings that don't follow on from each other. In my view the authors didn't read it themselves before sending it. That's my opinion, not a fact, but nothing in the document refutes it.
Worse than the formatting was the content. It talked about things nobody in that company could use, and given what I found out about the company, I strongly doubt that they themselves even understood or knew what they had sent her. She isn't technical herself and neither are her people in production. Now and then they tried something in GPT and it felt like sci-fi to them. The document lacked one single paragraph about who would actually use it there and why it should make sense.
They sent her a document that looked professional and technical, but was full of nonsense that couldn't be used or would cost hundreds of thousands, which for her size would be overkill. You always have to think about what the customer needs, will use, will understand and what makes sense. And above all the return, the saving and who in the company will be able to use it.
Why this happens
A lot is written about agents today, so they sell well. A company hears a word it knows from LinkedIn, and the supplier has something to offer.
The problem is that between “it can be built” and “people at your place will use it” there is a gap nobody pays to bridge. The AI Momentum 2026 survey by ČAUI and the Chamber of Commerce among 1 033 Czech companies says that 80 percent of companies lack qualified people and most admit they can't measure the return. Into that situation someone sells them an agent.
I don't want to claim this holds for everyone. I know people who build agents well and know exactly why. But very often it looks like this: the supplier goes after what sells well and doesn't care who will open it in a month.
What has changed in the tools in the meantime
This is the part most of those offers ignore. Models in the cloud already have agents inside.
Claude from Anthropic and GPT now look up facts on their own, check what they wrote on their own, and have connectors to common services. Claude Cowork or Codex can handle things for which a year ago you built your own agent and wrote code around it.
In practice that means one thing. A large part of what is sold as an agent project is now a feature of a tool you buy on a monthly subscription. No code, no integration, no development budget.
That's not an argument against agents. It's an argument for the order. First you use up what comes out of the box, and only what is left gets built.
That's why we do setup, not agents
AI Setup Day is one day at your company. Accounts and access for the people who will really use it. Rules on what may go into the tool and what may not, including what must never be uploaded. Three processes turned into a procedure that someone who knows nothing today can handle. And training for those specific people, not a presentation for a hall.
No code. No integration. Nothing that someone would have to maintain after us.
From 999.60 EUR for up to about ten people. For comparison, that owner paid 1 200 EUR for a document.
Four questions that decide instead of hype
With every company we go through them before any tool comes up. Not because we're thorough, but because without them you can't price it honestly.
What the company needs. Not what it wants. Wanting something and needing something are very different, and usually we find they need something other than what they called about.
Who will be able to use it. A specific name, not a department. I recommend what people will actually learn, not what is objectively best. If the team won't move to a better tool, I leave the worse one that they will use.
How it will work in a year. Who owns it when that person leaves or is on holiday. A procedure known only to the one who was at the training survives exactly until their holiday. Otherwise it's a dead folder.
How many hours it saves. In hours and in euros, measured before the start. A feeling is not a saving.
How I work out whether it makes sense at all
Deliberately boring. Hours a week times the hourly rate of the person who does it, times fifty weeks.
That's not an argument for buying. That's the size of the problem. But without it you decide by impression, and that is exactly how a thirty-page PDF for thirty thousand comes about.
The second anchor I give owners: an in-house AI specialist comes to roughly 5 556 EUR a month including contributions. An external team with us starts from 396 EUR a month. This comparison is more useful for a decision than any list of features.
When agents do make sense
When the process runs daily, has a clear input and output, and there is someone in the company who can tell when the agent made a mistake. The worst error is not a crash. The worst error is silence, when the system quietly does something else and pretends everything is fine.
With us, actions involving money are never done by AI alone. A payment, sending a contract or a quote is always confirmed by a person. Full automation without human control is not a system, it's a risk.
When these conditions hold at your place, we're glad to build an agent. It's just almost never the first step.
What one day won't do
It won't connect AI to your ERP or warehouse system. That's an integration, and it has its own budget and its own risks.
It won't change your processes. It speeds up the ones you already have. If a process is badly built, AI will do it badly, faster.
We measure before the start and after a month. If the numbers don't make sense, I'll tell you first.
If you want to talk it over
Fifteen minutes on the phone, no obligation, and I'll tell you straight whether setup makes sense for you, or whether you should keep the money: cal.com/transformuj.ai/30min
How much of what you paid for AI last year does someone in the company actually open today?
The article was first published on LinkedIn. Original article on LinkedIn
