On 15 October I talked about AI agents to a room full of people who live on public tenders: contracting authorities, companies, lawyers. It was a seminar organized by TenderBrain , the vertical of AIDeskPro for public tenders where I serve as CTO, inside the first national conference of FARE, the Italian association of procurement professionals, in Bari. The account of the day is on the Open Gate website , in Italian. Here I keep the four ideas I brought home, because they hold for tenders as for any job made of documents.
The chat answers, the agent works
That is the sentence I started from, and it explains everything else. A chat gives you an answer and stops. An agent plans, picks a tool, looks at the result and tries again, until the job is done. If you ask a chat to analyze a thousand documents, it will never really do it. An agent reads all thousand, one after the other, with a patience that wears a person down and costs the agent nothing.
In tenders you see it immediately: the specifications, the annexes, the previous offers, the CVs. Nobody rereads all of them every time, and the agent does. It is the same difference I described last week in Enterprise agentic systems : with a chat the work is done by the person, with help; with an agent the work is done by the system, and the person decides.
AI is not expertise, it is a multiplier of expertise
This is the passage that made the room smile, and I repeat it because it is the thing most often misunderstood. If the person using AI has no expertise, they multiply zero by zero and the result stays zero. If they have a hundred, the agent multiplies it by a hundred and gets ten thousand. The people who spent years learning the hard way are exactly the ones who get the most today: they do in days what used to take weeks.
The same goes for whoever writes the agents. An agent for tenders is configured by someone who knows tenders, not by a programmer. The programmer is needed for the platform, and that is what we are there for.
Knowing where every piece of information comes from
This is what we care about most. For every piece of information the agent tells the operator where it comes from and why it was used: whether the model knew it, whether the user said it, or whether it comes from a document, and in that case which document and which paragraph.
In the texts it produces, the points that need a look stand out. An improvement proposal is highlighted in blue, because a person has to approve it. A missing piece of data, such as the right CV for the project manager, is in red. Whoever reviews doesn’t have to search: they look at the colors and decide.
This is the practical way to control a piece of work too long to be reread in full. The same principle sits behind the independent checks we describe in How to write an agent : you don’t reread everything, you define beforehand what needs a look and make sure it stands out.
The agent prepares, it doesn’t decide
All the speakers were clear on this, including the lawyer who gave the legal framework. The agent doesn’t assign scores and doesn’t sit on the evaluation committee. It verifies, flags, brings the paragraph, and then it is up to the person to judge. I said it quoting Floridi: AI has the capacity to act, but not intelligence and understanding. Those stay with whoever uses the tool.
It is the same line you find at the bottom of the Open Gate website and in the post about moving from WordPress to an agent : the agent executes, a person decides. It is not a cautious formula. It is the rule that makes an agent usable wherever every choice has to be justified, and a public tender is the extreme case.
The video
The full talk, in Italian, with Gianluca Viganò, Amanda De Cosmo and Federico De Rovere at the table. Names, thanks and the extra chairs brought into the room are in the Open Gate post.
If you work with documents
If yours are public tenders, TenderBrain is the same platform with Configurations, Indexes and checks designed for specifications and technical offers: write to sales@opengate.biz . If it is another job made of documents, a questionnaire, a quote, a screening, the AI agents for business page is the place to start.