Why LLMs fall short on government tenders, and how to make them useful

A general LLM can summarise a notice you paste in, and it can suggest how to describe the company. It falls short of a bid until the boards for your market are checked, later amendments are read, and the selection lines in an annex are in the prompt. Give it those three before anyone drafts, and it can be useful.

One reply is not a watch on the boards

TED, a national portal, Find a Tender, SAM.gov and a development-bank site are different boards, and a notice can be on only one of them. A chat answers from the text in the conversation and from pages it opened in that sitting. It does not, as part of that reply, check those boards again tomorrow. A title, a buyer and a deadline in the answer still have to be opened on the official notice. If that notice was not retrieved, the details are unsupported.

The annex is where offers are rejected

The title and the short description are the pages a chatbot summarises well. Selection usually sits further in: a turnover figure, sometimes an average of stated years and sometimes only the field of the contract; a reference above a set value, for a set kind of client, finished inside a set period; a certificate in force, in the company’s name. Similar work under the figure does not clear the reference. A sentence about a quality system is not the certificate. Leave the annex out of the paste and it is out of the analysis, while the reply can still sound finished.

Do not let the model fill a hole in the company file

If the prompt describes the company in a paragraph, that paragraph is the file the model has. A certificate you did not mention should stay unknown. A project under the minimum should stay under the minimum. A larger firm in the same sector is not yet a partner. Under Article 63 of Directive 2014/24/EU, reliance on another entity’s professional experience is allowed only where that entity will perform the related work, and you can prove you will have the resources. A suggested name proves neither.

Give the agent three skills, in this order

A paragraph in the prompt is not a procedure. Give the agent these three skills, and do not let it draft before the first two are done.

• Configure a scope first: what you sell, where you bid, the certificates, and the reference projects with their clients, dates and values. A line that is not in that file stays unknown. Unknown is not a match.
• Then fit analysis, on the official notice, the amendments and the annex that holds the selection criteria. Each requirement is set against that file. A project under the minimum stays a gap. Where the notice allows another entity to cover professional experience, that entity has to perform the related work.
• Proposal draft starts only after that evaluation exists for the same scope and tender. The brief keeps the unmet lines. It is a note for the bid team. It is not sent to the buyer. Proposal briefs are on Teams.

How Tender Town makes the model useful

Tender Town is the platform that runs that order. It scouts public tenders, grants and funding calls from procurement portals, TED and development banks, and it keeps signals that a named buyer has announced a procurement that is not open yet. A pipeline runs every four hours, and each scope refreshes as often as its plan allows, so the check continues after the chat is closed. On a notice it has collected, it fetches the documents, lists the requirements, scores the fit against the company evidence on the scope, and prepares a proposal brief from that evaluation. It does not file the offer.

Connect your own agents