News
AI agents8 min read

An AI Agent for Public Procurement and Tender Analysis

Suppliers rarely lose tenders on price. They lose because they spotted the opportunity too late, or spent weeks preparing a bid they were never eligible for. Here is where an AI agent helps and where a human must have the last word.

Companies bidding for public contracts rarely lose on price alone. They lose much earlier - by hearing about a suitable contract three days before the deadline, or by spending two weeks preparing a bid for a tender where they never met one of the qualification requirements. The sales team did nothing wrong. It simply has dozens of sources to watch, hundreds of notices a month, and documentation that has a different structure, a different vocabulary and a different length in every single tender.

This is exactly the kind of work where deploying an AI agent makes sense: a lot of reading, very little actual decision-making, and a high price for a missed detail. The goal is not for the agent to submit the bid - that would be the wrong ambition. The goal is that the person who decides on the bid gets a briefing containing the things that actually matter, and gets it on the day of publication rather than ten days later.

Where the real bottleneck is

Ask a salesperson what eats the most time in tendering and they rarely say "writing the bid". They say two other things.

The first is monitoring. Relevant opportunities are not in one place. They sit in official bulletins, on contracting authorities' profiles, in email notifications, in dynamic purchasing systems, in lower-value contracts published only on one institution's website, and sometimes in subcontracting enquiries from larger players. Every source has a different structure and a different quality of description. Keyword filtering fails, because the same thing is named entirely differently in two notices: "development of an information system" in one, "software engineering services" in the other.

The second is a fast bid/no-bid call. You cannot make it from the notice alone. The real decision depends on five or six things buried deep in the documentation: submission deadlines and formats, qualification criteria, required references and their parameters, the evaluation method and the weighting of each criterion, bid security, contractual penalties. By the time someone digs all of this out and assembles it, several days have passed - and that happens for every tender, including the ones you eventually decline.

The cost of a tender is not incurred while writing the bid. It is incurred when the bid is written for a contract you should never have entered.

What the agent actually does with tenders

In this scenario the agent is not a chatbot over documents. It is a process that runs continuously and ends with a concrete output for a specific person. If you are unsure how such an agent differs from ordinary automation, a good starting point is the explanation of how AI agents work in companies.

Continuous monitoring of sources

At a regular interval the agent walks through a defined list of sources, downloads new notices along with attached documentation, and stores them with a timestamp. The boring part matters most here: keeping track of what it has already seen, detecting changes and addenda to existing tenders, and recording the source each document came from. Without that, nothing can be verified later.

Filtering by relevance

Instead of keywords, the agent compares the content of a notice against your company profile - what you actually deliver, which references you hold, which contract sizes and regions you go after. The output is not a binary yes/no but three buckets: clearly relevant, borderline, clearly out of scope. The borderline ones are the most valuable, because those are exactly the ones that get lost in manual triage.

Extracting parameters into a structured summary

For relevant tenders the agent reads the whole documentation and pulls out the same set of fields every time: deadlines and milestones, the subject and scope of the contract, qualification criteria, required references with their parameters, evaluation criteria and their weightings, required formats and annexes, and the contractual terms worth reading sooner rather than later. The result is a one-page summary with an identical structure for every tender - and that comparability is the main added value. Technically this is the same principle used for contract analysis and review in a legal department.

Flagging the conditions that disqualify you

This is the single most useful function of the whole solution. The agent compares the qualification criteria and required references against what you know about your own company and explicitly flags the points you probably do not meet. Not to decide anything, but so that it is visible on the first screen rather than on page 140. For the agent to answer from your own records - the reference list, certificates, staffing capacity - it needs access to them; we covered how to do that safely in the piece on connecting an AI agent to internal documentation via RAG.

A first pass at the repeating annexes

A significant part of any bid is nearly identical in every tender: the company profile, the reference list, statutory declarations, the methodology description, team CVs. From these blocks the agent can prepare a pre-filled first version matching the formal requirements of the specific tender. It is a draft and nothing more - it saves retyping, it does not replace review.

In short: The agent's job is not to write the bid, but to shorten the time from publication of a tender to a qualified decision on whether to bid at all.

A bid/no-bid score: an aid, not a verdict

Once your summaries have the same shape, scoring suggests itself. It makes sense, but only as a triage aid - not as a decision. A score is useful for knowing what to work on today and what can wait, not for rejecting anything on your behalf.

In practice, responsibilities split roughly like this:

Decision factorWhat the agent suppliesWho decides
Qualification criterialist of conditions with citations, marked as met / not met / uncertaina human, always for the final confirmation
Required referencescomparison of requirements against your reference recordsa human
Evaluation criteria and weightingsbreakdown of criteria, weights and calculation methoda human
Deadlines and milestonesa proposed calendar with remindersa human plus an independent calendar outside the agent
Available team capacitycannot assessa human only
Strategic intent and relationship with the authoritycannot assessa human only
Pricing strategycannot assessa human only

Notice that half the rows end with a human. That is not a weakness of the design, it is its point. The agent removes the mechanical work that precedes the decision, not the decision itself. This is the classic human-in-the-loop scenario we examined when comparing autonomous agents with agents under human supervision.

Every condition must carry a cited source

This is a hard requirement, not a nice-to-have. If the agent writes "a reference of comparable scope within the recent period is required", that sentence is unusable for decision-making unless it comes with the document, the page and the verbatim passage it was drawn from.

There are three reasons. The first is verifiability: a reviewer must be able to jump into the original in seconds and confirm or refute the claim. The second is precision of wording - in public procurement the exact wording decides, and a paraphrase changes the meaning even when it sounds the same. The third is team trust: after two unpleasant surprises, people stop opening a summary without citations and go back to reading the documents by hand.

The practical design consequence: the summary must never contain a sentence without a source reference. If the agent cannot say where a condition came from, it has to state that it did not find it - not fill it in based on what is customary. It must equally flag the places where the documentation is contradictory or unclear, because those are the prime candidates for a request for clarification.

Where the agent must not be the last check

The agent works with documents that change. An addendum is issued, a deadline is amended, an answer to another bidder's question shifts the reading of a condition. The model can also miss a passage in a scanned annex or in a table that converts badly to text. Two rules therefore apply without exception.

First: before submission, a human must verify every binding condition against the original documentation. Not against the agent's summary - against the source document in its current version, including all addenda and clarifications. The summary is navigation, not the source of truth.

Second: the agent must not be the sole guardian of a deadline. The deadline belongs in the company calendar with multiple reminders and a named person accountable for it. If an integration breaks, a container falls over or the structure of a source page changes, the agent simply stops sending notifications - and system silence is indistinguishable from "nothing new came in".

Caution: If an agent outage would mean a deadline quietly slips past you, the design is wrong no matter how well the agent reads documentation.

When it is not worth it

If you bid for four contracts a year and all of them come from two familiar authorities, an agent will not save you - the overhead of running it outweighs the benefit. It starts to make sense with dozens of tracked opportunities, several sources, and a team already producing these summaries by hand and falling behind. It equally fails in a company that has no order in its own references and certificates: the agent has nothing to compare against and its output becomes a mere description of the documentation. Similar entry conditions apply to most automation, as we wrote when discussing how to choose which processes to automate.

Summary

An AI agent for public procurement solves one specific thing: it shortens the path from publication of a tender to an informed decision on whether to bid. It does so through continuous monitoring of sources, filtering by relevance, extraction of the same set of parameters into a comparable summary, flagging of the conditions that disqualify you, and preparation of a first version of the repeating annexes.

Its usability depends on citing a source for every claim and on a clear division of labour: the agent prepares, the human decides, and before submission verifies every binding condition against the original. A deadline must never hang on the agent alone. At INTERFASE we build such solutions so that the output is always traceable back to the original passage - you can read more about our approach in the overview of AI and automation solutions.

If you are weighing whether something like this pays off in your case, get in touch and we will go through your sources and tender volume together - that is usually enough to say fairly quickly whether the project makes sense or not.

INTERFASE