Process Automation · Procurement

Screening public tenders with AI: see faster whether a bid is worth it

Across 16 portals, around 1,000 new public tenders appear every day. For a single company, perhaps ten of them fit. AI screening takes over the sifting and reading: it filters, reads the documents and hands over a summary with requirements, deadlines and risks for every tender worth a look. What happens next is your team’s decision.

Julia Rose 7 min read
Two people at a meeting table going through documents, one marking a passage with a pen

Key takeaways

  • The effort sits in the sifting: in our own portal, around 1,000 tenders a day come down to 10 to 15 after two filter stages.
  • The first stage works on keywords, the second reads the full documentation and judges whether the tender fits the company profile.
  • The summary covers requirements, certificates, bid bond, deadlines and risks, along with a recommendation and its reasons.
  • Doubtful cases deliberately land on the review pile. The decision about a bid stays with a person.

Anyone who bids regularly for public contracts knows the Monday morning: several portals, dozens of new notices, and each one comes with a specification, eligibility criteria, a draft contract and often hundreds of pages of annexes. By the time it is clear whether a bid is worth writing, half the day is gone.

Exactly that part can be handled by AI screening. How it filters, what it pulls out of the documents and where its limits are is what we show below. The figures come from our own portal, which Apollogic and theBlue.ai use for their own bidding work.

Why sifting costs so much time

Our portal searches 16 sources of Polish public tenders, and around 1,000 new notices appear there every day. The keyword filter brings that down to about 50. What is left only becomes clear once the AI has read the documents: 10 to 15 tenders that someone actually looks at. One percent.

Sifting gets expensive through sheer volume: almost everything you check ends up in the bin. At the same time the clock is running, because submission deadlines keep moving closer.

Without support, a team picks daily between two evils: sift thoroughly and lose hours on tenders that end up in the bin, or sift quickly and miss the ones that fit.

How AI screening works

Stage one filters the notices

From the title, the buyer, the category and the names of the attachments, the system forms a first assessment. It rates how well the tender fits the company profile, how confident it is, and gives a brief reason. Because this stage runs across everything, it has to stay fast and cheap. Whatever obviously does not fit drops out here.

Stage two reads the documents

For the rest, the system downloads the full documentation including annexes and evaluates it in context, because only then does it show what a tender really demands. A notice can look perfect at first glance and still require a certification in the annex that your company does not hold.

At the end there is a summary that is built the same way every time. That sounds like a small thing, yet it saves the most time: when the bid bond is always in the same place, you read twenty tenders in the time five would otherwise take.

FieldWhat it contains
RequirementsContractual, financial, staffing and performance obligations
EvidenceRequired certificates and references
Bid bondAmount and form of the security
DatesSubmission deadline, validity period, planned start of delivery
RisksPenalties and other points that can get expensive
RecommendationRecommend, review or decline, each with its reasons

Why the recommendation is not a decision

Whether a bid gets written depends on things that appear in no tender document: free capacity, running projects, how you have worked with the buyer so far, the question of whether this market is of strategic interest at all. A model knows nothing about any of it. That is why there is a middle category on purpose.

Everything the system places with any lingering doubt lands there. That setting deliberately costs time, because an error in one direction is more expensive than in the other: a missed tender stays invisible forever. One read too many costs ten minutes.

So that such errors surface at all, every recommendation states its reasons and the passage in the documents it relies on. Only then can the filter be checked and improved.

In practice: ApoTenderHub

Internal portal of Apollogic and theBlue.ai

ApoTenderHub searches for public tenders in our areas of interest, checks the requirements and drafts an initial offer. It currently covers 16 sources of Polish public tenders. The language models run on our own servers, because large volumes of text have to be read every day.

How we use local models for this

Most of the work sits not in the AI but in sourcing the tenders: every portal publishes its notices differently, so each source needs its own connection. The screening itself depends on neither language nor country: a model that understands Polish tender documents understands German ones too.

Cloud or your own servers?

For a handful of tenders a day, a model from the cloud is the simplest route. As soon as hundreds of documents with many pages go through the system every day, the arithmetic changes. Then the cost per document counts, as does the question of where these partly confidential documents are processed.

Which models suit the job, how fast they answer and what the hardware costs is covered in our article Local LLMs in practice. If you stay in the cloud, it is worth knowing which providers train on the data you enter, which is covered in Do AI providers train on company data?

Where to start

Best with one source and one type of contract. For that, the system needs three things from you. First your profile: what you offer, in which regions, with which references and certificates. Second your exclusion criteria, such as contract sizes or contract forms that are out of the question anyway; and examples from the past, both tenders you bid for and tenders you deliberately declined.

That last point is often skipped in practice although it matters most. Only against old cases can you see whether the screening really calls it better than the previous judgement. From those examples you can see after two weeks how many of the tenders you previously won the system would have found.

In a process analysis at a fixed price from €3k we look at your sources and criteria and propose an architecture with scope, timeline and cost. First working components arrive in six to eight weeks, built with real tenders from your market.

How many tenders does your team sift each week?

In the process analysis we look at your sources and criteria and propose screening that fits your bidding process.

Request a process analysis

Frequently asked questions

Yes. Screening filters the notices by keyword, then reads the full documentation and hands over a summary with requirements, deadlines and risks for every tender worth a look. In our own portal this brings around 1,000 tenders a day down to the 10 to 15 that a person reviews.

Yes. Specification, draft contract and annexes are read together, including PDFs and scans, because that is where the details that decide a bid sit, such as required certificates or the size of the bid bond. Illegible passages are flagged as uncertain.

Reading the documents does, independently of language and country. Every portal publishes its notices differently, though, so each source needs its own connection. Our portal ApoTenderHub currently covers 16 sources of Polish public tenders.

No. It takes over the sifting and reading and delivers a recommendation with reasons: recommend, review or decline. The team decides on the bid, because capacity, running projects and how you have worked with the buyer so far appear in no tender document.

Through the middle category: everything the system places with any lingering doubt goes to the review pile. On top of that, every recommendation states its reasons and the passage it relies on, and the filter is measured against tenders the company bid for in the past.

That depends on the deployment. In our portal the language models run on our own servers, because large volumes of text have to be read every day and the documents are partly confidential. Cloud or hybrid operation is equally possible.

Julia Rose

About the author

Julia Rose

Marketing Lead, theBlue.ai

Julia has been part of theBlue.ai since 2019 and has accompanied the development of AI applications in the enterprise environment since the company’s early days. In her role as Marketing Lead, she works closely with the engineering and consulting teams and makes complex technical topics understandable and accessible for decision-makers.

In her articles, she writes about practical experience from enterprise AI projects, as well as the challenges and opportunities of using AI in companies.