From BZP to a Win — How AI Scoring Helps You Pick Tenders with the Highest Chance of Success

Every day, more than 800 new notices appear in the Public Procurement Bulletin (Biuletyn Zamówień Publicznych, BZP). Your firm may see 40–60 that fit the profile, but it can realistically prepare 10–15 bids a month. Which ones do you choose? Traditionally the bid/no-bid decision rests on intuition, the experience of the head of bidding — and luck. In 2026 that is no longer enough. AI scoring changes the rules: the algorithm analyses thousands of tender features and your win history, then tells you straight: “This tender you have a 73% chance of winning. That one — 12%. Don’t waste the time.”

This article shows how AI scoring works in public procurement, what data it uses, how much it actually saves, and where its limits sit. No marketing smoke — tables, numbers, and a case study from the Polish construction market.

The cost of bad bid/no-bid decisions

Preparing a professional tender bid is not “filling in a form.” It is a process that pulls in many specialists for many hours:

Analysis of the SIWZ/OPZ (tender specifications / description of the subject of the contract) — 8–16 hours (lawyer + subject-matter specialist)
Cost estimate — 12–24 hours (estimator + project manager)
Formal documents — 8–16 hours (administration + compliance)
Site visit / due diligence — 4–8 hours (engineer + manager)
Review and corrections — 8–16 hours (board + quality control)

Total: 40–80 person-hours per bid. At an internal rate of PLN 120–250/h that means a cost of PLN 5,000 to 20,000 on every bid submitted — regardless of the outcome.

The brutal mathematics of win rate

The average win rate in public tenders in Poland oscillates between 15% and 25% — depending on industry, firm size, and type of contract. Broken down into concrete numbers:

Parameter	Value
Bids submitted per month	20
Average cost of preparing a bid	PLN 12,000
Monthly bidding budget	PLN 240,000
Win rate	20%
Bids won	4
Bids lost	16
Wasted budget per month	PLN 192,000

At 20 bids a month and a 20% win rate — fully 80% of your bidding budget is thrown away. Annually that is more than PLN 2.3 million spent on losing. Not on delivering contracts — on the mere attempt to win them.

The hidden opportunity cost

Direct cost is only the beginning. When your best estimator spends three days on a bid you will not win — they are not working on the bid you could have won. When a lawyer analyses the SIWZ of a tender doomed from the start — they are not reviewing the contract on a PLN 5 million job.

The opportunity cost of a bad bid/no-bid decision is a multiple of the direct cost. And that is exactly where AI scoring comes in.

What Is AI Scoring in Tenders?

AI scoring is a machine-learning algorithm that, on the basis of historical data and the features of a specific tender, computes the probability that your firm will win the bid. It is not a “magic ball” and not “artificial general intelligence” — it is a precise statistical model trained on real data from the public procurement market.

How it works, in brief

Input — tender features (value, CPV, contracting authority, award criteria, deadline, region) + your firm’s profile (bid history, competences, references, delivery capacity)
Processing — the ML model compares the tender’s features with patterns from past winning and losing bids
Output — a probability score on a 0–100% scale (e.g. “73% chance of a win”) + a list of risk and opportunity factors

The goal: bid ONLY where the chance exceeds the threshold

AI scoring does not say “you will win” or “you will lose.” It gives you a number — a probability. You set the decision threshold:

Score > 60% → definitely submit a bid (green zone)
Score 35–60% → consider it if you have spare capacity (yellow zone)
Score < 35% → do not bid, unless there are strategic reasons (red zone)

The effect? Instead of 20 bids “blind” — 8–10 targeted bids, with a win rate of 30–40% instead of 18–20%. Less work, more wins, better ROI.

What Data Does the Scoring Model Use?

The effectiveness of AI scoring depends directly on the quality and quantity of the data it learns from. A good scoring model uses three categories of information:

1. Your bid history (internal data)

This is the foundation — without it the model does not know “who you are” as a bidder:

A list of submitted bids (wins + losses) from the last 2–5 years
Prices offered vs. winning prices
Types of contracts in which you win most often
Contracting authorities with whom you have a history of cooperation
References and experience in specific CPV codes

2. Tender features (public data)

Every notice in BZP/TED contains dozens of features the model can analyse:

Contract value (estimated and budgetary)
CPV codes (classification of the subject of the contract)
Contracting authority (type of institution, procurement history, budget scale)
Delivery region (distance from the firm’s registered office)
Award criteria (weight of price vs. quality vs. deadline)
Procedure type (open tender, restricted, negotiations)
Bid submission deadline (how much time to prepare)
Formal requirements (bid bond, performance security, participation conditions)

3. Competitor data (market intelligence)

This is the element that delivers the largest edge — knowledge of who else will bid:

Historical participants in procedures at a given contracting authority
Typical prices offered by competitors on similar contracts
Current load on competitors (contracts being delivered)
Competitors’ win rates in similar CPV categories

Table: scoring-model features

Feature	Weight in the model	Data source
Historical win rate in a given CPV	High (18–22%)	Firm’s internal data
Number of potential competitors	High (15–20%)	BZP/TED — procedure history
Ratio of your price to the market average	High (14–18%)	Bid-opening protocols
Experience with the contracting authority	Medium (10–14%)	CRM + BZP
Contract value vs. your average	Medium (8–12%)	Internal data
Weight of the price criterion	Medium (7–10%)	SIWZ / notice
Geographic distance	Low (5–8%)	GUS + registered offices
Delivery deadline vs. capacity	Low (4–7%)	ERP / schedule
Required vs. held references	Medium (8–12%)	Internal data + SIWZ

The weights are not fixed — the model learns them automatically from the training data. The table above shows typical distributions for mid-sized construction firms (30–100 employees).

How Scoring Works, Step by Step

The full scoring process has five stages — from fetching the notice to a decision recommendation. Below we break down each of them:

Step 1: Ingestion of notices (BZP/TED/BIP)

The system automatically fetches new notices from three main sources:

BZP (Public Procurement Bulletin) — national procedures below the EU thresholds
TED (Tenders Electronic Daily) — procedures above the EU thresholds (via eForms)
BIP (Public Information Bulletins) — notices published locally by contracting authorities

Ingestion runs in real time (BZP API) or in 15–60 minute cycles (BIP scraping). Every notice is parsed into a standardised internal format — regardless of source, the model sees the same fields.

Step 2: Feature extraction (NLP on SIWZ/OPZ)

Notice metadata alone (value, CPV, deadline) is not enough. The real value sits in the body of the documents — the SIWZ (Specyfikacja Istotnych Warunków Zamówienia, tender specifications) and the OPZ (Opis Przedmiotu Zamówienia, description of the subject of the contract). An NLP (Natural Language Processing) model analyses the full text and extracts:

Technical requirements — which technologies, materials, certificates are required
Participation conditions — minimum experience, turnover, staff
Award criteria — how the contracting authority will score bids (not only weights, but the concrete sub-criteria)
Risk clauses — contractual penalties, deadlines, liability
Hidden preferences — wording that suggests the authority has a “favourite” contractor (e.g. highly specific reference requirements)

NLP extraction runs on language models fine-tuned on a Polish corpus of tender documentation — this is not general-purpose ChatGPT, but a specialised model that understands PZP terminology, SIWZ structure, and the context of public procurement.

Step 3: Matching against the firm profile

Extracted tender features are compared with your firm’s profile:

Do you have the required experience (references in the relevant CPV)?
Does your staff meet the participation conditions?
Does the contract value sit in your “sweet spot”?
Do you have spare delivery capacity in the performance window?
Is your typical price competitive in this segment?

Matching produces a fit vector — a numerical representation of how well your firm “fits” a given tender.

Step 4: ML scoring (gradient boosting / ensemble)

The heart of the system — the machine-learning model. In practice the best results come from ensemble models (a combination of several algorithms):

Gradient Boosting (XGBoost / LightGBM) — the main predictive model, trained on bid history
Logistic Regression — a baseline linear model as a benchmark and a stabilising element
Random Forest — an additional tree model that captures non-linear relationships
Ensemble — a weighted combination of the above, minimising variance and bias

The model takes as input the tender feature vector + the firm-profile fit vector and produces a probability score — the probability of winning the bid on a 0–1 scale (presented as 0–100%).

Step 5: Ranking + bid/no-bid recommendation

The last stage is aggregation and presentation of results:

Ranking — all new tenders ordered from highest to lowest score
Classification — a split into zones (green/yellow/red) according to thresholds set by the firm
Recommendation — a clear suggestion: “Bid” / “Consider” / “Do not bid”
Rationale — a list of the top-5 factors driving the score (positive and negative)
Alerts — warnings about unusual features (e.g. “In 90% of cases this contracting authority selects the incumbent contractor”)

The full pipeline — from the notice appearing in BZP to a recommendation on the dashboard — takes from 2 to 15 minutes, depending on the availability of SIWZ documents.

Scoring ROI: How Much Does a Firm Submitting 15 Bids a Month Save?

Let’s get to concrete numbers. Compare two scenarios — a construction firm bidding “the traditional way” (no AI) and the same firm using AI scoring:

Scenario: construction firm, 50 employees, road works

Parameter	Without AI	With AI (scoring)	Change
Bids submitted / month	15	9	−40%
Average cost of preparing a bid	PLN 14,000	PLN 14,000	—
Monthly bidding cost	PLN 210,000	PLN 126,000	−PLN 84,000
Win rate	18%	31%	+72% relatively
Bids won / month	2.7	2.8	+0.1
Average value of a won contract	PLN 1,800,000	PLN 1,800,000	—
Revenue from wins / month	PLN 4,860,000	PLN 5,040,000	+PLN 180,000
Cost per won bid	PLN 77,778	PLN 45,000	−42%

What happened?

The firm did not win significantly more contracts (2.7 vs. 2.8/month) — but it achieved a comparable result while submitting 40% fewer bids. The key benefits:

A saving of PLN 84,000 a month on bidding costs (−45% of the budget)
240 person-hours a month freed — time the team can spend on better preparation of green-zone bids
Win rate up from 18% to 31% — a selection effect (you bid only where you have a real chance)
Lower cost per win — from PLN 78,000 to PLN 45,000 for every contract acquired
Better team morale — less “work for nothing,” more successes

ROI at annual scale

Metric	Annual value
Saving on bidding costs	PLN 1,008,000
Additional revenue (better matching)	PLN 2,160,000
Cost of the scoring system (licence + deployment)	PLN 180,000
Net ROI (year 1)	460%

Even under conservative assumptions (win rate rises “only” from 18% to 26%, not to 31%) — first-year ROI exceeds 200%.

Case Study: A 60-Person Construction Firm — Scoring in Practice

Firm profile

Industry: construction works — road infrastructure and buildings
Headcount: 60 people (including a 5-person bidding department)
Region: Silesia + Lesser Poland (150 km radius from HQ)
Annual turnover: PLN 45 million
Bids per month: 12–18 (average 15)
Win rate before deployment: 20% (3 wins / 15 submitted)

The problem

The firm bid on “everything that fits the CPV.” The bidding department worked under constant deadline pressure, with no time for a deeper analysis of the odds. Bid/no-bid decisions were taken by the CEO on the basis of a “gut feel” and spare delivery capacity. The result: they regularly lost tenders they never had a chance of winning (too expensive, too little experience, too much competition) — and at the same time they did not bid on contracts they could have won (because they “didn’t make the deadline”).

Scoring deployment

Stage 1: Data collection (4 weeks)

The firm supplied a history of 45 bids from the last 3 years — 9 wins, 36 losses. For each bid they collected: procedure number, contracting authority, value, CPV, number of competitors, price offered, winning price, award criteria.

Stage 2: Model training (2 weeks)

The model was trained on the firm’s data + public BZP data (a history of 180,000 procedures in CPV 45.xx.xx.xx from 2021–2025). Accuracy on the test set: 74% (AUC-ROC: 0.81).

Stage 3: Pilot (4 weeks)

For a month the firm used scoring “informationally” — it saw the score but did not change the decision process. Verification: the model correctly predicted the outcome of 11 of 14 bids submitted in that period.

Stage 4: Full deployment (from week 5)

The firm set the bid/no-bid threshold at 45% and began applying the system’s recommendations.

Results after 3 months

Metric	Before	After	Change
Bids submitted / month	15	8	−47%
Win rate	20%	35%	+75% relatively
Wins / month	3.0	2.8	−0.2 (within statistical noise)
Bidding cost / month	PLN 195,000	PLN 104,000	−47%
Bidding-department workload	100% load	62% load	−38%
Bid quality (internal score)	6.2/10	8.1/10	+31%

The key insight: the firm does not win “more” contracts — it wins a comparable number, but at a radically lower effort. The freed time in the bidding department is spent on more thorough preparation of the selected bids, which further raises quality and the chance of winning.

Quotes from the deployment

“The biggest change is calm. Previously every Monday started with panic — 6 new tenders we ‘have to’ bid on. Now the system says: ‘These 2 make sense, drop those 4.’ And it is right in 3 out of 4 cases.” — Head of bidding

“The first month was hard — psychologically it is difficult to walk away from a PLN 3 million tender even if the algorithm says 18% chance. But when we saw we had won 3 of 8 submitted bids instead of the usual 3 of 15 — the mathematics spoke.” — CEO

Limits of AI Scoring (An Honest Assessment)

AI scoring is not a panacea. Like any tool, it has limits — and presenting them honestly is a condition of a responsible deployment. The most important:

1. It requires a minimum of 30–50 historical bids

An ML model needs data to learn. If your firm has submitted fewer than 30 bids in its history — the model will not have a sufficient sample to train. Practical minimums:

30 bids — a baseline model (low prediction confidence, wide confidence interval)
50–100 bids — a standard model (moderate confidence, operationally useful)
100+ bids — a mature model (high confidence, stable predictions)

Firms submitting 2–3 bids a month need at least 12–18 months of history to collect a sufficient sample.

2. It does not work on niche tenders (no market data)

If your firm specialises in a very narrow niche (e.g. conservation of class-0 historic monuments, cryogenic installations, tactical communications systems) — the model may not have enough comparative market data. Scoring works best where:

A given CPV has at least 50–100 procedures a year in BZP
At least 5–10 different contractors regularly bid
Award criteria are reasonably standard (price + quality/deadline)

3. It will not replace a human assessment of the relationship with the contracting authority

The algorithm does not “see” what an experienced salesperson knows:

That contracting authority X is unhappy with the current contractor and is looking for a change
That in municipality Y the administration changed and the new mayor prefers local suppliers
That in a month the authority will announce a larger tender in which it is worth having a reference from a smaller one

This “soft intelligence” is priceless — and AI scoring should be a decision-support tool, not a replacement for the human.

4. The model requires regular updates

The public procurement market changes — new players enter, old ones leave, material prices fluctuate, regulations change. A model trained on 2022 data may give wrong predictions in 2026. What is required:

Retraining every 3–6 months on new data
Drift monitoring — automatic detection of a drop in prediction quality
Feedback loop — every submitted bid (win/loss) returns to the model as a new data point

5. It does not eliminate risk — it only quantifies it

A score of 73% does not mean “you will win.” It means “in 73 cases out of 100 with similar features — firms with your profile won.” 27% of uncertainty remains — coming from factors the model does not see (e.g. an aggressive price from a new competitor, a formal error in the bid, a change of decision by the contracting authority).

BudOS — Scoring from 1.4 Million Historical Tenders

Where does the scoring model’s edge come from?

The effectiveness of AI scoring depends on two things: the quality of the algorithm and the volume of training data. Algorithms (XGBoost, LightGBM, ensemble methods) are publicly available — data is the real barrier to entry.

BudOS operates on a base of 1.4 million historical tenders from the Polish public procurement market — notices, bid-opening protocols, procedure outcomes, and SIWZ documentation from 2018–2026. These are not “raw” BZP data — they are processed, normalised, and enriched records on which ML models can learn effectively.

How does BudOS build scoring models?

Market database — 1.4M tenders, 4.2M submitted bids, 890,000 unique contractors. For every procedure: who bid, at what price, who won, what the criteria were.
Firm profile — on the basis of your bid history + data from KRS/GUS, BudOS builds a multi-dimensional profile: your “sweet spots” (CPV × value × region × type of contracting authority), your typical prices vs. the market, your competitive position.
A model per firm — scoring is not “one size fits all.” BudOS trains a separate model for each client, taking into account the specifics of their portfolio, region, and pricing strategy. A market-wide model serves as a prior (starting point), and the firm’s data personalises it.
Real-time scoring — every new notice in BZP is automatically scored in the context of your profile. On the dashboard you see a ranking of tenders sorted from highest to lowest score — with a rationale and risk factors.
Continuous learning — every new bid of yours (won or lost) automatically enriches the model. The longer you use the system, the more accurate the predictions become.

Data advantage = predictive advantage

A firm building a scoring model “from scratch” on its own 50 bids is operating on an extremely small sample. BudOS — thanks to access to 1.4 million historical tenders — can:

Precisely estimate the number of competitors (from the history of procedures at a given contracting authority)
Forecast the price range of competing bids (from their historical prices)
Identify patterns of “rigged” tenders (unnaturally specific requirements)
Detect seasonality and price trends in individual market segments

That is the difference between “guessing from 50 data points” and “predicting from 1.4 million data points + the context of your 50 bids.”

FAQ — The Most Common Questions About AI Scoring in Tenders

Is AI scoring legal in public procurement?

Yes — AI scoring is an internal contractor tool that supports the bid/no-bid decision. It does not interfere with the contracting authority’s procedure, does not manipulate bids, and does not breach any provision of the PZP. Analogy: a cost-estimating calculator is also “AI” (in a broad sense) — and nobody questions its legality. Scoring helps you make a better business decision — that is your right as an entrepreneur.

How long does it take to deploy AI scoring?

A typical rollout timeline:

Weeks 1–2: collect historical data (bids, outcomes, prices)
Weeks 3–4: model training and validation (backtest on historical data)
Weeks 5–8: pilot (informational scoring, no change to the decision process)
Week 9+: full deployment with automatic recommendations

In total 2–3 months from decision to full operational use. The key condition: orderly historical data (at least 30 bids with outcomes).

Does scoring work for small firms (5–10 people)?

Yes, with limits. Small firms typically:

Submit fewer bids (3–5 a month) — so they need a longer history to train
Operate in narrower niches — less comparative data
Have fewer resources for bid preparation — so every hour saved is proportionally more valuable

For firms submitting at least 5 bids a month, AI scoring already pays off at a win-rate improvement of 5–8 percentage points. For smaller firms — it is worth starting with a simpler rule-based model (filters: CPV + value + region + contracting authority) and moving to ML scoring after collecting 30+ bids.

How accurate is the scoring model?

Accuracy depends on data volume and industry specifics. Typical metrics:

Accuracy (overall correctness): 70–80%
AUC-ROC (ability to separate wins from losses): 0.75–0.85
Precision (how many of the recommended bids you actually win): 30–45%
Recall (how many of the winnable bids the system detected): 60–80%

For comparison: the intuition of an experienced bidding director has accuracy on the order of 55–65%. AI scoring is not perfect — but it is better than a “hunch” and delivers repeatable, measurable results.

Can I combine AI scoring with my own expert judgement?

Absolutely — and that is exactly how it should work. AI scoring gives you a “starting point” — a numerical estimate of the odds. Your expert knowledge adds context the algorithm cannot see:

Information about relationships with the contracting authority (+/−10–15 points on the score)
The strategic importance of the tender (entering a new market, building references)
Current team load and availability of key specialists
“Soft” market signals (budget rumours, personnel changes at the contracting authority)

The best results come from firms that treat AI scoring as a “first opinion” — with the final decision taken by a human who has the full context. The algorithm filters noise; the human adds nuance.

Summary: AI Scoring Is Not the Future — It Is the Present

The public procurement market in Poland is PLN 250 billion a year. Firms that bid “blind” — lose hundreds of thousands of zloty a year preparing bids they will not win. AI scoring is not a revolution — it is an evolution. The logical step from “we bid on everything and hope for luck” to “we bid where we have the highest chance and the best potential ROI.”

What do you gain from AI scoring?

A 40–50% reduction in bidding costs — fewer bids, but better chosen
A 50–100% relative rise in win rate — from a typical 18–20% to 30–35%
Team time freed — for better preparation of the key bids
A measurable, repeatable process — instead of “the CEO’s hunch”
Better morale — less “work for nothing,” more successes

When does AI scoring make sense for your firm?

You submit at least 5–10 bids a month
You have a history of at least 30 submitted bids
Your current win rate is below 30%
The cost of preparing one bid exceeds PLN 5,000
You want to scale the business without a proportional increase in the bidding department

If you meet these criteria — AI scoring may be the best investment in the development of your tender pipeline in 2026.

Next step

Want to see how AI scoring would work on your data? BudOS offers a free sample analysis — send the history of your last 10 bids, and we will show you how the model would have scored your chances and where unused opportunities sit.

Book a consultation →