Bid Preparation Costs and Competition in Public Procurement: A Field Experiment

Our intervention is three weeks of free access to an integrated public procurement platform, an artificial intelligence specialized for public procurement that covers tender discovery, documentation analysis, document preparation, workflow management and market analytics. Details in PAP.

2026-09-08

2027-01-09

Firm-Level Outcomes:

1) Number of distinct public procurement procedures for which the firm submits at least one bid.

2) A dummy taking the value of 1 if the firm submits at least one bid, and 0 otherwise.

3) Share of the firm's bids placed on administratively demanding tenders.

4) A dummy taking the value of 1 if the firm holds a paid subscription to the platform, and 0 otherwise.

5) Outcomes 1 to 3 re-estimated on the window that begins when free access ends, to test persistence.

Outcomes 1 to 3 and 5 are constructed by the researcher from the public bid records and contract notices of the Elektronicki oglasnik javne nabave. Outcome 4 comes from the vendor's sales records. None depends on a survey answer, and all are complete by construction, so a firm that does nothing scores zero rather than missing.

1) Distinct procedures rather than lots, so a firm bidding on four lots of one procedure counts once. A consortium bid counts for every member named, since each incurs preparation cost. Winsorized at the 97th percentile of the same variable on the recruitment frame over the four preceding quarters, a threshold computed on the frame and fixed before the draw so that treatment cannot affect where the cap falls. The unwinsorized count is reported alongside.

2) The same count before winsorization, taking the value 1 where it is at least one. Volume is registered on two scales because no single scale is well powered against both shapes the effect could take: under a proportional effect the count is sensitive and the extensive margin is blind, and under an additive effect the ranking reverses. The two are tested jointly.

3) A tender is administratively demanding when its published notice imposes three or more of four qualification requirement types: technical references with client confirmations, a certified quality norm, a performance guarantee, and a subcontracting declaration. The classification comes from the buyer's notice, published before any bid is prepared, so a treated firm can change which tenders it attempts but never how demanding a tender is. Defined for firms with at least one bid; bids whose notice cannot be parsed enter neither numerator nor denominator, and their number is reported.

4) Taken from vendor sales records for every study firm on the same basis.

A single summary index is also reported, formed by signing each primary outcome so that positive means the platform helped, standardizing on the control distribution, and combining by inverse-covariance weighting.

Firm-Level Outcomes:

1) Hours the firm spends on public procurement in total, and hours per month spent monitoring which tenders exist.

2) Hours spent preparing one named bid, split across six steps: discovery, reading the documentation, routine administrative papers, qualification evidence, drafting and submission, and pricing.

3) An index of preparation cost, and an index of uncertainty about requirements and about the winning price.

4) A dummy for bidding in a product division, or to a contracting authority, where the firm did not bid in the twelve months before randomization.

5) Money left on the table, the win rate, and the ratio of submitted price to the buyer's estimated value.

6) Stated willingness to pay per month for such a tool, and monthly spending on general-purpose artificial intelligence tools.

1) Asked unconditionally, with zero permitted, so these exist for every respondent whether or not the firm bid and carry no selection on bidding. If bids rise while these are flat, output per hour of procurement work has risen mechanically.

2) Asked about a specific, named, recent bid rather than in the abstract, so the total is reconstructed from its parts rather than guessed. A count of the distinct days on which anyone worked on that bid serves as a recall check. Because hours per bid depend more on the tender than on the firm, hours are also collected with the tender held fixed: where the register shows that one treated and one waiting firm bid on the same tender, both are asked the same grid for that named tender, so the documentation, requirements, deadline and buyer are identical and only assignment differs.

3) The preparation cost index combines four components measured at endline, each signed so that higher means lower cost, standardized on the control group and weighted by the inverse of their covariance. The uncertainty index combines reversed confidence that the documentation reveals what is required with reversed confidence in the firm's own estimate of the winning price. That price estimate is also scored against the record, as the negative absolute difference between the stated percentage and the median ratio of winning price to estimated value in the firm's main product division, so that accuracy is measured objectively rather than by self-report.

4) Measured from the bid records, to test whether cheaper preparation widens the set of markets a firm enters.

5) Money left on the table is own price minus winning price over winning price, computed within lot so that only prices for the identical object are compared.

Firms that already bid for public contracts sign up, complete a short baseline questionnaire, and are randomly assigned to receive free platform access either immediately or about three months later. Bidding behaviour is then read from the public procurement register for every firm. Details in PAP.

Recruitment begins only after this registration is approved. How many firms take part depends on how many respond to an invitation to try a product, so the realized sample may differ from the planned figures below.

Not available

Randomization done in office by a computer, seeded by a public randomness beacon. The eligible-firm list and the randomization script are fixed and cryptographically fingerprinted before the seed exists, so that neither the sample nor the seed can be chosen after the fact.

The firm. One level of randomization only, and the same unit at which the primary outcomes are measured.

No

500 firms (target). Recruitment opens only after this registration is approved, and the number recruited depends on uptake, so the realized figure may differ from this plan. The minimum considered adequate is 400 firms.

500 firms (target), the same as the number of clusters since the design is not clustered. The realized figure depends on uptake and may differ from this plan.

250 firms immediate access, 250 firms waitlist (target). Assignment is one to one within matched pairs, so the two arms are equal in size by construction up to a single firm per stratum. If uptake falls short of the target, both arms shrink proportionally and remain equal; the minimum considered adequate is 200 firms per arm.

All figures are at 80% power and alpha 0.05, with a planned total sample of 500 firms, 250 per arm, and are obtained by simulating the experiment on the national bid corpus rather than from a closed-form formula, because the bid count is skewed with a mass at zero and closed-form formulas overstate this design by roughly a quarter for count and share outcomes. Firm-level number of bids submitted per quarter: Mean in control group: 2.730, Power: 80%, Total sample: 500, Mean difference: 0.683, Standard deviation: 3.043, Cohen's d: 0.224, Alpha: 0.05. The standard deviation is the residual after analysis of covariance on the baseline count; the raw standard deviation is 4.439. Under a proportional rather than an additive effect the same power requires Mean difference: 1.037, Cohen's d: 0.341. Firm-level probability of submitting any bid: Proportion in control group: 0.564, Power: 80%, Total sample: 500, Proportion in treated group: 0.714, Effect: 0.150, Alpha: 0.05. Bid-level share of bids on administratively demanding tenders: Proportion in control group: 0.219, Power: 80%, Total sample: 500 firms, clustered by firm, Proportion in treated group: 0.314, Effect: 0.095, Alpha: 0.05. Firm-level probability of holding a paid subscription: Proportion in control group: 0.100, Power: 80%, Total sample: 500, Proportion in treated group: 0.188, Effect: 0.088, Alpha: 0.05. Firm-level preparation time per bid: Power: 80%, Total sample: 375 at an assumed 75% endline response, Cohen's d: 0.250, Alpha: 0.05. Preparation time on tenders where one treated and one control firm bid on the same tender: Power: 80%, Total sample: about 89 matched pairs, Cohen's d: 0.230 to 0.300, Alpha: 0.05. Bid-level money left on the table: Proportion in control group: 0.263, Power: 80%, Total sample: 500 firms, clustered by firm, Proportion in treated group: 0.308, Effect: 0.045, Alpha: 0.05. Bid-level win rate: Proportion in control group: 0.412, Power: 80%, Total sample: 500 firms, clustered by firm, Proportion in treated group: 0.502, Effect: 0.090, Alpha: 0.05. Bid-level probability that a submitted bid is rejected on procedural grounds: Proportion in control group: 0.030. Not detectable at any sample size this study can recruit, declared underpowered in advance and reported descriptively. These are targets. Recruitment opens only after this registration is approved and the realized sample depends on uptake, so power is re-computed at the realized sample size on the day recruitment closes, before any outcome is observed, and that curve is reported.