AI against fraud in public works procurement

Artificial intelligence is already helping Brazil detect bid rigging and fraud in public works procurement. The Administrative Council for Economic Defense (CADE), the country’s competition authority, has had screening algorithms to spot collusion among bidders since 2013, and the Office of the Comptroller General (CGU), the federal government’s internal audit body, reviewed 118,069 procurement processes in 2024. Construction firms have access to the same techniques and use them to decide where to bid and at what price.

The Brazilian cases are part of a broader shift. In a $13 trillion global industry that is fragmented, lightly digitized, and prone to delays and corruption, many problems take shape during procurement, before ground is broken, according to Francesco Decarolis, a professor at Bocconi University, and Emrah Ergelen, of the Turkish contractor Arge Construction. That is the stage where AI has begun to change outcomes, both for the agency that buys and for the firm that bids.

How does AI detect bid rigging?

Screening compares bidder behavior with what genuine competition would look like and flags deviations for an auditor to review. Cartels leave traces in the data: cover bids priced to lose, a rotation of winners among the same firms, prices that are too close across competitors, and firms that stop bidding in certain regions, as if they had split the market. These patterns emerge when thousands of tenders, bids, and bidders are cross-checked, and they go unnoticed by anyone who reviews one tender at a time.

Machine learning extends the reach of that analysis. Decarolis and Ergelen cite research based on firm-level data that Italian authorities opened to a group of academics in 2022, in which models such as random forests estimated corruption risk. Variables that a human analyst would usually discard were among the strongest predictors.

The tender format also shapes the risk. In earlier work, Decarolis showed that average bid auctions, in which the offer closest to the average of all bids wins, make coordination among bidders easier. Understanding the incentive each format creates makes it possible to design the statistical test that looks for collusion under that format. In any format, a screening result is an indication that guides an investigation and needs further evidence before any sanction.

What to do: gather in a single database the bids, bidders, and results of every public works tender your agency has run over the past five years, with each bidder’s tax ID. Without that history, no algorithm has anything to compare against.

What Brazil already uses: CADE’s Cérebro and CGU’s Alice

CADE’s General Superintendence created Project Cérebro (“Brain”) in 2013 to develop screening and data mining tools that point to suspicious behavior in public procurement, and the agency has shared its methods with competition authorities in Switzerland, England, and the Netherlands. In September 2021, CADE opened an administrative proceeding against 42 companies and 43 individuals over a suspected cartel in public and private tenders for outsourced services. The original 2015 complaint concerned an online reverse auction for firefighting brigades. Cérebro’s tools helped identify companies the complaint had not named, and CADE found the group coordinating in at least 4,700 procurement processes, with signs of collusion in more than 50. The evidence was shared with the Federal Police, and in 2021 the final decision still rested with CADE’s Tribunal.

Alice, short for Analyzer of Tenders, Contracts, and Calls for Bids in Portuguese, works before the contract is signed. In 2024, according to the CGU, it reviewed 118,069 procurement processes, and its alerts led to 206 preventive audits covering purchases worth R$25.57 billion. The completed audits saved R$257.25 million. That year, 126 public bodies were registered to receive its alerts: 32 federal, 33 state and Federal District, and 61 municipal.

A highway project shows the tool’s reach. In May 2024, an Alice alert led the CGU to audit two online reverse auctions run by the National Department of Transport Infrastructure (DNIT), the federal highway agency, for road safety devices and signage, with estimated values of R$2.24 billion and R$1.78 billion. The audit examined price formation and quantities in the reference budget, and the adjustments saved R$118.6 million, about R$112 million of it in the overhead and profit rate applied to materials such as metal guardrails. In that case, the problem lay in the agency’s own budget, and the screening corrected an error by the buyer.

What to do: ask the CGU, or the audit body that oversees your agency, how to receive alerts from this kind of tool, and decide, before the first alert arrives, who receives each one and how quickly they must respond.

What an agency needs beyond the algorithm

The tool shifts the auditor’s work and still depends on it. In a 2022 study published in the CGU’s technical journal, two of the agency’s auditors concluded that Alice meets its goal and that its financial benefits exceed its costs. The same study found that 76.4% of the alerts reviewed were classified as unfounded, and each of them took up audit time.

Decarolis and Ergelen list what an agency needs to use AI against corruption and collusion: data from multiple sources gathered in a central platform, cleaned and classified; a team with data skills throughout the process; rules that provide for the authorities’ use of the model; and cooperation with agencies in other countries to share methods and datasets. The Italian case they cite started with access: the authorities opened their data to researchers, and the research gave them a roadmap of which techniques to apply.

Brazil’s public procurement law, Law 14,133/2021, already gives this work an institutional home. Its Article 169 subjects public procurement to continuous risk management and preventive control practices, organized in three lines of defense: civil servants and procurement officers in the first; the agency’s own legal counsel and internal control unit in the second; and the government’s central internal control body and the court of accounts, the external audit institution, in the third. An algorithmic alert serves all three, provided each knows what to do with it.

What to do: record the outcome of every alert (substantiated, unfounded, or inconclusive) and the reason, and use that rate to review the model or the screening criteria every six months.

AI on the bidder’s side

Contractors use similar techniques on the other side of the table. Bidder-side models account for each competitor’s backlog, regional experience, specialization by project type, and alliances with local partners, and one study cited by Decarolis and Ergelen reached 96% accuracy in predicting tender outcomes. These models guide three decisions: whether to bid at all, how to raise the chance of winning, including through a joint venture, and what markup to apply without losing competitiveness.

Ergelen describes the case of a firm that was reluctant to bid for a consulate project but decided to compete when a forecasting model showed that its chances rose with a partner that had already worked for the same client. The joint venture also used the model to set its markup and submitted the lowest bid. Large language models also help firms read tender documents and specifications in other languages, with legal and technical staff reviewing every output before a bid goes in.

A bidder’s existing workload affects its price. Some banks prequalify only contractors whose ongoing commitments are below 1.2 times their annual turnover. In one study, Ergelen found that competitiveness rose by 24% on average as that ratio fell from 1.2 to 0.7. For the buying agency, bidders’ committed capacity is a data point that helps predict the quality of the competition. Agentic AI in procurement shows the same shift on the buyer’s side.

What to do: if your firm uses AI to prepare bids, document which data feed the model, where they come from, and who reviews the final price. That record supports the bid if a price calibrated by an algorithm triggers the agency’s screening.

When the tool helps collusion

The same technique that hunts for cartels can organize one. Decarolis and Ergelen note that in frequent, similarly structured tenders, AI can help colluding firms plan a rotation of wins and use shell companies to simulate competition, which makes detection harder. There are other risks: opaque decisions based on systems no one can explain; barriers for smaller firms without access to the tools; and the exchange of sensitive project data between building information modeling (BIM) and AI platforms, with incompatible formats and security exposure.

In Brazil, Article 337-F of the Penal Code, added by Law 14,133/2021, sets a prison term of four to eight years, plus a fine, for anyone who frustrates or defrauds the competitive nature of a tender to gain an advantage from the award. The text describes the conduct and its result without naming the means, so it also covers collusion planned with the help of a model.

On the oversight side, the recommendation is to keep final decisions under human review and never impose penalties on the basis of a statistical red flag alone. Tender design, from the auction format to qualification and transparency requirements, also creates or removes opportunities for coordination.

What to do: when revising your standard public works tender documents, assess the auction format and qualification criteria with your internal control unit, alongside the patterns your screening looks for. In companies, add a rule on the use of AI in bids to the compliance program, prohibiting the use of competitor data obtained outside public sources.

Conclusion

Algorithmic screening already works in Brazilian public procurement: CADE uses it to find cartels, and the CGU uses it to correct tender documents and budgets before contracts are signed. An alert is only worth something when the agency keeps its tender history in a single database, has a team that handles each alert on a deadline, and records the outcome to adjust the model. Bidders use AI too, and tender documents need to be designed with that in mind.

The risk continues after the tender. Law 14,133/2021 lists among the goals of public procurement the prevention of overbilling during contract execution, and that is where it shows up: in amendments, progress measurements, and payments. In public works, execution is already a known weak point, as the picture of stalled works in Brazil’s government transition shows.

Converge Gestão de Contratos® brings each contract’s term, amendments, sign-offs, budget, and payments into a single workflow, with alerts before every deadline, automatic reconciliation, and an audit trail. To organize the public works contract portfolio of your agency or company, talk to Takter.

Sources

• Francesco Decarolis and Emrah Ergelen, How the Construction Sector Is Using AI to Cut Waste and Fraud, Harvard Business Review, April 2025.
• CADE, CADE launches proceeding to investigate cartel in outsourced services procurement, September 2021.
• CGU, Órgãos federais economizaram em 2024 257 milhões de reais com apoio do robô Alice da Controladoria-Geral da União (Federal agencies saved R$257 million in 2024 with the support of the CGU’s Alice), December 2024 (in Portuguese).
• Agência Gov, Atuação preventiva da CGU em dois pregões do DNIT gera economia de R$ 118,6 milhões (CGU’s preventive work on two DNIT tenders saves R$118.6 million), May 2024 (in Portuguese).
• Daniela de Quadros Dantas and Leandro Barbosa Martins, Estudo sobre a eficácia e a eficiência do uso da ferramenta Alice (A study of the effectiveness and efficiency of the Alice tool), Cadernos Técnicos da CGU, November 2022 (in Portuguese).
• Law 14,133/2021, Brazil’s public procurement and contracts law, arts. 11, 169, and 178, April 2021 (in Portuguese).