AI won’t fix corruption in public procurement, but it may help you catch it

There is a reason public procurement is the government’s biggest corruption risk. Trillions of dollars are spent by governments through opaque, siloed and paper-based systems with no lever for effective oversight. But as public procurement moves from documents to data, we have a groundbreaking opportunity to transform both risk detection and how we act on those risks before procurement goes badly wrong.

Over the last decade, this problem has been increasingly recognized. The 2023 UN resolution, adopted by 193 countries, captures this shift, as governments agree to promote transparency and integrity in public procurement, followed by specific guidelines for using technology to combat corruption. 

So, on one hand, global norms and guidance are moving in the right direction: recognizing corruption as an issue to address, away from high-risk price-only selection criteria and single bidding, and placing a stronger focus on quality. On the other: (slowly) increasing availability of data, especially as a result of the growing adoption of the Open Contracting Data Standard (OCDS), and a better technology backbone including e-procurement and public investment systems that provide a complete picture covering planning to implementation.

Where does AI come in? 

So let’s look at what AI can bring to the table as it enters public procurement as part of the tech stack. In July, we brought together a group of experts and public procurement professionals as part of the World Bank’s Public Procurement in Anti-corruption Working Group to discuss the opportunities of AI in fighting corruption and draw some lessons from country experiences and international best practices.   

Firstly, procurement agencies are applying different types of AI across the public procurement process: 

• Large-language models (LLMs) are being used to generate text from static knowledge, such as drafting standard RFP templates or extracting data from existing documentation.   
• Retrieval-augmented generation against a defined corpus of legislation is being used to check bids against compliance with national regulations, 
• Machine learning and algorithms help compare patterns and data, such as red flags and fraud detection.
• Still in exploratory mode, we see how AI agents may use tools and software to facilitate the procurement process, auto-uploading contract documents or e-mailing winners. Self-directing AI could take the step even further by setting goals to screen sanctioned companies or fulfill sustainability plans. 

AI can now read and make sense of messy, unstructured documents. This is, in essence, about matching, finding patterns and risks in what exists. But this doesn’t mean we shouldn’t keep investing in specific, consistently collected structured data to measure how governments are buying and how the market is performing. Measuring is about designing systems to capture what matters and enable risk prevention strategies from the start.

Use cases for AI in procurement to fight corruption

What does this actually look like in practice? I’d like to highlight four main use cases.

1. Extracting, matching and linking data from structured and unstructured documents, including connecting data on asset disclosures, with campaign donations, and involvement of blacklisted firms.

In Kazakhstan, the tool Red Flags Management supports State Audit Institutions in prioritizing risky procurement transactions. It feeds unstructured text about a procurement transaction into a large language model (LLM) to extract specific data points, which are then processed using statistical methods and machine learning. It also uses tax data to detect shell companies, such as businesses with large sales volumes but no employees.

2. Flagging procurement anomalies based on corruption risk indicators to prioritize manual or automated audits. Remember, we’re talking about hundreds of procurement processes happening at the same time. 

Chile’s procurement agency, Chile Compra, runs an Observatory that emerged from the country’s legal framework mandating ChileCompra to monitor procurement processes, issue recommendations, and report compliance gaps to regulatory agencies such as the National Comptroller’s Office and Antitrust Authority. As a result, ChileCompra has implemented an integrity monitoring system that has evolved from manual expert reviews to robotic process automation and the development of LLM-based tools to detect large-scale compliance gaps, supporting coordination among regulatory agencies. Initially relying on manual expert reviews, who were only able to cover a small fraction of processes, ChileCompra has automated the detection of 15 types of compliance breaches, enabling real-time notifications and broader coverage. ChileCompra is also developing AI tools that use large language models to analyze unstructured documents and detect complex compliance gaps and infractions by encoding legal rules and prioritizing risks.

Similarly, Brazil is using ALICE (Análise de LICitações e Editais), an AI-powered automated oversight system developed by the Office of the Comptroller General of Brazil (CGU) to monitor public procurement, detect irregularities, and prevent fraud. It analyzes procurement processes daily and generates alerts when it detects risks. In 2023, the system reviewed nearly 191,000 processes and triggered audits linked to contracts worth over US$ 4.5 billion. 

3. Profiling supplier risks based on prior performance and external financial data, such as tax data, payroll records, and beneficial ownership registries.

Chile’s system integrates beneficial ownership data to identify potential conflicts of interest, cross-referencing public servant registries and contractor databases, with plans to include family member data in future stages. Using a hybrid approach, ChileCompra combines automation, predictive modeling, NLP, and human review, supported by robust whistleblower channels, to maximize coverage and continuously improve compliance monitoring.

4. Detecting bid-rigging patterns at scale, using advanced machine learning techniques that can dig across massive data sets more easily. Think multilayered price analysis and relationships among bidders and detecting evolving corruption schemes as loopholes are closed.

Countries like Chile are already exploring more advanced approaches, such as links between different companies. Tools such as Brazil’s MedicamentosTransparentes apply similar pattern-detection to medicine pricing. 

From detection to prevention

The best anti-corruption measures stop it before it starts. That means radically rethinking the procurement process and reducing friction for public officials as they navigate bureaucracy and systems.

Ukraine is currently reimagining how it manages public investment projects to rebuild the country while at war. With such a massive need and billions of dollars being spent rapidly on vital infrastructure, the risk of corruption is high, particularly as public officials are often uncertain about how to manage these large-scale projects. At the center of this reform is DREAM, the country’s open digital ecosystem for public investment management, which connects state-owned and external registers, supports project monitoring and data analysis, and creates a single digital pipeline for public investment projects. 

But more importantly, the DREAM AI embeds AI in the project management process from the start (using a RAG library), ensuring public officials have access to key legislation as it is updated and can address questions and potentially problematic issues as they emerge. An AI assistant handles user queries, supports feasibility studies, assists in project appraisal, and eventually models project impacts, all while maintaining human accountability and transparency.

Managing risks

But all of these use cases rely on data and AI models, which is where things can go wrong. 

• Automation & hallucination: Automating sensitive decisions raises questions about transparency, appeal rights & liability and hallucinations, particularly with LLMs.
  
  That’s why human oversight is critical. Humans need to stay in the loop for decisions. Where AI surfaces the risk or the anomaly, people investigate and act.

• Black box decision-making: Models can also provide results and information that may not be replicable, can’t be audited, or may be manipulated. Particularly when proprietary data is involved, oversight is limited.
  
  Models and algorithms need to be transparent, remain explainable, and clear reasoning needs to be established for AI system decisions, including audit trails & logging AI-driven actions.

• Poor quality & biased data: We’ve highlighted that there is increasingly more data available on procurement. But data quality is not perfect, and AI amplifies poor data. For example, models trained on low-quality, mislabeled or biased data will produce unreliable or harmful outputs.
  
  Data outputs are only as good as what you put in. Data standards, such as the Open Contracting Data Standard help structure that information. Structured data also helps measure more strategic, market-level questions that require data collected consistently. AI can’t conjure data that has not been recorded (unless it’s hallucinating)! 

• AI makes it easier to fabricate, falsify information & documents or impersonate executives, requiring strict access controls, digital identities and data governance standards. Participatory approaches such as public audits or crowdsourcing information can help mitigate this risk.

The promise of a new technology as a game-changer in the fight against corruption – and in development more broadly – is neither new nor a panacea, whether we were talking about Information and Communication Technologies (ICTs) or social media, open data or blockchain.

But these examples and our conversations with public officials also show that we can’t and shouldn’t do without technology; the integrity dividends from digital technologies are just too important, as highlighted well in a recent World Bank paper on GovTech for Public Sector Integrity. No one wants to go back to filing wheelbarrow loads of paper files to submit tenders, yet alone analyze them with a calculator in hand. 

Understanding AI’s strengths (and weaknesses) can help identify the digital solutions and tools that can deliver on what better procurement is all about: ensuring that public spending provides better services for people and communities.