MMOOZZGG.com

AI against corruption

AI corruption detection: from a red flag to a case for the auditor

8 min read5 sources

AI corruption detection starts with data, not with a model: procurement records, contracts, payments and company registers are brought together in one place. Then the machine marks risk indicators, finds links and ranks cases by amount and risk, and the auditor receives a ready case with documents. A flag at any step means risk, not proof of a violation.

On this page · 6 sections

In short

AI corruption detection works in five steps: gather the data, mark risk indicators, find links, set priorities and build a case for the auditor. The machine checks all records, and a person makes the decision. The World Bank’s ProACT platform collected open procurement data from 110 countries with risk indicators for them.

Step 1. Bring all the data into one place

A sign of a violation is almost never visible in one table. To notice a shared phone number between two bidders, you need to join the procurement register and the company register; to notice a payment without delivery, you need to join contracts, acceptance reports and payments.

The basic data set for checking procurement and spending:

1. Procurement notices, protocols and bids with prices.

2. Contracts, amendments, acceptance reports and delivery dates.

3. Treasury or bank payments: who paid whom, how much, when and for what purpose.

4. Company registers: directors, founders, addresses, phone numbers, registration dates.

5. Information about the buyer’s officials, if it is lawfully available to the auditor.

Sources can be open or internal. The World Bank’s ProACT platform collected open procurement data from 110 countries for 2006–2020, and Brazil’s ALICE robot has picked up more than 500 new procurement notices every day since 2015.

Matching reference lists takes the most time at this step. One supplier must have the same name in the procurement record, the contract and the payment; otherwise links break at the next steps and the scheme hides in typos.

Data is collected regularly, not as a one-off export: new procurement and payments are loaded every day, or the picture is out of date by the time of the audit. Arrange access in advance: who provides the exports, in what format and with what limits on personal data.

How a signal reaches the inspector
  1. Data

    procurement, contracts, payments, registries

  2. Risk indicators

    red flags for every record

  3. Links

    people, companies, accounts in one graph

  4. Priority

    risk amount and strength of signs

  5. Case file for the inspector

    documents and links for the case

Step 2. Mark the risk indicators

A risk indicator works like a rule that the machine applies to each record. The Open Contracting Partnership described 73 red flags for all stages of procurement: planning, tendering, award and contract performance.

Examples of indicators from the World Bank’s ProACT platform:

1. A tender has only one bidder.

2. A short deadline for submitting bids.

3. The first digits of amounts deviate from Benford’s law, which natural amounts follow.

4. One supplier has a large share of the buyer’s purchases.

5. The supplier is registered offshore.

6. The contract is delivered late, and its price grows after signing.

Indicators can be simple or composite. A simple one looks at one field, such as the number of bidders. A composite one compares a purchase with similar ones: the price with prices for the same product code, the bid deadline with this buyer’s usual deadline.

ALICE in Brazil checks each purchase against about 40 types of risk. Each triggered indicator is saved with a link to the document, so the auditor sees why the case is on the list.

A detailed list of signs by procurement stage is on the page about red flags.

Step 3. Find links between people and companies

Indicators look at one purchase, while links show the relations between participants. The machine builds a graph: its nodes are companies, people, phone numbers, addresses and accounts, and the links are labeled: “manages”, “same number”, “signed”, “transfer”, “bidder”.

What the graph flags: two bidders in the same tender with a shared director or phone number; a supplier linked to an employee of the buyer; a chain of transfers through which money returns to a party to the contract.

Colombia showed the scale of such matches. The Oceano platform, working on 8,678,464 public contracts from 2014–2019, flagged 54,155 cases of one representative in several companies and 8,749 companies with one phone number.

Each such link is not a violation by itself: firms may share an accountant, and a director may run two companies. It is a reason to check whether these companies competed with each other in the same tenders.

The graph also finds indirect links. If there is no shared director, the machine follows a chain of two or three steps through a founder, an address or a subsidiary. By hand, such chains are traced one case at a time; the machine traces them for all participants at once.

How links come together in the scheme of a specific case is covered on the page about the case model.

Step 4. Set priorities by amount and risk

There are always more alerts than people. So the machine gives each case a score: how many indicators fired, how strong they are, how much money is at risk and whether there are links from step three.

The queue is sorted by score and amount. Large purchases with several signs at once are at the top, single weak flags at the bottom. They do not disappear: they are kept in the history and move up if the same supplier gets new ones.

The amount matters in the score, but it does not decide everything: small purchases from one supplier by one buyer can add up to a large sum. So the machine calculates risk both for a single purchase and for the buyer-supplier pair over a year.

Prioritizing pays off. A study of Brazilian municipalities (American Economic Journal: Economic Policy, May 2025) showed that if municipalities are chosen for audit by a model built on budget data, auditors find almost twice as many municipalities with corruption for the same number of audits as with random selection.

The manager sets the thresholds: the queue should hold as many cases as the team can review. In Brazil, more than 500 auditors receive ALICE alerts.

How the risk score works and how to check that it works is described on the page about risk scoring.

Step 5. Build a case for the auditor

In the end the auditor gets not a flag but a case: which indicators fired, on which documents, which links were found, how much money is at risk and on what dates the decisions were made.

A good case answers three questions: what exactly is wrong in the data, where it shows, and what to check by hand first. Each answer comes with links to the documents, so the auditor does not have to search for them again.

Agree on the case format with the auditors in advance: a summary page, a table of signs, a link graph and a timeline. Then a case can go to a colleague or to investigators without retelling it.

Then a person takes over: requests documents and explanations, visits the site, and compares the data with what they see on the ground. The machine does not do this part.

After the review, the auditor records the outcome: the violation is confirmed, there is an honest explanation, or there is not enough data. These marks go back into the system, and the thresholds become more accurate over time.

The decision is made by the auditor, the investigator or the court. The system gives an alert for review and collects materials, but it does not conclude that a violation happened: in the words of the Open Contracting Partnership, a red flag means risk, not proof.

Where AI makes mistakes

The machine’s errors are predictable, and you account for them from day one. Here are the main ones.

1. False alerts. A single bidder in a tender may have an honest reason, such as a narrow market. If such cases are not reviewed and the rules are not fixed, analysts get used to skipping alerts.

2. Missed cases. The model sees only what is in the data: it will not notice cash payments, verbal agreements or documents outside the system.

3. Dirty data. An error in a tax ID or different spellings of a name break links and hide the scheme, and duplicate records create false double payments.

4. Biased signs. If the model gets a sign that describes a person rather than a deal, such as origin, it starts to suspect whole groups of people. An indicator must describe behavior in the data.

5. Outdated rules. ProACT indicators such as the bid deadline or the growth of the contract price are revised when the law or the market changes.

Quality is checked all the time. Once a quarter, take a sample of closed alerts and see how many were confirmed and how many violations audits found without any alert.

So a flag remains an alert for review, not an accusation. How to account for these errors during rollout is covered on the page about risks and errors.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. OECD OPSI, Robot Alice, 24 June 2024 (oecd-opsi.org)
  2. El Espectador, 25 October 2019 (elespectador.com)
  3. Ash, Galletta, Giommoni, American Economic Journal: Economic Policy, May 2025 (aeaweb.org)
  4. OCP, Red flags in public procurement, 2024, 2024 (open-contracting.org)
  5. World Bank Blogs, 30 June 2021 (blogs.worldbank.org)

Found a mistake? Tell us

Publications

For government bodies

We build such systems for government bodies and state companies, from data analysis to a pilot on one scheme. The data stays on the organization's own servers.

Contact us