MMOOZZGG.com

AI against corruption

Risk based inspection with AI: who to audit first

7 min read5 sources

Risk based inspection sends auditors where a violation is more likely, not by lottery or by calendar. A model gives each purchase, municipality, or payment recipient a risk score, and auditors start at the top of the list. On Brazilian budget data, this selection finds almost twice as many corrupt municipalities as random selection with the same number of audits.

On this page · 6 sections

In short

Risk based inspection decides whom to audit first: a model calculates a risk score from data, and a person checks the top of the list. In the Brazilian study, this selection finds almost twice as many corrupt municipalities with the same number of audits. A score means risk, not guilt: the auditor makes the decision.

Why random audits lose

A random audit is impartial but wasteful: an inspector is as likely to land on a clean place as on a risky one. When there are dozens of auditors and hundreds of thousands of purchases and payments, most audits go to cases with no violations.

Brazil shows the scale of the task well. Since 2015, the ALICE robot of the CGU audit agency has collected more than 500 procurement notices every day, and in 2023 it reviewed 190,923 purchases. Nobody can read such a flow by hand, so the machine checks about 40 types of risk and sends signals to more than 500 auditors.

Selection by risk changes the order of work: the machine looks at every record, and a person looks at the top of the list. The OECD case page on the system says that in 2019–2022, tenders worth more than 9.7 billion reais were suspended or canceled.

Random selection does not disappear. A small random sample is needed to check the model itself: if it reveals violations that the model rated low, the model is retrained.

Brazil: twice as many findings

The economists Elliott Ash, Sergio Galletta, and Tommaso Giommoni trained a model on budget data of Brazilian municipalities. It uses gradient boosting on decision trees: an algorithm that estimates the probability of corruption in a municipality from its budget items. The paper came out in American Economic Journal: Economic Policy in May 2025.

The main finding comes from a simulation of audit policy. If municipalities are chosen by the model’s forecast, the same share of audits finds almost twice as many corrupt municipalities as random audits. The same number of auditors, almost twice as many findings.

The authors compared the forecast with earlier corruption studies: they repeated those findings with the model’s score and extended them. This showed that the score measures the same thing that on-site audits found.

A caveat: this is a simulation, not a report on a running program. The result shows how much can be gained if the budget data is complete and the model is trained on cases that people have already checked.

The practical lesson for a government body: start with data that is already collected. Municipalities file budget reports anyway, and a model on that data adds no new duties for those being audited. Other work on predicting violations is collected on the research page.

Spain: a forecast for 1–3 years

Félix López-Iturriaga and Iván Pastor Sanz of the University of Valladolid trained a neural network on real corruption cases in Spain from 2000–2012. For each province, the model estimates the probability that cases will appear in one, two, and up to three years. The paper came out in Social Indicators Research in November 2017.

Among the features that the authors link to corruption: property tax, a sharp rise in housing prices, new bank branches, and new companies. Another factor is political: the longer one party stays in power, the higher the probability of corruption, whether or not it has a majority.

Such a forecast is for planning, not for accusation. The authors themselves write that a high probability does not mean corruption will actually happen. The forecast shows where to strengthen prevention and collect data in advance over the next year or two.

The difference from Brazil is the question the model answers. The Brazilian model says whom to audit now; the Spanish one says where risk will grow over the next 1–3 years.

What makes up a risk score

A risk score is built from features that can be counted from data. In procurement, these are red flags: one bidder, a short deadline for bids, a contract price that rises after the tender, a large share of one supplier at one buyer. Each flag adds to the score, and the model weighs them and gives a total rating.

The ALICE robot checks about 40 types of risk in each procurement notice. This is a rule-based score: each feature is described in words, and the auditor sees which one fired.

Machine learning models go further and find combinations of features that predicted violations in the past. In the Brazilian study these are a municipality’s budget items; in the Spanish one, taxes, housing prices, bank branches, new companies, and how long one party has been in power.

What must not be in the score: features that describe a person rather than an operation. Citizenship, ethnicity, and place of birth say nothing about a violation, but a model easily turns them into grounds for suspicion. The Dutch case in the next section shows how this ends.

One more rule: the model learns from past audits. If the same districts or groups were audited before, it will learn this habit as a pattern. So results of random audits are added to the training data, and the features are checked against the list of red flags.

Explainability: why the model decided this

Explainability means that each score has a reason the auditor understands: which features fired and with what weight. Without it, an employee gets a number and can neither check it nor object to it.

The Netherlands showed the price of an opaque model. From 2013, the tax authority used a risk scoring algorithm to check childcare benefit applications, and nationality was a risk factor in it. According to Amnesty International, applicants without Dutch citizenship got a higher score.

The algorithm was self-learning, and its errors repeated without real human control. The employee who reviewed a flagged application got no explanation of why the score was high.

The result: tens of thousands of parents, mostly from low-income families, were wrongly accused of fraud. Families had their benefits suspended and fell into debt. The scandal led to the government’s resignation in January 2021.

Those being audited also need the explanation. If an organization lands at the top of the list, it has the right to know by which features, and to present documents that clear them.

The lesson for a government body is not that models are dangerous, but that a score without an explanation cannot be checked. A signal card with features and documents makes an error visible before it hurts people. Other errors of control systems are covered on the page about risks and errors.

How not to turn a score into a verdict

A risk score shows where an audit is needed more, not who is guilty. The auditor decides on a violation, and in a criminal case, the investigator and the court. The system only puts a case in the queue.

After the Dutch scandal, Amnesty International set out requirements for government risk scoring algorithms. They fit any risk scoring:

1. Do not use data on nationality and ethnicity in risk scoring.

2. Before launch, carry out a mandatory assessment of the system’s impact on human rights.

3. Do not use closed self-learning systems where a decision strongly affects people’s rights.

The audit body adds its own rules to this. A signal does not become a basis for sanctions without a check of documents. Some audits are assigned at random, to compare the model against them. Model errors are recorded and reviewed in the same way as auditors’ errors.

The Brazilian experience shows that such rules do not hurt the result. ALICE does not punish by itself: it sends signals to more than 500 auditors, and they take the work from there.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Ash, Galletta, Giommoni, American Economic Journal: Economic Policy, May 2025, 2025 (aeaweb.org)
  2. López-Iturriaga, Pastor Sanz, Social Indicators Research, 2017 (Agencia SINC), 2017 (agenciasinc.es)
  3. OECD OPSI, Robot Alice, 24 June 2024 (oecd-opsi.org)
  4. Amnesty International, 25 October 2021 (amnesty.org)
  5. Ash, Galletta, Giommoni, A Machine Learning Approach to Analyze and Support Anticorruption Policy, American Economic Journal: Economic Policy, vol. 17, no. 2, May 2025 (aeaweb.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