What a conflict of interest looks like in data
A conflict of interest arises when a public servant has a personal gain in a decision they make at work. In data it is a pair: the decision, and a link between the person who made it and the one who received the money or the right.
Decisions differ: a signature under the technical specification, a vote on a procurement committee, acceptance of work, the award of a subsidy, the issue of a permit.
Links differ too. Direct: the official or a relative owns the receiving company. Indirect: a shared address, phone, email, a former employer, or a shared representative in several companies.
What AI flags first:
1. A committee member and the owner of the winner match by surname, date of birth or address.
2. An official used to work at a company that now receives contracts from their department.
3. The receiving company’s phone or email matches the official’s contacts in HR data.
4. A company was registered shortly before the tender, and its founder is linked to an employee of the customer.
5. An official signs the acceptance of work from a company whose head is a relative named in the official’s declaration.
Colombia shows the scale of such matches. The Oceano platform of the Comptroller General processed 8,678,464 public contracts from 2014–2019. It found 54,155 cases where one representative was listed in several companies, and 8,749 companies with one phone number.
A match is not a violation in itself. It shows where to look: a company and an official with one phone number call for a question, not an accusation.
Lists of officials and owners
The check needs two data sets. The first answers who makes decisions: staff lists, committee members, signatures on documents, and declarations of income and interests if they are available in machine-readable form.
The second answers who stands behind the recipients: the register of companies with founders and heads, data on ultimate owners (beneficial owners), and the register of sole traders.
Decisions also need to become data. Committee minutes, acceptance orders and subsidy orders are often stored as scans. A language model extracts who signed and who voted, and only then does the check cover all decisions, not a sample.
The main difficulty is not the model but the matching. One person is written differently in different databases: with a typo, in another transliteration, after a change of surname. AI joins records by several features at once: name, date of birth, tax ID, address.
Family links are usually not visible in open registers. They come from officials’ declarations or an auditor’s request, and the system works only with what the law allows.
The European Commission uses exactly such sources in Arachne+: data on company owners, sanctions lists and data on public officials.
Direct and hidden links between a customer and a supplier are covered in more detail on the page on related parties.
EU: Arachne and beneficial ownership data
The European Commission runs the Arachne system to find fraud, conflicts of interest and irregularities in EU budget spending. The new version, Arachne+, counts 78 risk indicators; the previous one had 107.
The system answers whether a recipient of EU funds is linked to a person who influences the decision. To do this, it joins data on the ultimate owners of companies, sanctions lists and lists of public officials.
Arachne+ went live on 27 April 2026. From the 2028–2034 budget period, EU countries must send data to it, and the question of mandatory use will be reviewed in 2028 after a readiness assessment.
The European Commission states plainly that the system makes no automatic decisions. It suggests which project to check, and national bodies and auditors do the checking.
A signal from such a system leads to an ordinary audit: the auditor looks at which indicators fired, requests documents on the recipient’s owners and decides whether an on-site visit is needed.
The cut from 107 indicators to 78 also tells a story. More flags do not mean more accuracy: extra signals take auditors’ time and hide the important ones.
How the European Union checks budget spending in general is covered on the page on the European Union.
China: Zero Trust and 8,721 officials
China tested a system called Zero Trust. According to the South China Morning Post, it compared big data on public servants in 30 counties and cities and flagged 8,721 officials.
Officials resisted the experiment. The report does not say how many flags audits confirmed, so the number shows the reach of the check, not the number of violations.
The practical lesson: a check of officials works when its rules are clear and approved in advance. Otherwise the system itself becomes the subject of a dispute instead of helping the audit.
For a public body, this gives two rules. The order of the check is approved and announced before launch. An official has the right to explain a match before anyone other than the auditor learns about it.
Where privacy begins
A conflict of interest check touches the personal data of officials and their relatives. The line is set by the purpose: the check covers only links with decisions the person makes at work.
Practical rules:
1. Check only data that the law allows you to collect for this purpose: declarations, registers, HR records.
2. Do not build a full profile of a person: the system looks for the overlap of a decision and a recipient, not every link it can find.
3. Keep the data inside the agency and log who opened a result and why.
4. Show a match to the auditor, not to the official’s manager or colleagues.
A false match on a namesake hurts the reputation of an honest employee. So a person checks every signal, just as in Arachne+, where the system makes no automatic decisions.
Why AI-based control raises concerns and how to answer them is covered on the page about why people fear AI. Model errors and their cost are collected on the page on risks and errors.
How to set up the check
The check is easy to build in four steps.
1. Collect the lists: who signs decisions on procurement, subsidies and acceptance, and who sits on committees.
2. Join them with the company register and data on the recipients’ owners.
3. Set up the signs: direct ownership, a shared address or phone, a shared representative, a former employer.
4. Send matches to the person responsible for corruption prevention, who requests an explanation and decides whether an audit is needed.
It pays to run the check before the decision, not after. The official declares the conflict and steps aside, and the tender runs clean. After the contract is signed, the same match turns into a long investigation.
A hypothetical example. Customer B announces a tender, and a head of department signs the technical specification. Company A wins, and the system notes that its founder is listed in the head of department’s declaration as a brother. The auditor requests an explanation and finds out whether the head declared the conflict. If the conflict was declared and the head took no part in the decision, there is no violation, and the signal is closed with a note.
You can start with one area, such as acceptance of work or one subsidy program. The Colombian example shows why order matters: across 8,678,464 contracts, matching phones and representatives run into thousands, and without a risk score an auditor drowns in them.
A flag means a risk, not proof. A phone shared by an official and a supplier may turn out to be the number of a business center. The conclusion about a violation is made by the conflict of interest committee, an auditor or a court.