MMOOZZGG.com

AI against corruption

Procurement fraud detection: how AI finds fake contracts and paper deliveries

7 min read3 sources

Fake contracts leave traces in data: a delivery certificate with no delivery, a company with no staff, a contractor who is no longer alive. Procurement fraud detection with AI matches each contract with certificates, payments and government registers and shows the reviewer only the mismatches. A flag from the system is a sign of risk, and a reviewer, an investigator or a court decides on a violation.

On this page · 6 sections

In short

Fake contracts show in data as a gap between documents and reality: the certificate is signed, but the contractor could not have done the work or died before the contract date. AI matches the contract, the certificate, the payment and registers and flags mismatches. In Colombia, such a check of 8,678,464 government contracts found 45 contracts with dead people.

Types of fake contracts

A fake contract is signed by all the rules, but there is no real delivery, work or contractor behind it. In data, such contracts fall into four groups, each with its own signs.

1. Delivery on paper only: the acceptance certificate is signed, but the goods never reached the warehouse, the work was not done, the service was not provided. Sign: the certificate leaves no trace in the warehouse records, the work log or the request system.

2. A contractor that does not exist: the individual has died, or the company was liquidated or could not operate on the contract date. Sign: the contract date is later than the date of death or of removal from the register.

3. A shell company: the legal entity exists, but it has no staff, assets or track record, and its address and phone number are shared with other suppliers.

4. Inflated volume: part of the work is real and part is added on paper. Sign: the volume in the certificates exceeds what the contractor’s capacity or the contract term allows.

The Colombian platform Oceano searched government contracts for exactly these signs: contracts with dead people, shared company phone numbers and the same representatives in different firms.

The fourth group overlaps with ghost projects, where a facility is paid for but not built or not working. Satellite images of sites and sensor data are covered there too.

Matching the contract, the certificate and the payment

The core check rests on three documents. The contract says what was bought and for how much, the certificate says what was accepted, and the payment says how much was paid and to whom. In an honest deal they agree on amounts, dates, volumes and the recipient.

The system matches them line by line and flags mismatches:

1. The certificate was signed before the delivery period began, or on a day when the site access log shows nobody there.

2. The total of payments exceeds the total of certificates, or a payment went out without a certificate.

3. The payment recipient does not match the contractor in the contract.

4. The volume in the certificates exceeds the contract volume, and there is no amendment.

A person does such a check selectively, because one contract takes hours. A machine goes through all contracts for a period and links the documents into one case model: who signed, who received the money, who runs the contractor company.

In its ProACT platform, the World Bank included execution delays and contract price growth after signing among its risk indicators. These signs also show only when you match documents over time, not in one contract.

If a payment repeats an invoice already paid or goes to someone else’s bank details, that is a different scheme: duplicate and false payments.

Fly-by-night firms and empty addresses

A company set up for one contract almost always has a thin history: registered shortly before the tender, no staff, no assets, no earlier contracts. Each of these facts sits in open or government registers.

Signs the system takes from registers:

1. The registration date is close to the tender or contract date.

2. The company has zero or one employee, while the contract needs a crew or a production site.

3. A mass registration address: dozens of companies are registered there.

4. A phone number, email or representative shared with other bidders.

The last sign stands out at the scale of a country. The Oceano platform found 8,749 companies with one shared phone number and 54,155 cases where one representative was listed in several companies. One representative appeared in 46 companies, and another 8,324 companies shared one auditor.

Each sign alone proves nothing: honest companies can also share an accountant or an address in a business center. The risk grows when the signs add up at a supplier with a large contract and a short deadline. How the system checks such links is covered on the page on links between procurement bidders.

Colombia: 45 contracts with dead people

Colombia’s Comptroller General analyzes the state’s contract and budget data on the Oceano platform. According to the newspaper El Espectador, the platform reviewed 8,678,464 government contracts for 2014–2019.

Among its findings were 45 contracts with people who had died by then. In essence this is a simple match: the contract date against the date of death in the register. Nobody does it by hand, because it means matching millions of records from different agencies.

The platform flagged other signs too. It found 119 cases where contractors worked with the state although they were barred from it, and 7,078 state debtors among contractors, 98% of them legal entities.

The Comptroller estimates that almost one peso in three in the country’s government contracts goes to networks of linked companies. This is not proof of collusion but a map of where to look: the networks show which contracts to check first.

What matters is how these figures are used next. A list of 45 contracts is not a verdict but a review queue: behind each one stand a document request, the buyer’s explanations and a decision by the control body. More about the platform is on the page about Colombia.

Register checks before payment

Finding a fake contract after payment is useful, but the money is gone and hard to get back. So the check moves to the moment before payment: the buyer or the treasury checks the contractor and the certificate against registers while the money is still in the budget.

What the system checks before payment:

1. The register of legal entities: the company is active, not being liquidated, and its director is the person who signed the documents.

2. Death records: the individual in the contract was alive on the signing date.

3. Registers of unreliable suppliers and restrictions: the contractor is allowed to work with the state.

4. Tax and pension data: the contractor has employees who could have done the work.

US experience shows the cost of skipping such a check. The Pandemic Response Accountability Committee (PRAC) estimated that checking applications against data before payment could have prevented about USD 79 billion of potentially fraudulent payments. In a sample of 662,000 records out of 67.5 million applications, the committee counted about 1.4 million questionable Social Security numbers.

A check before payment does not stop a payment automatically. The system sets a flag, and an accountant or a controller decides whether to request documents, delay the payment or pay.

What AI will not replace: site visits and expert review

Documents and registers show where the risk is. They do not show whether the goods are in the warehouse, the facility is built or the equipment works: you see that only on site.

So a signal from the system leads to a check by people:

1. The reviewer gets a case card: the contract, certificates, payments, contractor data and the reasons for the flag.

2. The reviewer asks the buyer for acceptance documents, waybills, photos and work logs.

3. If the mismatch has no explanation, a site visit, a stock count or an expert review is ordered.

4. The result of the check goes back into the system: confirmed and cleared flags help it set priorities more precisely.

A flag means risk, not proof of a violation. A new employee may have signed a certificate with a wrong date, and a company with no staff may have a lawful subcontractor. A reviewer, an investigator or a court decides on a violation, and the machine saves them search time.

Large systems work the same way. The ProACT platform shows risk indicators for procurement in 110 countries, not a list of offenders.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. El Espectador, 25 October 2019 (elespectador.com)
  2. World Bank Blogs, 30 June 2021 (blogs.worldbank.org)
  3. Pandemic Response Accountability Committee, 4 June 2025 (pandemicoversight.gov)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min

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