MMOOZZGG.com

AI against corruption

How to detect ghost employees on a payroll and among benefit recipients with AI

7 min read3 sources

Ghost employees appear when a payroll keeps people who do not work, have died or never existed, while someone else collects the money for them. To detect ghost employees, a machine matches every payroll record against population registries, bank and biometric data and shows the reviewer only the mismatches. A flag means risk, not proof: HR, an inspector or a court makes the decision.

On this page · 6 sections

In short

Ghost employees on a payroll and among benefit recipients become visible when you match the payroll against registries: a person has died, does not exist or is paid twice. AI matches the payroll against registries, biometrics and bank data and flags the mismatches. In Nigeria, such matching removed 23,846 non-existent workers from the payroll.

Where ghost employees appear

Non-existent recipients show up in three places, and each has its own data source for matching.

1. The payroll of a government body or state company: an employee quit or died, but the salary keeps being paid; or the payroll lists a person nobody has ever seen at work.

2. Social payments: benefits and pensions keep going after the recipient’s death or go to a made-up person.

3. Applications for state support: the application holds someone else’s or non-existent data, including data of dead people. In the US, the PRAC committee found more than 11,000 records linked to dead people in a sample of pandemic aid applications.

The sign in the data is the same: a record about a person exists in one system and is missing from another or contradicts it. An employee on the payroll who for years has no building passes, no logins to work systems and no vacations needs a check.

Each such record takes money from the budget and distorts headcount reports. So matching is done regularly, not once. How similar checks work in support programs is covered on the page about subsidy fraud.

Matching against registries and biometrics

The main tool here is not a model but record matching. Each line of the payroll or recipient list is matched against independent sources:

1. The population register and death records: the person was alive on the payment date.

2. The identity document and tax number: the number is valid, belongs to this person, and the name and date of birth match.

3. The bank account: the account belongs to the recipient, not to someone else, and it does not receive payments for several people.

4. Biometrics: fingerprints or a photo confirm that a living person stands behind the record and that there is only one.

5. HR and accounting systems: building passes, system logins, timesheets.

A machine is needed because matching runs across millions of records from several agencies at once. For example, in the US the PRAC committee sent the Social Security Administration the Social Security number, name and date of birth for each record in the sample. The Administration checked whether the number was valid, whether the name and date matched and whether the number belonged to a dead person.

AI adds fuzzy matching to exact matching. It finds one person under different spellings of the name, shared addresses and phones among different recipients, and groups of records created by one operator on one day.

Matching runs both ways. The payroll is checked against registries, and registries against the payroll: if a person who is still being paid a salary or benefit appears in the death records, the system flags it the same day, not at the next audit.

Nigeria: 23,846 workers out of thin air

The government of Nigeria matched the federal civil service payroll against biometric data and bank verification numbers (BVN). The bank number ties an account to a specific person, so it shows who really receives the money.

The result: 23,846 non-existent workers were removed from the payroll. The check began in December 2015, and by February 2016 the wage bill was already 2.29 billion naira lower than in December.

What matters in this experience for control:

1. Matching used two independent sources: biometrics confirm the person, and the bank number confirms the account.

2. The effect shows up in money quickly, in the next payrolls, not after years of investigations.

3. A record with a mismatch is not always fraud. The list may include people with errors in their data, so records are removed after a check, not at the moment of the flag.

The main lesson for control: a one-time cleanup is not enough. After it, matching is repeated at every payroll run; otherwise new non-existent records will appear again and stay unnoticed until the next big audit.

India: checking benefit recipients

India moved social payments to direct transfers to the recipient’s account (DBT) with identity checks through Aadhaar, the national ID. The system has worked since 2013.

The government estimates that cutting leaks saved 3.48 trillion rupees, based on data for 2009–2024. Aadhaar checks removed non-existent recipients, so coverage could grow without the same growth in spending. In April 2025, the Minister of Electronics and Information Technology cited an estimate from a policy brief: the number of recipients covered grew 16 times.

The lesson for control: matching against an identity register works as a filter at intake. A non-existent recipient does not get onto the list, and there is no need to search for them later.

Savings estimates for such programs come from the governments themselves. It helps to read them as estimates, not as results checked by an audit.

The experience of other countries where data matching helped find non-existent recipients is collected on the page about global experience.

USA: applications in the names of the dead

US pandemic programs paid money fast, and applicants were often checked after payment. The Pandemic Response Accountability Committee (PRAC) randomly selected 662,000 records from 67.5 million funded applications and checked them through the Social Security Administration.

According to the committee, more than 11,000 records in the sample were linked to dead people. The committee did not include them in the fraud estimate: some applicants may have been alive when they applied.

The overall result of the matching: about 1.4 million questionable Social Security numbers across all applications and about 79 billion dollars of possible fraudulent payments that a check before payment could have prevented.

Two rules follow from this example. First, match against death records before payment, not years later. Second, a match with death records is not proof, because the date matters: whether the person died before or after the application.

For a government body or state company, the same approach means matching the payroll and recipient lists against death records and registries before each payment run. An annual audit finds the same records, but by then the money has already been paid.

How to protect legitimate recipients

A check that wrongly removes living people from payments does no less harm than a ghost on the payroll: a person is left without a salary or a benefit.

Rules that protect legitimate recipients:

1. A flag does not stop a payment automatically. An employee checks the record and contacts the person.

2. The person gets a clear reason: which mismatch was found and which document will clear it.

3. There is a fast way to confirm yourself: come in person, repeat the biometric check or correct the data in the registry.

4. The system does not use citizenship, ethnicity or other traits that lead to discrimination as risk signs. What happens otherwise is covered on the page about the risks and errors of algorithms.

5. A log is kept: who removed or confirmed a flag, and why.

A flag means risk, not proof. A person decides to remove someone from payments or pass materials to an investigator, and in a dispute a court decides. In Nigeria, records were matched against two independent sources, biometrics and the bank number: this means fewer errors, but checks by people are still needed.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Africanews, 29 February 2016; Reuters in The Irish Times, 29 February 2016 (irishtimes.com)
  2. News On AIR, 21 April 2025 (newsonair.gov.in)
  3. Pandemic Response Accountability Committee, 4 June 2025 (pandemicoversight.gov)

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