What a ghost project looks like
On paper, a ghost project looks like a full set: a contract, completion acts and payments. One thing is missing: a result you can see or measure on site.
There are three variants. The object does not exist at all. The object exists, but it was built earlier under another contract. The object is built but does not work: the equipment is not connected, or the building stands empty.
In its ProACT platform, the World Bank collected open procurement data from 110 countries for 2006–2020 and calculates risk indicators from it. Two of them concern performance directly: delays and contract price growth.
In its list of 73 red flags, the Open Contracting Partnership marks flags for the performance stage: failed delivery, payments above the contract amount and amendments that raise the price. Each one is calculated from data, without a site visit.
The data a check needs: a register of contracts with amendments, acceptance acts and payments, the coordinates of the object, satellite images from several dates and, if available, meter readings. The more sources are matched to one contract, the fewer false signals the inspector gets.
An inspector needs not a verdict from the machine but a short list of objects where the documents and the facts differ most. How separate signs add up to one case is explained on the page about the case model.
Price and deadline growth after signing
The first sign is visible even before a visit: the contract changes after signing. The price grows through amendments, the completion date moves again and again, or the scope of work shrinks while the sum stays the same.
Each change is legal on its own. A risk sign appears when there are many changes, they come one after another, and they repeat with one supplier or one buyer.
The Open Contracting Partnership includes in its list a large gap between the price at the award and the final contract amount. The World Bank’s ProACT calculates delays and price growth from the open procurement data of 110 countries.
AI compares a contract with similar ones: the same category of work, the same region, a comparable scale. If the price grew more than for its neighbors and the handover moved further, the project rises in the risk list.
Price growth after signing and an inflated price at contract award often go together. The second case is covered on the page about inflated prices.
Matching acts against the facts
The act says what was done. The payments say what was paid for. The data about the object says what exists on site. AI matches the three sources and looks for where they differ.
What the system checks first:
1. The act was signed before the materials were bought, or fewer materials were bought than the volume in the act requires.
2. The object was handed over, and soon after a new contract was opened for the same work.
3. The same volume of work was closed by acts under two contracts, with different buyers or in different years.
4. The object was accepted but not put on the balance sheet, not connected to utilities, or did not get a commissioning permit.
The Open Contracting Partnership has a separate flag for this stage: a gap between the work done and the contract terms. Language models read scans of acts and estimates and turn them into tables. Then the gap is calculated across all contracts at once, not for one folder.
Similar signs, but at the level of payments and deliveries with no real result, are covered on the page about fictitious contracts.
Images and sensor data
Images answer a question that documents cannot: is the object on site? The method: take satellite images for the contract’s coordinates before the work, during it and after handover, and compare them.
Images show that the site did not change after the handover date, that the structure stood there before the contract, or that the work is on a different site from the one in the documents. A computer vision model picks out such pairs of images, and the auditor looks only at them.
An example from abroad. In September 2025, the Commission on Audit of the Philippines (COA), the country’s supreme audit body, sent four fraud audit reports to the Independent Commission for Infrastructure. They covered flood control projects worth more than PHP 360 million.
At all four sites, the auditors were led to places that did not match the location in the contract. Historical satellite images showed that on the approved site of one project, a structure already stood on 29 February 2024, months before the notice to proceed of 23 April 2024. COA pointed to possible double counting of completed work and called its list of those responsible preliminary.
In December 2025, COA said it was preparing mandatory geotagging for all infrastructure projects of government agencies, local authorities and state corporations: photos with GPS coordinates and a time stamp. For checks with satellite images, the commission is working out an agreement with the country’s space agency.
Sensors give the same check for working objects. Water, power or traffic meters show whether an object is used after handover. Zero consumption six months after a pumping station opens becomes a signal for a check.
The project’s result, not a report about it
The contractor writes the project report. The result can be measured without the contractor: how many cars use the road, how many children study at the school, how much water the station supplies.
AI links the contract with result data from other systems: registers of objects, attendance records, meter readings and citizen complaints. If the object was handed over but the area’s indicators did not change, the project goes on the list for a spot check.
The earlier a signal appears, the less money goes into an object that will never exist. Since 2015, Brazil’s robot ALICE has reviewed more than 500 procurement notices every day, checked about 40 types of risk and sent alerts to more than 500 auditors.
According to the OECD, Brazil suspended or canceled tenders worth more than BRL 9.7 billion in 2019–2022. This is control before the contract, and it adds to the performance check rather than replacing it.
The model itself needs feedback. Each object checked on site becomes an example: the risk was confirmed or not. With these examples, the system picks the next projects for a visit more precisely.
Who decides on a check
A person decides on a visit, a check and a violation. The Open Contracting Partnership says directly that a flag means risk, not proof of a violation.
A gap can have honest reasons. The object was moved by agreement, clouds cover the image, the work was accepted in parts, or a sensor broke. The inspector requests documents and looks on site before drawing conclusions.
The usual order of work:
1. The system puts the project on the risk list and shows which signs fired: price growth, a shifted deadline, an image with no change.
2. An analyst checks the data and removes record-keeping errors.
3. The inspector requests documents from the buyer and, if needed, visits the object.
4. The audit body concludes whether there was a violation, and then an investigator or a court takes over.
This order protects both the budget and honest contractors: checks go where there are signs, not to everyone.