What is a red flag
Open Contracting Partnership (OCP) defines red flags as indicators that help spot possible irregularities, corruption and abuse along the whole procurement process. They do not prove corruption on their own, but they signal a risk.
In its 2024 guide, OCP describes 73 flags: each has a definition, a procurement stage, the data fields it needs and a formula. The flags fall into groups: low transparency, weak competition, fraud, bidder collusion and a rigged process.
In its ProACT platform, the World Bank computes similar indicators from open procurement data of 110 countries for 2006–2020: a single bidder, bidding periods, Benford’s law, supplier share, offshore registration, delays and contract price increases.
Flags also work in government systems. Italy’s National Anti-Corruption Authority (ANAC) launched a red flag tool in 2022 that applies more than 70 indicators, 17 of them for procurement.
The calculation needs procurement data in machine-readable form: notices, bids with prices, minutes, contracts with amendments, payments and the company register for owners. For each flag, OCP lists the fields of the Open Contracting Data Standard, so you can check the data set before work starts.
One flag means almost nothing. A signal for a check comes from a combination: one bid, plus a short deadline, plus a winner who wins from this customer year after year. How flags add up to an overall score is covered on the page on risk scoring.
Planning stage
Risk is built in before the tender is announced. At the planning stage, the system looks at how the purchase was designed.
Flags of this stage in the OCP list:
1. The procurement plan or planning documents are not publicly available.
2. A single-source purchase without a tender bypassed the procurement plan.
3. The amount is set just under a threshold: there are clearly more purchases just below the competitive tender threshold than usual.
4. A purchase is split into parts so that each one passes without a tender. There is a separate page on contract splitting.
5. The starting price is clearly higher or lower than the average for this product category.
6. Bidder requirements or the technical specification narrow the pool of suppliers for no clear reason.
Here AI compares a purchase not with a rule but with the market: with thousands of similar purchases by other customers. A language model reads the technical specification and flags requirements that fit only one manufacturer.
Bidding
At the bidding stage, the flags concern how the tender was announced and who came to it.
1. The notice was poorly published or lacks key documents.
2. The bidding period is too short, or little time passed between the notice and the opening of bids.
3. Bidders’ questions went unanswered.
4. The fee to take part or to get the documents is high or grows from tender to tender.
5. Only one bid came in, or there are fewer bidders than usual for this category.
An example from abroad. In the Dominican Republic, the procurement agency DGCP monitors all tenders in real time with 21 red flags. According to OCP, the share of single-bid tenders there fell from 63% in 2020 to 60% in 2023, and more than 60 suppliers were barred for violations in 2023.
When there are several bids but no competition, collusion flags apply. They are covered on the page on bid rigging.
Choosing the winner
Here the flags look at the bids and the committee’s decision.
1. The winner’s price is very close to the starting price or above it.
2. All bids except the winner’s were rejected, or the lowest bid was rejected with no clear reason.
3. Bidders’ prices are suspiciously even: multiples of one sum, identical, or right next to the winner’s price.
4. Bids from different bidders look alike, or the bidders share an ultimate owner or a major shareholder.
5. A bidder’s contact details match those of an employee of the customer.
6. Bidders’ prices break Benford’s law: the frequencies of first digits differ from the natural distribution.
7. A late bid won, or the decision was made too fast or too slowly.
Some of these flags cannot be counted without AI. A language model looks for similarity between documents from different bidders, a graph of links shows shared owners, and statistics across all of a customer’s tenders show the price distribution.
This is how Brazil’s ALICE robot works: since 2015 it has processed more than 500 procurement notices every day and checked about 40 types of risk. In 2023 it analyzed 190,923 tenders, and more than 500 auditors receive its signals.
Contract performance
After signing, the flags track whether performance matches what was promised.
1. A large gap between the price at award and the final contract amount.
2. Amendments that raise the price or change the scope of work.
3. Payments under the contract exceed its amount.
4. A missed delivery, or a gap between the work done and the contract terms.
5. The winner hands all or almost all the work to subcontractors, and losing bidders are among them.
6. A small first purchase from a supplier, followed by much larger ones.
The World Bank’s ProACT also has indicators for this stage: delays and contract price increases.
This stage needs performance data: amendments, acceptance acts, payments. If open data ends at contract signing, AI works on the internal data of the customer or the audit body. How auditors check that a facility was built and works is covered on the page on ghost projects.
A flag is a risk, not proof
Open Contracting Partnership says it plainly: a flag means a risk, not proof of a violation. Every flag has honest explanations.
A single bid happens in a narrow market where there really is one supplier. A short deadline happens after an emergency. A price rise happens when materials get more expensive.
So flags work as a queue, not a verdict. The system ranks tenders by risk, an analyst checks the top of the list, an auditor requests documents, and the conclusion about a violation is made by the audit body, an investigator or a court.
A hypothetical example. Customer B buys equipment: one bid came from Supplier A, the bidding period was shorter than usual, and the price equals the starting price. Three flags lift the tender to the top of the list. The auditor finds that only one manufacturer makes the equipment and that the deadline was cut because of an emergency, and closes the signal with a note. The note goes back into the model, and it scores similar tenders lower from then on.
OCP advises deciding in advance what a user does when a flag fires: prepares a report, cancels the tender or passes it to another body. Without such a procedure, flags pile up and nobody works through them.
Flags need tuning to your own market. OCP warns that weights taken from someone else’s system may not fit. The results of each audit go back into the system, and it flags the next tenders more accurately.
The model draws no conclusions about people on its own. It shows tenders where a combination of signs is rare and explains which flags fired, so the auditor sees the grounds for a question.