Why sampling checks come too late
A sampling check looks at part of the procurement and comes after payment. By then the contract is done or failed, the money is gone, and it is recovered through claims and courts.
A sample misses a lot simply because of volume. Brazil’s ALICE robot analyzed 190,923 procurements in 2023: a flow like that can be checked by hand only by sampling, and a violation often stays outside the sample.
The cost of being late shows in pandemic payments in the USA. The PRAC committee reviewed a sample of 662,000 records from 67.5 million claims and found about 1.4 million questionable Social Security numbers. If claims had been checked against data before payment, the committee estimates, about $79 billion of possible fraudulent payments could have been prevented.
There is a second loss: reviewers’ time. A sampling check spends it on documents that are in order, because it does not know in advance where the risk is.
The third problem is predictability. The plan of sampling checks is known in advance, while a full check is the same for every procurement, and nothing can slip out of it by chance.
procurement, contracts, payments, registries
red flags for every record
people, companies, accounts in one graph
risk amount and strength of signs
documents and links for the case
How continuous monitoring works
Continuous monitoring follows the same five steps as a one-off check, only every day and over all records:
1. Data: the system collects new notices, records, contracts and payments right after publication.
2. Risk indicators: each record passes a set of rules; ALICE has about 40 types of risk.
3. Links: new participants are matched against company registries and known cases.
4. Priority: each case gets a score based on the amount and the number of signs triggered.
5. A case for the reviewer: only what is above the threshold enters the queue, together with the documents.
Working systems are built this way. ALICE of Brazil’s audit body (CGU) has collected more than 500 procurement notices a day since 2015. Since 1 August 2025, Russia’s Federal Antimonopoly Service (FAS) has used the state system Anticartel: according to the agency, it analyzes data from the Unified Information System for procurement and the state Torgi system every day with elements of machine learning, covering 100% of procurement procedures. FAS launched the first module of its predecessor, the Big Digital Cat system, in autumn 2019.
Until a signal enters the queue, the work runs without people. The analyst does not export data or run checks by hand: in the morning there is a ready list.
Each step, from data to a case, is covered in detail on the page about how AI finds violations.
A queue of signals instead of a queue of documents
With manual control, the analyst works through a queue of documents: opens each one and looks for what is wrong. With monitoring, they work through a queue of signals, and each signal already explains which sign was triggered and where to look.
A signal contains the procurement, the triggered indicators, the amount at risk, the links found and links to documents. The analyst chooses one of three options: take it for review, close it with an explanation or put it aside until new data arrives.
The machine joins signals on one supplier or buyer into one case. The analyst sees not a dozen separate flags but one picture with a history.
South Korea has run such analysis since 2006. The BRIAS system of the Korea Fair Trade Commission (KFTC) collects data on public tenders above thresholds, from KRW 100 million for goods and services, and ran 40,170 analyses in 2023 against 25,520 in 2018. Since December 2023, more than 700 organizations must send data to it.
The KFTC states plainly that BRIAS does not prove collusion but gives investigators important leads. A signal means risk, not proof of a violation; the reviewer, investigator or court decides.
Every closed signal leaves a trace: who reviewed it, what they found and why they closed it. Over time these notes make the rules more precise, and empty signals are removed.
Reacting before payment, not after
The main gain of monitoring is reaction time. If a signal comes on the day the notice is published, the tender can be suspended, the terms corrected and the payment held until the check.
The signal goes to whoever can stop the process: the buyer, the audit body or the treasury, depending on the stage where it arose.
Brazil shows the scale: in 2019–2022, after checks with ALICE, tenders worth more than BRL 9.7 billion were suspended or canceled. Stopping a tender before the contract costs less than recovering money after payment.
The same logic works for payments. PRAC’s conclusion on pandemic aid in the USA is about exactly this: checking claims against data before payment, not after, could have prevented about $79 billion of possible fraudulent payments.
Reacting before payment does not mean refusing without review. The signal postpones the decision for the time of the check, and the supplier or recipient can give explanations and documents. If the sign is confirmed, the talk happens before the money moves, not in court.
How to count the money saved and reviewers’ time is covered on the impact page.
What the head sees every day
The head does not need every signal. They need a daily summary on one dashboard:
1. Coverage: how many procurements and payments were checked. For the FAS Anticartel system, according to the agency, it is 100% of procurement procedures.
2. New signals and the amount at risk in them and in cases under review.
3. The queue by analyst: how many cases each has and how many are overdue.
4. Review results: how many signals were confirmed and how many closed with an explanation.
5. New links of suppliers who have already been in reviews.
6. Signals that have sat without action longer than the set term: the head sees them at once, not in a quarterly report.
The share of confirmed signals shows whether the thresholds are set well. If few are confirmed, the rules are noisy and need review; if the queue grows faster than it is cleared, raise the threshold or add people.
The summary does not replace decisions. It shows where work is stuck and where to send people tomorrow, while conclusions on each case stay with the reviewers.
How many people you need for a thousand cases
You need as many people as there are signals above the threshold, not as many as there are procurements. The head sets the threshold: the queue gets as many cases as the team can review.
The scale of working systems: ALICE analyzed 190,923 procurements in 2023, and more than 500 auditors receive its signals. BRIAS in Korea ran 40,170 analyses of tender data in the same year.
That is why one analyst can handle thousands of cases. The machine keeps all cases under watch, and the analyst opens only those where the risk has grown. The rest stay in the model and move up the queue when new data arrives.
Besides analysts, you need an owner of the rules. They review false signals with the team and decide which indicator to fix; without this person, the queue turns into noise within a few months.
The launch order is simple. First one scheme and one group of procurements, then past data to tune the thresholds, then daily mode. This shows how many signals fall on one analyst before the system covers all procurement.
How the cases that sit in the model are structured is described on the case model page. How audit bodies in Russia and the CIS countries run similar monitoring is covered on a separate page about Russia and the CIS.