Why purchases are split
The threshold for a competitive procedure separates simple purchases from complex ones. Below the threshold, the buyer may purchase from one supplier or run a simplified procedure; above it, open tenders with more bidders are required.
The Open Contracting Partnership describes the risk this way: what should be one contract is divided into two or more parts, each below the threshold, so the contract can be awarded directly or with less competition. This is one of the 73 red flags in the OCP guide.
Why a split purchase harms the budget: without competition, nobody pushes the price down, and the choice of supplier stays with the buyer. So check splitting together with a purchase with an inflated price and with the share of a regular supplier at that buyer.
But not every series of small purchases means splitting. A need may arise during the year, money may arrive in parts, and different units may buy for themselves. The line between normal and risky is covered below.
Signs in the data
The OCP guide has several flags for splitting. They look at the same fields from different angles:
1. Two or more purchases by one buyer in one product category are announced within a short period, for example three months, and each is just below the threshold. The purchases and the buyer are flagged.
2. One supplier gets several direct contracts from one buyer, and their total over the period is above the competitive threshold. The buyer and the supplier are flagged.
3. Many of a buyer’s purchases cluster just below the threshold. You see this not in one purchase but in the distribution of all its amounts.
4. After a direct purchase, contract changes raise the amount above the threshold at which a competitive procedure is required.
The data you need: the procurement method, the subject with its classification code, the amount and currency, the date of announcement or signing, the buyer and the supplier. OCP adds that thresholds must come from local law: without them, the flag cannot be computed.
All these fields exist in procurement registers. The hard part is that the same need is described in different ways, and a group of contracts comes together only after the subject of each one is brought to a common form.
How AI groups similar contracts
Grouping takes three steps. First, AI sets the subject of each purchase: the classification code if it is given, and the meaning of the description if the code is too broad. A language model understands that “printer cartridges” and “office equipment supplies” can be one need.
Then the system builds groups: one buyer, a similar subject, a sliding window of dates, one or several suppliers. For each group it adds up the amounts and compares the total with the threshold.
Finally, the group gets a risk score. It is higher if the amounts sit close to the threshold, if there is one supplier, if the purchases came in a row days or weeks apart, and if the same need did not recur after the group.
An example from the OCP guide: a buyer announces a direct purchase of a product for USD 999, and two weeks later a second purchase of the same product for USD 995, with a threshold of USD 1,000. Both purchases and the buyer get the flag.
The reviewer gets groups, not a list of separate contracts: the total amount, the threshold on the date, the suppliers, the dates and links to the documents of each purchase. This way a split purchase is checked once as a whole, not in parts in different reports.
A made-up example: Buyer V signed several contracts with Supplier B within six weeks to repair one building, each just below the threshold. The system puts them into a group, shows the total amount, the dates and the estimates side by side and gives the group a high score. The reviewer decides whether this is one job or several.
Thresholds and time windows
The flag has two settings: how close to the threshold and over what period. OCP leaves both to the user: the window can be three months, and the distance to the threshold 1% or 2%, depending on the market and the procurement rules.
Thresholds change over time, and the system must know which threshold applied on the date of each purchase. Otherwise, after a threshold rises, old purchases will look suspiciously close to it, and new ones will drop out of the check.
A study by Jan Palguta and Filip Pertold (American Economic Journal: Economic Policy, May 2017) on procurement in the Czech Republic shows how thresholds change buyer behavior. After a reform that let buyers choose bidders themselves below the new thresholds, purchase amounts began to cluster just below the thresholds: more in construction and services, less in goods.
Among such purchases, the probability that a firm with hidden owners wins the contract tripled. For control, this is a signal: check clusters of amounts below the threshold together with the suppliers’ owners.
Timing matters in another sense too. Splitting is easier to stop before the second and third contracts are signed, so a check is useful on the day a purchase is published, not at the year-end audit.
Where the line between splitting and normal practice lies
Several purchases of one product in a year are normal in themselves. Paper is bought every month, repairs follow accidents, and program money arrives in tranches. The flag does not tell these cases apart; a person does, from the documents.
What the reviewer checks after a flag:
1. The procurement plan: was the whole need known in advance.
2. The justification for each purchase and the internal notes on the need.
3. The technical specifications: is this one job or different ones, one site or several.
4. The delivery dates: did the contracts follow each other for one job.
5. The suppliers: the same one or different ones, and are they linked.
If the need arose in parts and the documents show it, the flag is closed with a note. If the job is one but the contracts are several, the materials go to the internal control team. A flag means risk, not proof: the reviewer, a control body or a court decides whether a violation took place.
What to check in your own organization
You can run the first check on your own data, without a new system.
1. Export all purchases below the threshold for the last year: subject, classification code, amount, date, supplier, unit.
2. Group them by subject and supplier in a three-month window and add up the amounts.
3. Mark the groups where the total is above the threshold and each purchase is close to it.
4. Plot the distribution of all purchase amounts and look for a cluster just below the threshold.
5. For the marked groups, pull the procurement plan and the justifications.
If there are many groups, start with those where the supplier is one and the amounts are closest to the threshold. The result of the first check will show whether to run such a check continuously, so that each new purchase is matched with its group on the day it is published.
Open data helps you compare yourself with others. On the ProACT platform, the World Bank collected procurement data from 110 countries for 2006–2020 and calculated risk indicators from it, among them a supplier’s share at a buyer and bidding periods. For buyers who want to check more documents with the same people, there is a page on AI for control and audit, and a separate section on red flags covers the topic in detail.