About abilities
AI does what people have no time for: it checks the whole flow of procurement and payments every day. The Brazilian robot ALICE collects more than 500 procurement notices every day and checks them against about 40 types of risk.
It sees signs a person would miss in a table: the same participants in different tenders, a shared phone number among competitors, a price that rises after the contract is signed, a link between a buyer and a supplier through company managers.
AI does not decide whether there was a violation, and it does not replace the inspector. Its work ends with a signal and a reason: which sign fired and on which documents.
The main thing the machine gives is a case model: people, companies, accounts, contracts and dates are linked in one graph and updated as new data arrives. A person cannot hold a thousand links in their head, but the system collects them in minutes.
Signals are sorted by risk, so one analyst can handle a large flow: they see large sums and matches of several signs first, not thousands of small marks.
How these steps add up to a chain from data to a case on the inspector’s desk is shown on the main page of the section.
About data
For the first scheme, what you already have is usually enough: a contract register, payments, a counterparty directory and open procurement data. ALICE works on procurement notices, that is, on open data.
Company details come from open registers: founders, managers, addresses. Without them, the links between a buyer and a supplier are not visible.
Data quality matters more than volume. If tax IDs are written in different ways and some payments have no dates, the system gives false signals, so before the pilot the data is checked and brought to one format.
You do not have to send the data to a contractor or abroad. The models run on the organization’s servers, as described in the section on sovereign AI.
About errors and responsibility
There will be errors: the system will mark an honest purchase or miss a violation. So a mark is not a decision but a reason to check the documents.
The cost of wrong rules shows in the case of the Netherlands. The tax authority’s model used nationality as a risk factor, tens of thousands of parents were falsely accused of fraud, and the scandal led to the government’s resignation.
To prevent this, banned features are excluded before launch, each mark is linked to a document, and a person has the right to know the reason. These rules are covered in detail on the page about risks and errors.
The model itself is also checked: every quarter, the share of false signals is counted, and on past inspections its selection is compared with random choice and with current practice.
Responsibility stays where it is without AI: with the inspector who made the decision and with the manager who approved the rules for handling signals. The system is responsible only for making the reason for each mark visible.
About adoption
You start with one scheme and a pilot on your own data. Even large systems rest on one clear flow: since 2015, ALICE has worked with procurement notices and sent signals to more than 500 auditors.
A pilot needs a data export, a list of signs agreed with the inspectors, and one responsible employee to review disputed signals.
The rules for handling signals are written before launch: who sees the signals, how soon they are reviewed, and what an employee learns about the reason for a mark. Without these rules, even an accurate system meets resistance.
Scaling comes after the pilot, one scheme at a time: payments are added to tenders, and company links to payments. The effect is measured on your own data: how much was stopped before payment, how much was returned and how many hours were saved.
The timing, the pilot setup and the order of scaling are described on the page about how to introduce AI control.