Choose one scheme to start
Choose the first scheme by three conditions: its data already exists in electronic form, the risk signs can be calculated from this data, and the auditors know what to do with an alert. Most often this is procurement: bid rigging, split purchases, inflated prices.
Other countries started this way too. Brazil’s audit agency CGU has checked procurement notices with the ALICE robot since 2015, and Korea’s KFTC has analyzed public tenders with the BRIAS system since 2006. Both systems work with tenders, not with all spending at once.
A good sign for the first scheme: the auditors have already reviewed such cases by hand. Then you can compare the system’s flags with their findings right away and see where the model helps and where it errs.
Schemes that need data from other agencies you do not have, or where each case needs a site visit, do not fit the start. Add them later, when the review of alerts is already running smoothly.
The signs for each scheme are described in the section. For example, bid rigging shows in prices and the set of bidders, and split purchases in the dates and amounts of similar contracts.
Data: what you already have
For procurement schemes the base is already public: notices, protocols, contracts and performance records. ALICE works with exactly this kind of data: every day it collects more than 500 procurement notices and checks them against about 40 types of risk.
Inside the organization you need three sources: the contract register, payments and the list of counterparties. If they sit in different systems, an export to spreadsheets for the last one or two years is enough to start.
From public registers you add company details: founders, managers, addresses, registration date. Links between the buyer and the supplier show only when these details are joined with the contracts.
Before the pilot, check the data quality: are tax ID numbers written the same way, do all payments have dates, are any contract amendments missing? This check takes a large part of the time, and without it the alerts will be false.
A pilot on one type of violation
The pilot tests one scheme on your data in 14 days. The price is from $1,499, and the metric is written into the contract before the start.
1. The first days: access to the export for a separate account, and agreement on the risk signs with the auditors.
2. The middle of the pilot: the system goes through past procurement and flags alerts, and an auditor reviews a sample and marks which alerts are useful and which are false.
3. The last days: a daily check of new data and a report in numbers: how many records were checked, how many alerts came up, how many of them a person confirmed, and how much time the review took.
The pilot metric is counted in hours and alerts, not in violations found: an auditor or a court recognizes a violation, not the system. An example metric: the time to check one purchase before and after.
The Accounts Chamber of Russia shows time savings at scale: according to its deputy chair, a system that reads audit reports and payment documents saves tens of thousands of working hours.
How to procure the pilot and write the specification is described on the page on how a public body can order an AI project.
Who reviews the alerts
An alert without a person is useless. ALICE sends alerts to more than 500 auditors, and an auditor, not the robot, reviews each alert.
BRIAS does not prove collusion either. In the KFTC’s own words, the system gives investigators important leads, and they draw the conclusions.
The pilot needs one responsible employee: about half an hour a day for disputed alerts. After the pilot, roles go into a procedure: who receives an alert, how fast they review it, to whom they pass the materials, and who sees the log.
Alerts are sorted by priority: large amounts and several matching signs come first. This way one analyst reviews what matters most and does not drown in low-risk flags.
False alerts are part of the work too. The auditor closes them with a short explanation, and these explanations become material for refining the signs and thresholds.
Who sees the alerts and how an employee learns the reason for a flag is covered on the page on risks and errors.
Data protection and AI requirements
Procurement and payment data is processed on the organization’s servers or in a data center in Russia. The model runs inside the closed environment, and the data is not passed to the vendor.
Information about people (company managers and founders, recipients of payments) falls under the personal data law. So the pilot runs on open, anonymized or already approved data, and the security service approves the list.
Russia’s Federal Law 243-FZ on supporting the development of AI technologies came into force on 1 September 2026, and the rules on cases where only sovereign or national models are allowed apply from 1 March 2027. For systems where such models already run on that date, the limits do not apply until 1 September 2032, if the data is processed and stored in Russia.
The system’s account gets read-only rights to the data it needs. All actions are written to a log, and the access, the data and the system itself stay with the customer after the project.
A detailed review of the law and of what to check before the pilot is on the page on AI requirements in the public sector.
How to scale
After the pilot, add schemes one at a time: payments to tenders, company links to payments. An AI department of 3–5 agents, one per scheme, is built in 4–6 weeks and costs from $4,999, and the cost of the pilot counts toward it.
The volume grows step by step. BRIAS ran 25,520 analyses in 2018 and 40,170 in 2023, and since December 2023 more than 700 organizations must send data to it.
In 2023, ALICE analyzed 190,923 purchases. In 2019–2022, tenders worth more than BRL 9.7 billion were suspended or canceled in Brazil.
Russian audit bodies are taking the same path. The Accounts Chamber announced the launch in 2026 of a state digital audit system with AI for risk-based auditing.
Support after launch costs from $480 a month: updating the signs, checking model quality and reviewing false alerts.
You can start with a free 30-minute review: we discuss which scheme fits the start and what data you already have for it. The formats of work are described on the page on consulting and development for public bodies.