480 million pounds in a year
The Cabinet Office and the PSFA announced the figure on 24 September 2025: in the twelve months from April 2024, more than 480 million pounds did not reach fraudsters. The government called it a record.
More than a third of the sum, 186 million pounds, came from finding and recovering money that fraudsters received during the pandemic: small business loans and other support measures.
The rest came from routine data matching. More than 37,000 false claims for the single person discount on council tax were stopped, worth 36 million pounds. More than 2,600 people were removed from social housing waiting lists they had no right to be on. Wrongful pension payments of 68 million pounds were prevented, including payments after the recipient’s death.
The government itself gives these figures, and the announcement mentions no independent audit of the sum. For a leader, the structure matters more: the savings add up from many small matches, not from one high-profile case.
Checking a program before launch
The PSFA created the Fraud Risk Assessment Accelerator. The tool helps public bodies find fraud risks in programs, policies and payment schemes while they are still on paper.
The work runs in three steps. An employee uploads the program documents, such as grant terms and guidance. A language model drafts a risk assessment and breaks each risk into three parts: who could cheat, what they would do and what it would lead to. Then fraud experts review the draft.
According to pilot users, an assessment takes about half a day instead of several days. In September 2025, a minister spoke of an 80% cut in time while keeping human control.
The tool makes no decisions. In the algorithmic transparency register, the government writes that the output is not a final document, and a qualified assessor checks every risk line.
Before processing, the tool finds and hides personal data, and it deletes the files after the work. In May 2026 it had 143 users, and the target for the open testing stage is up to 1,000 assessments a year.
Weak points most often hide in the terms of grants and subsidies: who can apply, how they prove the right to the money and who checks the report. Signs of subsidy fraud in data are covered separately.
Companies dissolved to avoid paying
A separate part of the UK work concerns Bounce Back loans that small businesses took during the pandemic. The risk sign here shows in the companies register: a company with an unpaid or doubtful loan applies to be dissolved.
According to the government, its measures stopped hundreds of thousands of such companies from closing before they repaid the money. Once a company is struck off the register, there is no one left to collect the debt from, so the check is needed before closure, not after.
This took a law. The Directors Disqualification (Dissolved Companies) Act of 15 December 2021 allowed the disqualification of directors of companies that closed without insolvency.
For an analyst, this is an example of linking two data sets: a register of loans or payments and the companies register. When a debtor applies for dissolution, the system flags the event, and an employee decides whether to object to the closure.
The same register shows a second sign: the director of a dissolved company soon registers a new one in a similar line of business and receives public support again. The link is visible from the director details in the companies register, and an inspector assesses it.
Sharing the tool with other countries
The United Kingdom plans to license the tool abroad. This was announced at an anti-fraud summit of the Five Eyes alliance countries in London, and, according to the government, alliance partners are considering adopting it.
At home, the tool is available to all public bodies as a cloud web app. The government published guidance on it on 20 February 2026 and updated its entry in the algorithmic transparency register in May 2026.
A public record of the algorithm is useful to other countries too. It states which documents the tool receives, how it hides personal details, who checks the result and why it makes no decisions. Such a description is a convenient starting point for a conversation with lawyers and the security team.
The tool itself runs on a foreign vendor’s cloud language model. So the method is what should travel, not the product: a draft risk assessment, the breakdown of each risk into “who, what, with what result” and a mandatory expert review. A model for this work can run on your own servers.
How similar systems work in other countries is collected in the review of global experience.
What Russia can take from this
1. Check a program before launch. The rules for granting a subsidy or grant can be analyzed by a model just as the PSFA does: which conditions are easy to meet with a false document, where there is no check against a register, who checks the report.
2. Match the registers of different agencies. Much of the UK savings came from such matches: pensions against death records, loans against the companies register, benefit claims against the right to them.
3. Write down the human role. The UK tool only prepares a draft, and an expert signs off the result. This principle should travel together with the method; otherwise a model error becomes the agency’s error.
In Russia, the data for such matching already exists: the dissolution mark in the unified register of legal entities (EGRUL), death records in the civil registry, and registers of support recipients. Matching needs no new data, only links between systems and rules for who reviews the signals.
You can start with one subsidy program: the model analyzes the granting rules, an expert selects the real risks, and by the program’s launch the agency has a list of checks. How to plan such a pilot is described on the rollout page.