Optimization
Close the leaks in tax and spending before you raise a tax.
A government should close the leaks it can measure before it raises a tax or cuts a service. We would find them payment by payment and set the recovered money against every proposed rate change. HMRC estimates that 6.4% of the tax owed in the UK went uncollected in the 2024 to 2025 tax year. The IMF estimates that governments could get one-third more value from their spending, on average, by adopting best practices.
How can AI reduce tax leakage and wasteful public spending?
AI helps when it checks every payment against a tested model, instead of an audit sampling a few after the money has gone. We would combine budget, payment, tax and register data to find uncollected tax, wrong payees and programs that cost more than their peers for the same result. Microsimulation shows what a rule change would raise or cost. Investigators review every flag, and a named official signs off each action.
Budgets are debated before the leaks are counted
Every leak is paid for with other taxes, new borrowing or fewer services. We think closing leaks should be costed first, because it asks only for money already owed or already budgeted.
The European Commission puts the EU's VAT compliance gap at €128 billion in 2023, or 9.5% of the VAT due. The US Government Accountability Office estimates $186 billion in improper payments in fiscal year 2025, mostly overpayments. The IMF's October 2025 Fiscal Monitor finds that public spending falls short of the best results possible with the same resources by about 31% in advanced economies and 39% in low-income developing countries.
What we would model, and with which data
Published data comes first, and internal records come in only where they change the answer.
- Line-item budgets. The World Bank's BOOST program has released more than 40 such datasets.
- Tax gap estimates, which tax authorities such as HMRC publish every year.
- Payment, supplier and benefit records, to find duplicates and prices far above the norm.
- Registers of people and companies, linked under a data-sharing agreement, to check that each payee exists and qualifies.
- Tax and benefit rules, written as code.
How the model is built and checked
A model should pass a back-test before anyone uses it for a budget decision. A back-test runs it on a past year and compares its output with what audits confirmed. A microsimulation model applies tax and benefit rules to a representative sample of households or firms and adds up the results, so a rule change is costed person by person. Record linkage matches records about the same person or company across registers with no shared ID, which exposes duplicate payees. Anomaly detection ranks payments that look unlike their peers.
What decision-makers get
- A ranked list of leaks, each with its size, evidence and confidence range.
- Costed fixes, such as a match between two registers or a new payment control.
- A what-if model of tax rules, so a proposed rate cut can be set against what closing the leaks would recover.
- A scorecard each cycle that compares predicted and confirmed savings.
We would start with one leak, not a program for the whole government. A first project proves one leak with a named owner, such as duplicate supplier payments. We would build the model from published data and a sample of records, back-test it on a past year and hand over the scorecard. The public body then decides whether to continue.
How it runs safely
A wrong flag can harm a real person or company. The model never stops a payment or opens a case. An investigator reviews each flag, and a named official signs off any action.
- Data stays in the public body's systems, and personal details are masked before any language model sees them.
- Fairness checks compare flag rates across regions and groups.
- We log every run, so an auditor can rerun it.
- In the EU, the AI Act classes AI that judges eligibility for public assistance benefits as high-risk, and the GDPR limits decisions based solely on automated processing. This page is not legal advice.
Open-source building blocks
Open tools let a finance ministry inspect, rerun and keep its model.
- OpenFisca, an engine for writing tax and benefit rules as code. Its site lists tools built on it by Barcelona City Council and the French National Assembly.
- PolicyEngine, tax and benefit microsimulation models for the UK and the US.
- Splink, for record linkage across datasets that lack a shared ID.
- PyOD, a library of anomaly detection methods.
- Fairlearn, for checking whether a model's errors fall unevenly across groups.
Has extendfuture built this for a tax authority or finance ministry?
No. Optimization is a new extendfuture offering. Our production work so far is AI for companies, with accuracy measured as a number, human review and an audit trail.
Will this lower taxes?
Not by itself. Recovered money can fund lower taxes or better services, and elected officials make that choice.
Do you need personal data to start?
No. A first model runs on published budgets and tax gap estimates. Register matching comes later, inside the public body's systems.
How is this different from an audit?
An audit checks a sample after the money is spent. The model scores every payment, so auditors can focus where problems are most likely. The audit office still decides what to examine.
Talk to the people who build.
One call. An honest read on what AI can do for this, and the number it has to beat.