Chapter 4: PolicyEngine - Proof of Concept

In September 2021, a website launched that let anyone in the United Kingdom design their own tax and benefit system.

The premise was simple: enter a policy reform—raise income tax, increase child benefits, introduce a carbon dividend—and see the effects. Not just abstract estimates, but specific numbers: the cost to the Treasury, the change in poverty rates, the impact on inequality. And then enter your own household circumstances and see what the reform would mean for you personally .^[1]

PolicyEngine UK was, as its creators claimed, "the world's first product allowing anyone to design policy reforms and see both the effects on specific households, and on UK-wide outcomes like poverty, inequality and the budget" .^[1]

Behind this launch was a frustration that had been building for years.

The UBI Center Problem

Max Ghenis had founded the UBI Center in 2019 to conduct rigorous, open-source research on universal basic income policies .^[2] The idea was to bring quantitative analysis to UBI debates—not advocacy, but numbers. How much would different UBI designs cost? Who would gain, who would lose? How would a basic income interact with existing taxes and benefits?

The problem was immediate: UBI is a deceptively simple policy that interacts with everything else. A $1,000-per-month basic income doesn't just add $1,000 to everyone's income. It affects tax liabilities. It changes benefit eligibility. It alters work incentives. To model UBI properly, you needed to model the entire tax-benefit system.

In the United States, Tax-Calculator could handle federal income taxes, but not benefits like SNAP or Medicaid. OpenFisca-US was nascent. And for state-level analysis—crucial since many UBI proposals targeted specific states—the tools barely existed at all.

Then Ghenis met Nikhil Woodruff, a young developer based in the UK with a talent for building things quickly .^[1] They wanted to analyze UBI policies across both countries: the US where Ghenis lived, and the UK where Woodruff lived.

But there was no open-source model of the UK tax and benefit system.

So they built one.

Building OpenFisca UK

The foundation was OpenFisca, the French framework that had proven legislation could be encoded as executable code .^[3] But OpenFisca was a framework, not a complete model. Someone still had to encode every benefit rule, every tax bracket, every eligibility threshold.

Woodruff did the work. Over months in 2020 and 2021, he built OpenFisca UK from scratch: Income Tax, National Insurance, Universal Credit, Child Benefit, Council Tax Reduction, dozens of interacting programs. The codebase grew to encode the complex reality of British tax and benefit policy—means tests, tapers, interactions, special cases .^[1]

"Over the past year, we developed OpenFisca UK, the UK's first open source tax and benefit microsimulation model, and used it to produce four reports and three conference presentations on UBI in the UK."

The model could run simulations, but only for programmers comfortable with Python. That wasn't good enough.

From Model to Product

The leap from research tool to public product was deliberate. The UBI Center had built interactive tools before—visualizations that let users explore different UBI parameters. But PolicyEngine would go further. It would put the full power of the microsimulation model in a web browser, accessible to anyone .^[4]

The technical architecture reflected this goal. The model ran on Python in the cloud, called through an API. The frontend was a React application that managed complex state: household definitions, policy parameters, simulation results. Users could define reforms without writing code, and see results update in seconds.

In October 2021, the creators spun PolicyEngine off as a new nonprofit organization. The mission statement evolved: no longer just "Make Everyone a Policymaker," but "Help People Understand and Change Public Policy" .^[1] The second verb mattered. Understanding was passive; changing required agency.

Within months of launch, the public policy community had started using the tool. The UBI Lab Network embedded PolicyEngine in their Resilience UBI proposal. Parliamentary groups experimented with policies in real time during presentations. When the Chancellor announced Autumn Budget changes to Universal Credit, PolicyEngine had analysis published within a day .^[1]

Crossing the Atlantic

In March 2022, PolicyEngine expanded to the United States .^[5]

The US model was different—not just in the policies encoded, but in the data challenges and institutional landscape. The UK has a single national tax system, though Scotland sets its own income tax rates (Wales and Northern Ireland have more limited fiscal powers). The US, by contrast, has fifty states with entirely separate income tax codes, plus a federal system of bewildering complexity.

The team started with household impacts: enter your circumstances, see your taxes and benefits. Then in July 2022, they added population impacts using the Current Population Survey—the same microdata foundation that official government estimates relied on .^[5]

State by state, the model grew. Maryland. Massachusetts. Oregon. New York. Pennsylvania. Washington. Each state had its own income tax structure, its own EITC variants, its own quirks.

"The US is not just one launch: each state has their own benefit program rules, and most have their own income tax as well."

The expansion was exhausting and incomplete. By the end of 2022, PolicyEngine had modeled six states with dozens more to go. The gap between ambition and capacity was palpable.

What "Open" Actually Means

PolicyEngine inherited the open-source ethos of its predecessors, but extended it deliberately. The code was on GitHub—anyone could see the formulas, trace the calculations, identify bugs .^[6] The methodology was documented. The data sources were cited.

But open-source means more than accessible code. It means choices about governance, funding, and sustainability.

The organization was structured as a nonprofit. This wasn't accidental. Tax and benefit analysis is politically charged; an organization funded by ideological donors or structured for profit would face questions about neutrality. The nonprofit structure was a commitment: analysis without advocacy.

Funding came from foundations and individual donors. End Poverty Make Trillions provided early grants for US development. Innovation Network for Communities and Gary Community Ventures followed. The Policy Simulation Library Foundation served as fiscal sponsor, offering tax-deductible status .^[5]

The contributors were largely volunteers, supplemented by a small paid team. Dozens of developers contributed to the various repositories: the US model, the UK model, the web app, the API, the documentation. Open source meant distributed maintenance—anyone could fix a bug, add a benefit program, update a threshold.

The Use Cases Emerge

As PolicyEngine grew, patterns emerged in how people used it.

Researchers found a tool that could answer questions they couldn't answer before. The Center for Growth and Opportunity used PolicyEngine to analyze how targeted cash assistance affects work incentives. The Social Market Foundation modeled cost-of-living responses. Academic economists cited PolicyEngine results in papers .^[5]

Advocates discovered they could make quantitative arguments without paying consultants. The Maryland Child Alliance used PolicyEngine to analyze child poverty policies. UBI Lab Northern Ireland modeled recovery basic income proposals. Organizations that had previously relied on government estimates or expensive think tank reports could now run their own simulations.

Journalists appreciated the speed. When UK Prime Minister Liz Truss announced tax cuts in her short-lived administration, PolicyEngine produced distributional analysis within hours—estimates that appeared in news coverage before official government figures .^[5]

Policymakers themselves used the tool. The team presented to the US Congressional Budget Office, to UK Parliamentary groups on Universal Basic Income, to the Green Party of England and Wales at their PolicyFest .^[5] Staff members and elected officials ran scenarios, explored alternatives, asked "what if."

And increasingly, developers built on the API. The Fund for Guaranteed Income integrated PolicyEngine to show participants how pilot programs would affect their benefits. Gary Community Ventures used the API in their tools. The boundary between PolicyEngine as product and PolicyEngine as infrastructure was blurring.

Recognition and Validation

In April 2023, the Digital Public Goods Alliance—a UN-endorsed initiative involving UNICEF, the Norwegian Agency for Development Cooperation, and others—added PolicyEngine to their registry .^[7] The recognition affirmed that PolicyEngine met their standard for digital public goods: open-source, privacy-respecting, do-no-harm, and supporting the UN Sustainable Development Goals.

The specific SDGs cited were telling: ending poverty, achieving gender equality, promoting inclusive economic growth, reducing inequality. PolicyEngine wasn't just a technical achievement; it was infrastructure for the kind of analysis those goals required.

By late 2023, the team had grown from two cofounders to a staff of five, with a researcher network and dozens of open-source contributors. The models had expanded to include Canada and Nigeria prototypes. The data science had improved with machine learning-based reweighting that made estimates more accurate than ever before .^[8]

And the vision was evolving. "We've come to see another opportunity," the 2021 review had noted. "The same open source policy simulation models that power our reform analysis can also show people their tax liability and benefit entitlement under current law" .^[1]

It wasn't just about policy reform analysis. It was about helping ordinary people understand the complex systems that governed their financial lives.

The Gaps Remain

Despite the progress, honest assessment revealed limitations.

Coverage was incomplete. The US model handled federal taxes and major benefit programs, but not every state, not every local quirk, not every edge case. The UK model was more comprehensive but still had gaps—housing benefit calculations, legacy benefit interactions, Scotland-specific provisions.

Accuracy varied. The team invested heavily in validation, comparing PolicyEngine results to official calculators and published statistics. But microsimulation is an approximation, and the further you got from typical cases, the more uncertainty crept in .^[9]

Usability challenged non-experts. The interface was simpler than programming, but modeling a complex reform still required understanding of policy terminology, tax concepts, and the limitations of the tool. The dream of "everyone a policymaker" remained aspirational.

Resources were thin. A small team trying to maintain and improve models for multiple countries, respond to user requests, produce original research, and build new features faced constant tradeoffs. Open source meant anyone could contribute, but maintaining a production-quality policy analysis tool required sustained professional effort.

The proof of concept had worked. PolicyEngine demonstrated that open-source, accessible, rigorous policy analysis was possible. But the concept was still far from complete.

Toward a Platform

By the end of 2022, the founders were already thinking about what came next.

"We've been exploring opportunities to leverage the recent explosion of artificial intelligence tools to make policy analysis more robust, accessible, and even delightful," they wrote .^[5]

The microsimulation engine was the foundation. But the engine could power many things: chatbots that explained policies in plain language, personalized calculators for specific use cases, automated research assistants that could answer policy questions.

The question was whether PolicyEngine would remain a product—a specific tool with a specific interface—or become infrastructure: a platform on which others could build.

That question, and the larger question of what AI would mean for policy analysis, is the subject of later chapters. But first, we need to understand what PolicyEngine actually does—how it works at the level of individual households and entire populations.

PolicyEngine: Proof of Concept