Why We Built CentreAI Policy

As our Executive Chairman Tony Blair has written: “The centre is the place where policy comes first and politics second. You work out the correct analysis, then the correct answer, and shape your political strategy around it.” CentreAI Policy is our tool to help leaders and policymakers conduct the right analysis, find the answers and write better policy.

AI is already widely used in policy research. But policymaking is specialist work, and it needs specialist tools. Analysis has to be rigorous, traceable, built on trusted data and designed around how policy is actually shaped. AI used naively, with poor data, will not produce better policy. There is a real risk it produces worse.

But good policy also depends on the quality of human decision-making. As analysis becomes more abundant, the scarce thing is no longer information but the ability to cut through it, giving policymakers the optimal conditions in which to exercise their judgement.

CentreAI Policy is built for both problems. It is designed with and for policymakers, giving them evidence they can trust, and freeing up their time for the discussions and judgements only they can make. It keeps the policymaker in command throughout, and it is built to counter the human and organisational pitfalls that can often turn good analysis into bad policy.

What Gets in the Way of Good Policymaking

From our years of experience embedded at the centre of government in more than 45 countries, we’ve observed three recurring problems with policymaking. We designed CentreAI Policy with each of them in mind.

1. Unclear objectives

Perhaps the most common way that policymaking fails is when the outcomes are unclear or don’t exist. Or even worse: when a policy tries to be the solution to everything.

This is the core of the “Everything Bagels” argument made by political commentator and journalist Ezra Klein. A new railway ends up also becoming a regional jobs plan and a habitat-regeneration project. Each additional objective sounds reasonable on its own. But every one that’s added dilutes the others, and by the end, nobody can say what the policy was really for.

An example used by Klein is the Tahanan supportive-housing project in San Francisco. It provides permanent housing for the chronically homeless and was built in three years for under $400,000 a unit, roughly half the time and cost of a typical affordable-housing project in the Bay Area. However, it achieved this in large part by using private philanthropic funding, letting it skip the onerous public requirements and reviews otherwise required. If faced with the option of doubling the value for money at the expense of losing those additional and often hard-to-measure benefits, how many citizens would choose the latter?

This is not to say that policymakers should not consider the knock-on effects of their policies or view policy in isolation, but policies should not try to be all things to all people. They should have a clear target outcome. The more outcomes a policy seeks to address, the weaker the policy typically becomes and the more difficult it is to measure its success or failure.

2. Poor use of evidence

Even when the objective is clear, policymaking routinely fails to make good use of evidence. This happens in three ways.

The first is where evidence is not used at all. Leaders sometimes pursue policies for political reasons, resulting in policies written with little attention paid to alternatives or what has gone before. In more technocratic governments, the same tendency exists but is better disguised as “policy-based evidence” – the leader asserts what they want, and the policy researcher’s job is justifying it with cherry-picked evidence.

The second is capacity. Leaders who wish to pursue evidence-based policy often lack the staff to pursue it effectively. This is particularly true for smaller or regional governments that may not have access to the same depth of policy expertise. Carrying out in-depth assessments of possible policies, analysing historical and international precedents, and conducting academic research is time consuming and expensive.

The third is the illusion that, with enough evidence, a perfect policy can be found. Unfortunately, there is no such thing; the world is messy, the future unknown, and no volume of analysis changes that. An excess of evidence may overwhelm, rather than help, a busy decision-maker. Endlessly commissioning more reviews and assessments can also result in missing the window where effective action is possible.

3. Lack of constructive challenge

The best policies emerge from robust debate. In a well-functioning policy team, there should be well-informed debate and discussion. Too often, however, policy is written by cosy consensus, and the first real challenge it meets is a painful collision with reality. In the United Kingdom, the Covid-19 Inquiry found that “advisers and advisory groups did not have sufficient freedom and autonomy to express dissenting views and suffered from a lack of significant external oversight and challenge. The advice was often undermined by ‘groupthink’.”

When leaders neglect to challenge their officials, policies can drift towards the path of least resistance or, worse, towards an institution’s own interests. When officials fail to push back on leaders, policies can become impractical or contradict policy elsewhere, producing confusion and poor outcomes.

One reason there is not more challenge is information asymmetry. Ministers and senior officials often do not know enough to interrogate what is put in front of them, so they accept recommendations without stress-testing them or asking whether they really serve the core objective.

The problem is compounded by a tendency to present a single recommendation rather than letting the decision-maker hear a range of options and arguments on each side. Where it does happen, it often takes the form of three options, one of which is framed as the only sensible choice a leader could make.

A good options paper should provide the basis for a real debate between different but viable choices. This is the insight behind research psychologist Irving Janis’s work on “groupthink” and political scientist Alexander George’s related “multiple advocacy” model: both show the danger of reaching consensus too early, without a rigorous assessment of competing perspectives. The classic illustration of groupthink is the disastrous Bay of Pigs invasion in April 1961, where American President John F Kennedy’s advisors converged on a deeply flawed plan without any serious internal challenge. By the time of the Cuban missile crisis 18 months later, JFK had changed his approach, convening his top advisors to surface multiple options, each with a real advocate. The result was a much better, if heated, decision process that achieved America’s main aim while sparing the world a further escalation.

What We Built, and Why

In short, with the help of the leaders and expert policymakers in our network, we set out to build an evidence-driven tool that addresses each of the failures described above.

CentreAI Policy takes the user through a series of deliberate stages. The policymaker must first define the problem and the key outcomes they want to achieve. The tool then produces a “State of Play” to outline the current situation, what has been tried and an assessment of the current political context. Only after aligning with the user will it proceed to suggesting interventions.

The research and evidence is then assembled by AI agents, each designed for a specific task. The evidence comes from a curated catalogue of hundreds of thousands of official policy documents, strategies, think-tank reports and academic studies, supplemented by web search constrained to authoritative domains. Every claim is cited and traceable back to its source. CentreAI Policy specifically looks across other countries and jurisdictions for case studies, with implementation timelines, evidence of impact and an assessment of how transferable each case is. This is the in-depth, evidence-based work that many leaders lack the resources to do. CentreAI Policy pauses after each stage, allowing the user to query or redirect so they remain in control.

To help leaders arrive at the best policy, CentreAI Policy promotes challenge. It follows the multiple-advocacy model: each option has a distinct premise and posture that an advocate might plausibly argue for. Each option is also put through a stakeholder simulation. AI agents identify groups the policy would affect and predict how each will likely respond based on what they have previously said. The purpose of this is to force the decision-maker to consider a wide range of perspectives, to catch oversights and to work out how the policy could be better drafted to negate avoidable opposition.

None of this replaces the judgement of the policymaker. CentreAI Policy is designed to present trade-offs and uncertainties honestly, and let human decision-makers decide. Nor does CentreAI Policy remove the need to talk to actual stakeholders and experts. If anything, it can help identify who they should talk to and free up a leader’s time to actually meet with those people. It is our goal to empower political leaders and policymakers to make better and more informed decisions and to write better policy that delivers for citizens.