How to Redesign a Public Service So More People Use It: A Behavioral Playbook for Program Uptake

Last updated October 30, 2026

When US states rechecked Medicaid eligibility after the pandemic, at least 25 million people lost their health coverage by September 2024, according to KFF's Medicaid Enrollment and Unwinding Tracker. About 7 in 10 of them lost it for procedural reasons, such as an unreturned renewal form or outdated contact details, and only about 3 in 10 had been found ineligible.

To redesign a public service so more people use it, define the exact behavior you want to increase, find where eligible people drop out, remove barriers in the service itself, test messages and designs against a comparison group, and measure uptake across priority groups before scaling. At The Decision Lab, an applied behavioral science firm, we do this work for governments and public health organizations, including Health Canada, the City of Rome, the World Bank in Uganda, Kids Help Phone and Wellness Together Canada, Canada's government-funded mental health platform.

Low uptake rarely comes down to awareness alone. Eligible people drop out because a service is hard to reach, poorly timed, socially risky or designed around the agency's own processes. Below are the five steps we follow, with examples from our public-sector work at each stage.

1. Define the behavior. Pick the single action that counts as success, such as a dose given or a session attended, and record its current level.
2. Diagnose the barriers. Map every step eligible people have to take and find where they stall, and why.
3. Redesign the service. Fix the points where people drop out inside the service before spending on promotion.
4. Test behavior change campaigns. Compare messages and designs against a comparison group before choosing one to scale.
5. Measure equitable uptake. Report results for each priority group, and check whether gains hold after the campaign ends.

Program teams often start with goals like "awareness" or "engagement," and those goals make it hard to tell later whether anything changed. We ask for the single action that counts as success: a vaccine dose given, a form submitted, a first counselling session attended, a vote cast. When the City of Rome asked us to help with its #RomaDecide participatory budget, the outcome was concrete from the start. The city had set aside 20 million euros for local projects and wanted as many residents as possible to vote on how it was spent.

A precise behavior also tells you what data you will need. If success is a completed booking, someone has to be able to see bookings by week and by neighbourhood before the project starts, or there will be nothing to compare against at the end.

Most low uptake comes from a handful of specific points where people stall, and those points are rarely where the program team expects them. We run a behavioral diagnostic, combining interviews and surveys with the people the service is meant for, and map what we learn against each step they have to take.

In Uganda, the World Bank asked us why clean cookstoves had reached under 5% of the market as of 2018, even though models costing less than $10 could cut a household's fuel spending in half. Our behavioral diagnostic of clean cookstove adoption produced seven core themes, such as trust and access to the stoves. We mapped those themes onto the COM-B model, a standard behavioral science framework for sorting the barriers that stop people from acting, and ended with 26 specific barriers. Some were about knowledge, like not knowing the health risks of indoor smoke. Others were physical: the cheapest stove was too heavy for most women to carry home on their own, a problem no awareness campaign could solve. We then tested responses through randomized trials delivered by text message.

The same diagnostic logic applies to services people avoid for social reasons. When the Indiana National Guard asked us why at-risk members were not seeking mental health treatment, the barriers we focused on were a culture of silence and the fear that asking for help would damage a career.

If the diagnostic shows that people drop out inside the service itself, a better campaign will mostly send more people into the same dead end, so we fix the service before promoting it.

For Kids Help Phone, we led the redesign of how young people reach mental health support online, from inception to launch. That meant research with hundreds of young people across Canada, interviews with staff from crisis responders to service innovation specialists, a review of web analytics and an audit of the existing site. The design problem was specific: millions of young people visit each year with very different needs, and the site has to stay easy to use for someone in distress while still offering several levels of support. We did similar work redesigning the Wellness Together Canada platform, where users arrive from different cultural and linguistic backgrounds and with different levels of readiness to engage.

Uptake problems also sit with the professionals who deliver a service. CASEL, which supports educators and policy leaders in putting social and emotional learning into practice, found that uptake of its recommendations was low. We identified the behavioral barriers holding implementation back and redesigned CASEL's digital tools to reach those audiences more directly.

Once the service works, the campaign that promotes it should be tested before it is scaled, because the message that looks strongest in a meeting is often not the one that moves people.

Two campaigns run side by side in England show how far apart reasonable ideas can land. In a national randomized trial published in The Lancet, Michael Hallsworth and colleagues sent a letter from England's Chief Medical Officer to 1,581 general practices whose antibiotic prescribing was in the top fifth for their area, telling them most nearby practices prescribed fewer antibiotics per person. Over six months, those practices dispensed 3.3% fewer antibiotic items per 1,000 people than practices that got no letter, about 73,000 fewer antibiotic items in total, where fewer is the goal. A patient-facing education campaign run alongside it had no measurable effect.

Testing many versions at once helps because expert intuition is a weak guide. In a pharmacy vaccination megastudy led by Katherine Milkman, 689,693 Walmart pharmacy patients received one of 22 text reminders or no message at all. On average, the texts raised flu vaccination over the following three months by about 2 percentage points compared with patients who got no message. The best message, two texts three days apart saying a vaccine was waiting for the patient, was one that neither experts nor lay people predicted would win.

A controlled experiment is the strongest test, but it is not always possible or necessary. In Rome, we tested a wide range of message framings with nearly 10,000 residents before the campaign ran. Some appealed to family and community, and others used the endowment effect, the tendency to value something more once it feels like yours, to build a sense of ownership over the city budget. For Health Canada's organ donation and transplantation governance, the people whose choices mattered were the decision makers themselves. We used a discrete choice experiment, a survey method that asks people to choose between options with different features, to help the steering committee settle on a governance model, then ran a facilitated workshop so members could refine the result together.

Measure the behavior you defined in step 1 against the comparison you set up in step 4, then break the results down by the groups the program most needs to reach. An average rise can hide a group that did not move at all. In Rome, we compared how different communities responded to each framing, so the city could reach groups whose voices are often underrepresented. In Nepal, where we worked with the non-profit Aerosan and the Government of Nepal to raise public toilet use, public toilets in some areas were used by 7 times as many men as women, so a rise in total visits would have said little about whether the program was working for women.

Many public programs also face a mandated evaluation. We find it pays to design that evaluation at the start, alongside the intervention, so the comparison group and the data collection are in place before the first message goes out. Report the results that did not work too. The patient campaign in the English antibiotics trial is useful to every health agency precisely because it was reported.

AI is changing government health communication in two ways. Citizens now ask AI tools for health guidance before, or instead of, contacting a public service, and agencies are beginning to use AI to draft and personalize their outreach. We treat AI as a communication channel to test like any other, with accountability and human support staying with the agency.

On the citizen side, about 1 in 4 US adults used an AI tool for health information or advice in the 30 days before they were surveyed, according to a West Health-Gallup survey of 5,660 adults in late 2025. About 1 in 7 of those recent users said the AI's answer led them to skip a visit to a health provider. Their trust is low and divided: about a third trusted the information and about a third distrusted it, and only 4 in 100 strongly trusted its accuracy. Public agencies cannot control what those tools say, but they can keep their own guidance current and easy for AI systems to find and cite.

On the agency side, the evidence points in two directions at once. In a study of AI-written vaccination messages, Elise Karinshak and colleagues found that participants rated pro-vaccination messages written by GPT-3 as more effective than messages from the US Centers for Disease Control and Prevention, yet liked messages less once they were labelled as AI-written. A study published in Science by Thomas Costello, Gordon Pennycook and David Rand had more than 2,000 people who believed a conspiracy theory talk it through with GPT-4 Turbo. Their belief in that theory fell by about a fifth on average after the conversation, and the drop was still there two months later.

We set four conditions before an agency puts AI in front of citizens:

1. Test AI-written or AI-delivered messages against a human-written version with the intended audience.
2. Keep a named public body visibly responsible for what is said.
3. Give people a way to reach a person when the AI cannot help them.
4. Check results separately for each priority group before large-scale deployment.

The fourth condition follows the World Health Organization's 2024 guidance on large AI models in health. Among more than 40 recommendations, it asks governments to require independent audits after large-scale deployment, with results broken down by user characteristics such as age and disability. In our public-sector AI work, including predictive modelling with Aerosan and the Government of Nepal, we treat AI-enabled interventions as pilots and measure their effect on behavior before recommending wider deployment.

The Decision Lab provides applied behavioral science consulting for public health programs where awareness has not turned into participation. We help public agencies and health organizations find the barriers to uptake, then redesign and test services and communications before they are rolled out across a population. Our core services for public health programs are:

• Behavioral diagnostics and journey mapping
• Interviews and surveys with the people a service is meant for
• Service and digital-experience redesign
• Behavioral campaign and message design
• Randomized controlled trials and other comparison-group evaluations
• Analysis of results by community or demographic group
• Discrete choice experiments to support policy and governance decisions

Government behavior change campaigns are designed and tested by in-house behavioral insights teams, specialist behavioral science consultancies, academic research groups and communications agencies. The Decision Lab is a behavioral science consultancy that helps governments find the barriers to uptake and test redesigned services and interventions before they are scaled.

In-house behavioral insights units, found in many national and city governments, know the policy context and can run trials on their own data. Specialist consultancies suit agencies without an in-house unit, or teams that need extra capacity for a specific program. Academic research groups run rigorous large trials, often over longer timelines. Communications agencies produce campaigns at scale but rarely test them against a comparison group.

We are probably not the right fit when the service already works well and what you need is media buying or production volume. We are a good fit when people who could benefit from a public service are not using it and nobody can yet say why. For a related problem, see our guide on how to increase public benefits enrollment, and for building testing into an agency's routine, our guide to building a behavioral experimentation program.

Each summary below links to the full published case study.

Health Canada: organ donation and transplantation governance

The problem: a pan-Canadian steering committee with many stakeholders needed to agree on how organ donation and transplantation would be governed. The method: interviews, surveys, a discrete choice experiment and a facilitated prioritization workshop with committee members. The outcome: a governance model and implementation recommendations the committee shaped together. Read the Health Canada organ donation case study.

City of Rome: participation in the city budget

The problem: fewer residents were voting in the #RomaDecide participatory budget than the city hoped. The method: we audited the city's existing messages against the research literature, then tested framings with nearly 10,000 residents across different communities. The outcome: a behaviorally informed messaging strategy for the campaign. Read the Rome participatory budgeting case study.

World Bank in Uganda: clean cookstove adoption

The problem: clean cookstoves had reached under 5% of the market in Uganda as of 2018, despite clear health and cost benefits. The method: field research mapped onto the COM-B model, followed by randomized trials delivered by text message. The outcome: 26 specific barriers identified and responses tested in the field. Read the Uganda clean cookstove case study.

Kids Help Phone: youth mental health support

The problem: millions of young people with very different needs had to find the right level of support on one website. The method: research with hundreds of young people across Canada, staff interviews, web analytics and a site audit. The outcome: a redesigned gateway to support that we led from inception to launch. Read the Kids Help Phone case study.

Wellness Together Canada: a national mental health platform

The problem: people arriving at the government-funded platform differed widely in background and in readiness to engage, and distress can make choosing between options harder. The method: user research and behavioral design with the consortium of mental health organizations behind the platform. The outcome: a redesigned platform built around how people in distress make decisions. Read the Wellness Together Canada case study.

Aerosan and the Government of Nepal: public sanitation

The problem: in some areas, public toilets were used by 7 times as many men as women, partly because of stigma around menstruation. The method: behavioral research combined with predictive modelling. The outcome: behaviorally informed recommendations to promote better sanitation practices. Read the Aerosan public toilet case study.

What is applied behavioral science consulting for public health programs?

It is consulting that uses research on how people actually make decisions to raise participation in health programs such as vaccination or mental health services. A typical engagement first diagnoses why eligible people do not take part, then tests a redesigned service or message against a comparison group before it is scaled.

Can behavioral science improve uptake of government health programs?

Yes. Behavioral science helps agencies find why eligible people do not use a program, whether the cause is friction, low trust, stigma, confusing eligibility rules or a poorly designed service. Agencies can then test targeted changes to the service and the way it is promoted with a comparison group before rolling them out widely.

How long does a public service redesign project take?

A focused diagnostic and redesign can take a few months, and a field test with a comparison group usually adds at least one full program cycle, such as a flu season or an enrollment period. The timeline depends most on how quickly the agency can share data and run a test inside its existing systems.

Do we need a randomized controlled trial to know if a campaign worked?

Not always. A randomized trial gives the most reliable answer, and is worth the cost when a decision will be scaled widely. When a trial is not possible, a well-matched comparison group or a rollout staged across regions can still show whether uptake moved. Testing framings with the audience before launch also helps.

What should governments test before using AI to communicate health behaviors?

Governments should test AI-written or AI-delivered messages against human-written versions with the intended audience, and check the results for each priority group before large-scale deployment. A named public body should stay responsible for every message, and people should always be able to reach a person when the AI cannot help.

Can AI-written health messages replace human-written public communication?

Early evidence suggests AI can draft messages that people rate highly, but audiences say they prefer public health messages that come from human institutions. We recommend treating AI drafts as candidates to test against human-written versions, with a named public body responsible for every message that goes out.

How do you measure program uptake?

Count the one behavior the program exists to change, such as doses given or sessions attended, and compare it against a baseline or comparison group over the same period. Count at the point of action, such as attendance, rather than registrations, and check whether the rise holds once the campaign stops.