Behavioral Science Consulting Methodology: How We Turn Research Into Business Results

Last updated October 19th, 2026

In a 2021 field experiment published in Nature, Katherine Milkman and colleagues had 30 scientists design 54 four-week programs to get 61,293 members of a US gym chain to exercise more, and just under half of the programs raised weekly gym visits over a control condition, by between 9% and 27%. Only about 1 in 12 still had a measurable effect after the program ended, and impartial judges asked to forecast the results could not tell in advance which programs would work best.

At The Decision Lab, our behavioral science consulting methodology is designed to reduce the risk of investing in interventions that sound promising but do not change business outcomes in your setting. We treat published findings as hypotheses to test with your own customers or employees, in seven steps: define the target behavior and the business metric, grade the research, diagnose the behavior in your setting, design several candidate interventions, test them against a comparison group, measure business outcomes over time, and build and hand over what worked. We have used it across the more than 60 client and research projects in our published case study library, from a behavioral diagnostic for the World Bank in Uganda to a sales-call experiment for a Canadian insurer that raised sales by 12% against a control group.

Each step answers one question before the next one starts, and each produces something the client can check.

Academic findings describe a specific behavior in a specific setting, so translation starts by describing your own behavior with the same precision. A goal such as improving engagement has to become something countable, for example the share of new users who return for a second session within seven days. We then tie that behavior to the business number it should move and agree on a guardrail metric, a number that must not get worse, such as cancellations or complaints.

When several behaviors compete for attention, we rank them before any research starts. For one of North America's largest insurers, we ranked candidate pilot projects by expected return and feasibility, and weighed how far each result would carry to other teams. Agent sales calls came out on top and became the first project.

• What you receive: a written definition of the target behavior, with its primary metric and its guardrail metric.
• How we decide whether to move forward: we continue once both sides agree the behavior can be counted with data you already hold or can collect.

Published effects vary widely. Dennis Hummel and Alexander Maedche reviewed 317 nudge effects from 100 publications and found a median change of 21% relative to control groups, with some nudges proving ineffective or even backfiring.

We rate each relevant finding on who the participants were, the setting, the behavior measured and the strength of the study design, since a lab study of students choosing snacks says little about employees choosing software. For a Fortune 500 HR technology company, our review covered 48 sources on technology adoption. For Money Guided, an employee financial-wellbeing product, it covered more than 50 studies.

• What you receive: an evidence review in which each finding is rated for how well it transfers to your setting.
• How we decide whether to move forward: we continue when the review yields hypotheses worth testing, and when the evidence is thin we say so and widen the diagnostic in step 3.

A behavioral diagnostic finds out why the target behavior is or is not happening with your people, using interviews, surveys, observation and the data you already hold. We often organize the findings with the COM-B model, which Susan Michie and colleagues built from a review of 19 earlier behavior-change frameworks. It separates barriers that sit inside the person from barriers that sit in their surroundings, so each one gets a matching fix.

For the World Bank in Uganda, where fewer than 1 in 20 households used improved cookstoves, our team worked with government officials and stove distributors and collected data from hundreds of households. We grouped what we found into seven themes and 26 specific barriers. One of them was physical: the cheapest stove was too heavy for most women to carry home on their own, a barrier that better information alone could not remove. The World Bank published the work as its Uganda clean cooking behavioral diagnostic in 2019.

At the Fortune 500 HR technology company, the diagnostic found five barriers to AI use and two employee groups who failed in opposite ways. Enthusiasts were already experimenting and kept running into friction, while skeptics doubted AI's value and stopped at first contact.

• What you receive: a map of the specific barriers behind the behavior, with the groups each one affects.
• How we decide whether to move forward: we continue once every barrier on the map is supported by data from your own customers or employees.

The gym megastudy showed that impartial judges could not pick the winners in advance, so we design more candidates than we plan to test. Each candidate is aimed at a named barrier for a named group. We then score the candidates against criteria agreed with the client before anyone has a favorite: the strength of the evidence behind the idea, its fit to the diagnosed barrier, the cost to build it, and whether it serves the person's own interests as well as the business's.

For the Fortune 500 HR technology company, we designed 20 candidate programs and selected eight for piloting. They ranged from hands-on tasks built into employees' real workflows to a mission-based challenge series run in Slack.

• What you receive: a shortlist of candidate interventions, each scored against the agreed criteria.
• How we decide whether to move forward: a candidate goes to testing only if it clears every criterion, including the check that it serves the person's own interests.

A comparison group tells you whether the change caused the improvement, or whether the number would have moved anyway. Controlled experiments give the clearest answer, but the right test depends on the decision at stake. A small, easily reversed change may only need careful before-and-after monitoring, while a costly rollout deserves a randomized test. Before any test starts, we write down the result that would justify shipping the change.

• Before launch, an online experiment can compare versions with a sample of target users. For Money Guided, a 400-participant controlled experiment tested four ways of framing the product's value against a control, using validated wellbeing scales.
• In the field, a pilot runs with real employees or customers next to a control group. At the Fortune 500 HR technology company, we piloted all eight programs for one month with more than 100 employees. Confidence in exploring AI tools, measured before and after, rose 41% among pilot participants and fell 26% among control-group colleagues over the same period.
• In live operations, a change is rolled out to part of the business while the rest carries on as before. For the insurer, we piloted new call scripts and agent training against a control group, and sales rose 12% from the first pilot, with a projected $35 million increase in annual revenue.
• Inside a digital product, an A/B test shows different versions to randomly split groups of users and compares what each group does next.

The four formats differ in cost and speed, and every one of them compares results with people who did not receive the change.

• What you receive: results against a control group, read against the decision rule agreed before the test.
• How we decide whether to move forward: the change ships only if it meets that decision rule, and otherwise we revise it or drop it.

Effects found in research usually shrink in practice. Stefano DellaVigna and Elizabeth Linos found that nudges run at scale by two US government nudge units raised take-up, the share of people who did the target behavior, by about a sixth as much as nudges published in academic journals.

We plan for this by measuring the business metric and the guardrail metric past the end of the intervention, and by stating what has not yet been measured. Our Money Guided results, for example, validated the product design before launch, and in-product behavior remains the next test.

• What you receive: business-outcome data over a realistic time window, including the guardrail metric.
• How we decide whether to move forward: scaling starts when the effect lasts past the end of the intervention and the guardrail metric has not worsened.

A finding changes a business only when someone builds it into the product or process that shapes the behavior. We stay involved through that build, and we hand over the materials the client's teams need to run the intervention without us.

At the Fortune 500 HR technology company, each program came with written protocols and guides for the client's own teams, and the programs are now scaling to the full 20,000-person workforce. For the insurer, we designed a permanent in-house behavioral science team, from the organizational blueprint to the hiring profiles, and proved its value on the sales-call project before handing it over.

• What you receive: the assets and protocols your team needs to run the intervention at scale.
• How we decide whether to move forward: the handover is complete when your team runs the intervention without us and results stay close to the pilot, allowing for the shrinkage described in step 6.

AI speeds up two of the slower steps. Language models can produce a first pass of the evidence review in step 2, and machine learning can find patterns in behavioral data too large to read by hand in step 3. For the insurer, we analyzed more than 170,000 hours of sales calls, using process mapping to locate the critical moments in each conversation and machine learning to cluster where and how calls went wrong. We built the script changes and agent training from those clusters.

The judgment calls in the method stay with people who know the setting. A model can summarize what a study found, but rating how well it transfers to your employees depends on the diagnosis, and an AI-generated idea still has to beat a control group in step 5. We use the same seven steps when the behavior under study is AI use itself, as in our enterprise AI adoption work.

What is the difference between academic behavioral research and applied behavioral science?

Academic behavioral research asks whether an effect exists, usually under conditions chosen to isolate it. Applied behavioral science asks whether a change will move a specific behavior and business metric in one organization, with its own people and constraints. Applied work treats academic findings as starting hypotheses and tests them where the real decision gets made.

Do you need a randomized controlled trial to apply behavioral science?

A randomized controlled trial is not always necessary. It gives the clearest evidence that a change caused a result, and it is worth running when a rollout is costly or hard to reverse. For small, reversible changes, careful before-and-after monitoring with a guardrail metric is often enough. The test should match the size of the decision it informs.

Why do behavioral interventions that work in studies often fail in business settings?

Effects measured in studies often shrink when an organization runs them at scale. Stefano DellaVigna and Elizabeth Linos found that nudges run by two US government nudge units averaged about a sixth of the effect reported in academic journals. Selective publication of strong results, made worse by small samples, explained about 70% of that gap.

What is a behavioral diagnostic?

A behavioral diagnostic is a structured study of why a specific behavior does or does not happen in a given population. It combines a review of existing research with interviews, surveys, observation and behavioral data, then lists the barriers and the groups each one affects. Interventions come after the diagnostic, so each targets a barrier the research actually found.

How do you measure the return on a behavioral science intervention?

Measure the change in the business metric the target behavior was chosen to move, against a comparison group, over a period long enough to show whether the effect lasts. Then set that change against the cost of building and running the intervention. For one insurer, a 12% rise in sales over a control group translated into a projected $35 million increase in annual revenue.