Behavioral Science for Financial Services: Improving Customer Decisions Without Manipulation

Last updated October 9, 2026

About 2 in 5 UK adults have used a generative AI tool such as ChatGPT or Gemini for personal finance advice, according to a nationally representative Censuswide survey for Finder's June 2025 research on AI and personal finance, while about a third said they oppose using AI this way. In an online experiment the Ontario Securities Commission ran with 7,771 Canadian adults, participants showed no material difference in how closely they followed investment suggestions attributed to an AI tool or to a human provider, and they acted on suggestions that were deliberately unsound.

Behavioral science for financial services is the use of evidence about how people actually make money decisions to design products, disclosures, sales conversations and reminders that help customers act in their own interest, tested against a control group before they scale. At The Decision Lab, we do this work for financial regulators, insurers, banks and debt-relief nonprofits, and we judge every design on the customer's outcome as well as the firm's.

The line between helpful choice architecture and manipulation comes down to whether a design still works once customers understand how it was built, and whether they end up better off by their own standards. We design and test for both. In research conducted with the Ontario Securities Commission, we co-authored a published study that paired a survey of 655 Canadian retail investors with a randomized controlled trial of 1,465 Canadians, measuring how posts from financial influencers change trading and which protections reduce their pull. In a client engagement with a top Canadian insurer, a redesign of agent sales calls lifted sales 12% over a control group, and in a project with American Financial Solutions, a nonprofit debt management program, the dropout rate fell by half after we redesigned its client touchpoints.

Choice architecture is the way a decision is presented to someone: the default option, the order of the choices, the timing of a reminder, the wording on a form. Every financial product already has one, so the practical decision for a bank or an insurer is which design to use and what to measure it on. We put every design through four tests before it goes into a trial.

1. Success is defined by a customer outcome. A design measured only on conversions or trading volume can succeed while customers lose money, so we set a customer outcome, such as debt repaid or a scam avoided, next to the business metric before testing begins.
2. The design survives being explained. If customers who learned how a design works would feel tricked, it fails. Reminders and sensible defaults usually pass, while fake countdown timers and cancellation paths buried several screens deep do not.
3. Leaving is as easy as joining. Cancelling or switching should take no more effort than signing up did.
4. It holds up for the customers with the most to lose. An average improvement can hide harm in a smaller group, so we report effects separately for groups such as first-time investors and older adults. In our finfluencer trial, people with no investing experience were more swayed by promotional posts than experienced investors, and they also gained more from the protections we tested.

These tests leave room for persuasion. A reminder sent shortly before a savings contribution deadline is persuasive by design, and it passes because it serves a goal the customer set for themselves.

At The Decision Lab, we treat financial-services design as ethical when it improves an outcome the customer would choose for themselves, stays defensible when explained, keeps exit easy and has been tested for harm across vulnerable groups. Manipulation is design that depends on customers not understanding it, or that wins for the firm while the customer loses. A fake countdown on a loan offer and a cancellation flow several screens longer than the sign-up both fail on these terms, however well they convert.

Financial well-being

Financial well-being products work better when they are designed around the moment and mindset in which people make money decisions, because income alone does a poor job of predicting how confident or stressed people feel. In research we conducted with Capital One on financial stress, the negative effects of financial stress held even after we controlled for household income and FICO credit score, and our findings pointed to simple prompts that help people step back to their longer-term goals before a financial choice.

In a client engagement with Money Guided, an employee financial well-being product, we surveyed 533 UK employees and 264 HR managers, then ran a 400-person controlled experiment comparing four ways of framing the product's value against a control. Perceived financial position, rather than income, was the strongest predictor of reported financial confidence and stress in that study, and the winning frame now shapes how the product presents itself to employees and to the HR managers who buy it.

Disclosures

A disclosure protects customers only under some conditions, because its source and timing can decide whether it warns them or makes the promoter seem more trustworthy. The literature review in research we conducted with the Ontario Securities Commission on finfluencers found studies showing that disclosing a conflict of interest can make the messenger seem more credible and raise purchase intent. Many of the 655 retail investors we surveyed said they were inclined to trust finfluencers who provide a disclosure.

In the trial, we had the disclosure appear as a label from the social media platform itself, to test whether an independent source avoids that credibility boost. That version reduced the share of participants who bought the promoted asset. For firms writing disclosures, the source and timing of the message deserve as much testing as the wording.

Investing

Investment decisions increasingly happen in apps and social feeds, close to the content that sways them. In our randomized trial with the Ontario Securities Commission, 1,465 Canadian social media users traded in a simulated platform with $10,000 in play money. We measured the share of participants in each group who bought a promoted asset in the round after seeing a post about it.

• Promotional posts changed what people bought. With no protection in place, 38% of participants who saw posts promoting an asset bought it, nearly 2 in 5, compared with 8%, about 1 in 12, of the control group, who saw unrelated posts.
• People with no investing experience were more susceptible. After seeing a post, about 3 in 10 non-investors bought the promoted asset, compared with about 1 in 5 experienced investors.
• Prebunking and inoculation reduced purchases. Prebunking is a short general warning about misleading content placed before the post, and inoculation shows a weak example of a misleading post and explains its flaws. Both lowered purchases, though about twice as many participants still bought the promoted asset as in the control group.
• A risk-confirmation message at the point of purchase showed no measurable effect, and combining protections performed about the same as a single one.

Timing also shaped our redesign of the Ontario Securities Commission's investor education library, GetSmarterAboutMoney.ca. We recommended organizing its resources into courses with a clear path and tying reminders to decision points, such as a message shortly before the contribution deadline for users who said they invest in a tax-sheltered account every year.

Insurance renewal and sales

Renewal can expose insurance customers to inertia, particularly when comparing and switching take more effort than staying put, and regulators have started to draw lines around it. The UK Financial Conduct Authority found that 6 million home and motor insurance customers would have paid £1.2 billion less in 2018, roughly £200 each, had they been charged the average price for their risk. Since January 2022, the regulator's general insurance pricing rules bar insurers from charging renewing customers more than equivalent new customers. At renewal, the designs that pass our four tests make comparison and exit easier, for example by showing last year's premium beside this year's and offering a one-step way to change cover.

Our insurance work has focused on helping customers who want cover reach a decision. In a client engagement with a top Canadian insurer, we analyzed more than 170,000 hours of agent sales calls, using process mapping and machine learning to find where conversations went wrong. We then piloted targeted script changes and agent training against a control group, and sales rose 12% over the control group in the first pilot. For any sales or renewal intervention, we recommend tracking early cancellations and complaints alongside sales, because a sales lift without customer fit should not scale. The same call analysis that shows where a sale stalls can show where a customer is being hurried past a question about what the policy covers.

The pilot carried a projected $35 million increase in annual revenue, and the insurer now runs a permanent in-house behavioral science team that we designed.

Fraud prevention

Scams work by pushing people to act before they think, and the most promising defenses add a moment of perspective. In a controlled experiment from our own research program on financial scams and older adults, we showed 102 North American participants an investment email carrying common red flags of fraud. Adults aged 60 and over who first completed a short personal financial risk assessment found the pitch less trustworthy and were less willing to invest than those who went straight to the email. The same step had no effect on 18- to 25-year-olds, consistent with research showing that older adults lean more on fast, automatic thinking. The sample was small, so the result is a promising design lead and does not yet show that the intervention reduces fraud in live settings.

Our survey with the Ontario Securities Commission points to who is exposed online. Retail investors who had made a financial decision based on a finfluencer's advice were 12 times as likely to report having been scammed on social media as those who had not. A survey cannot show which came first, but it identifies a group to protect early.

Debt repayment

A debt management program only helps people who stay in it. In a project with American Financial Solutions, a nonprofit whose program has helped more than 450,000 people repay over $9 billion in debt across 20 years, the problem was early dropout. We analyzed records from thousands of its clients, with more than 100 characteristics per person, and the barriers our model identified accounted for 4 in 5 program dropouts. The analysis also identified features of the first onboarding call that were associated with later dropout risk, months before clients quit.

We redesigned the counselor call scripts to draw out each client's own reasons for getting out of debt at enrollment, and paired them with an early-warning process that reaches at-risk clients sooner. After the redesigned touchpoints were introduced, the program's dropout rate fell by half. Reminding someone of a goal they named themselves passes all four of our tests.

The four tests define what an ethical intervention must prove, and the five steps below show how a financial institution can build and test one. Our financial services engagements follow them whether the client is a securities regulator or a bank.

1. Define the customer decision and outcome. Start with a decision customers get wrong by their own standards, such as missed savings contributions or abandoned repayment plans, and name the outcome that would show they got it right.
2. Diagnose the barrier. Combine the firm's own records with interviews or surveys to find where the decision goes wrong. In our debt repayment work, the risk showed up in the onboarding call, months before clients left.
3. Set ethical guardrails. Write the four tests into the test plan and agree in advance what would stop a rollout, such as a rise in complaints or worse results for older customers.
4. Test against a control group or a simulation. A field trial gives the strongest evidence, though it is not always possible or appropriate with real money at stake. A well-built online simulation, like the trading platform in our finfluencer study, can answer questions that would be unethical to test on live accounts.
5. Segment results before scaling. Scale what improves the customer outcome for the groups most at risk, as well as on average.

AI is changing both who gives financial advice and who writes the scams. The Ontario Securities Commission's regulator-led experiment found no material difference between how closely Canadians followed AI and human investment suggestions, so an AI assistant that recommends a product carries the same choice-architecture responsibilities as the advisor it stands in for. How an AI recommendation is labelled, and whether the label changes how far customers trust it, is now something firms and regulators can test directly.

Machine learning is also how much of our financial services work finds where decisions go wrong, from clustering 170,000 hours of insurance sales calls to the dropout model we built for American Financial Solutions. Our four tests apply to models as much as to people. A model trained only to raise conversion will probably learn to find the customers least likely to read the terms, so we make the customer outcome part of what the model is judged on.

AI raises the stakes of choice architecture, because a single design decision inside an AI assistant repeats across every customer it serves. When an AI assistant recommends a financial product, sets a default, ranks options or decides when to step in, firms should test whether customers understand the recommendation and whether it improves their own outcomes, including for the customers most at risk. The trial designs we used for finfluencers carry over directly to AI-generated advice and AI-assisted scams.

What is the difference between a nudge and a dark pattern in financial services?

A nudge changes how a choice is presented while leaving customers free to choose and better off by their own standards, such as a reminder before a contribution deadline. A dark pattern works against the customer's interest, such as a cancellation path far longer than the sign-up. The practical test is whether the design still works once customers understand it.

Do financial disclosures protect consumers?

They can, depending on who delivers them and when. Research reviewed in our study with the Ontario Securities Commission found that conflict-of-interest disclosures can make the messenger seem more credible. In our trial, a disclosure shown as a platform label reduced purchases of promoted assets without removing the effect, so firms should test a disclosure's source and placement before relying on it.

How can a bank or insurer tell whether a behavioral intervention is ethical?

Set a customer outcome next to the business metric before testing, such as debt repaid or cover kept. Compare results against a control group and break them down by segment, including older and first-time customers. An intervention that lifts revenue while worsening the customer outcome for any group should not scale, however well it performs on average.

Does The Decision Lab work with financial regulators?

We co-authored the Ontario Securities Commission's 2025 study of finfluencers, which paired a survey of 655 Canadian retail investors with a randomized trial of 1,465 Canadians, and we redesigned the regulator's investor education library. We also work with insurers, banks, fintechs and debt-relief nonprofits on customer decisions, from sales calls to debt repayment.

Can behavioral science reduce investment fraud?

Controlled tests suggest it can. In our experiment, a short personal risk assessment made adults aged 60 and over less willing to invest in a fraudulent pitch. In our trial with the Ontario Securities Commission, prebunking and inoculation messages reduced purchases of promoted assets without removing the effect. Both kinds of intervention work alongside platform controls and enforcement and cannot stand in for them.