How to Build Customer Trust in AI Products: Designing for Appropriate Reliance

Last updated October 30, 2026

Just over 1 in 4 US adults (27%) say they trust businesses "some" or "a lot" to use AI responsibly, down from 31% a year earlier, according to the 2026 Bentley-Gallup Business in Society study of 3,270 adults. Among adults under 30, the share fell from 3 in 10 to 1 in 5.

Building customer trust in AI products means designing for appropriate reliance: customers rely on the AI when it is likely to be right, and pause to check or override it when it may be wrong. At The Decision Lab, we design customer-facing AI around four behavioral design principles that shape that judgment, and we test how customers act on the AI's output before launch.

1. Disclose AI use clearly, at the point where customers act on its output.
2. Give customers real control to adjust an output or hand the task to a person.
3. Signal uncertainty at high-stakes moments, instead of relying on explanations alone.
4. Prepare for the first visible error with a quick way to correct it and a clear route to a person.

Before launch, we check that customers accept the AI's correct outputs far more often than its incorrect ones.

We have worked on this problem since before generative AI reached consumers. For Mila, the Quebec AI institute, we turned COVI, a contact tracing app whose AI model calculated a graded COVID-19 risk score, into a consumer app people could understand and act on, and we carried our decision model for the app into working sessions with Health Canada and public health leaders.

This guide is for teams building AI into products customers use directly, such as advice tools, chatbots, recommenders and risk scores. Our guides on workforce AI adoption cover the employee side.

In their review of trust in automation, published in Human Factors in 2004, John Lee and Katrina See argued that designers should aim for trust calibrated to what a system can actually do. Too much trust leads people to rely on automation when they should not, and too little leads them to reject it when it would have helped. Appropriate reliance is the behavior that calibrated trust produces.

Trust versus reliance

Trust is an attitude, and reliance is what customers do with the AI's output. A survey can show the first, and only product behavior shows the second. For customer-facing AI, more trust is not always the goal. A bank whose customers accept every AI-generated budgeting suggestion has a problem as serious as one whose customers ignore them all, and the first problem is harder to spot because high acceptance rates look like success.

1. Undertrust. Customers avoid or ignore an AI feature that would serve them well. This one tends to show up early, as low uptake.
2. Overtrust. Customers accept the AI's output without checking, including when it is wrong.
3. Confusion. Customers do not know what the system is doing or whether a person is involved, so they have no basis for deciding when to rely on it.
4. Collapse after error. Customers who trusted the system abandon it after one visible mistake. In a series of forecasting experiments, Berkeley Dietvorst, Joseph Simmons and Cade Massey found that people lost confidence in an algorithm faster than in a human forecaster after seeing both make the same mistake, even when the algorithm was the more accurate of the two.

We start each project by working out which of these is happening for which group of customers, because the fixes pull in different directions. Adding friction helps with overtrust and makes undertrust worse.

1. Disclose AI use clearly and at the right moment

Across 13 experiments with more than 5,000 participants, Oliver Schilke and Martin Reimann found that people who disclosed using AI in their work were trusted less than people who did not. Being found out was worse: the steepest drop in trust came when hidden AI use was discovered by someone else.

Those studies looked at individuals, so for products we treat the same pattern as a design hypothesis to test. A typical test compares up-front disclosure with disclosure only on request, and tracks whether customers keep using the feature after they learn how it works.

In practice, we recommend disclosing at the point where a customer will act on the output, in words that say what the AI did with what information, such as a note that a spending summary was generated from the customer's last 12 statements. A generic AI badge gives customers little they can use to decide how much to rely on the answer.

Disclosure is also a legal requirement in some jurisdictions. Since 2 August 2026, Article 50 of the EU AI Act has required providers of AI systems that interact directly with people, such as chatbots and voice assistants, to make sure users are told they are dealing with AI by their first interaction, unless it is already obvious from the context. The Act covers providers outside the EU whose systems are used there, so North American companies with EU customers are in scope.

Disclosure does not shift responsibility away from the company. In February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay a customer C$812.02, including C$650.88 in damages, after its website chatbot gave him wrong information about bereavement fares. The tribunal rejected the airline's argument that the chatbot was responsible for its own statements and found Air Canada "responsible for all the information on its website."

2. Give customers control over AI outputs

In a later set of experiments, Dietvorst, Simmons and Massey found that people were much more willing to use an imperfect forecasting algorithm when they could adjust its forecasts, even when the adjustment allowed was severely limited. They also performed better and were more satisfied with the process.

In a customer-facing product, control can be small: editing a draft before it is sent, moving a recommended budget up or down, excluding a category from recommendations, or asking for a person. What counts is that the customer's input visibly changes the result.

We treat one further point as a design hypothesis to test: a control that changes nothing, such as a feedback button with no visible effect, may make a system feel less responsive than having no control at all. The test compares a real adjustment control with a passive feedback control, and measures whether customers notice and use each one.

3. Communicate uncertainty when it changes the decision

Explanations are the most common response to the trust problem, and the evidence on them is mixed. In a study presented at the CHI 2021 conference, Gagan Bansal and colleagues found that AI explanations made people more likely to accept the AI's recommendation whether or not it was correct, and did not improve how well people and AI performed together. Yunfeng Zhang, Q. Vera Liao and Rachel Bellamy found that showing a confidence score with each AI prediction helped people calibrate their trust, while the explanations they tested did not deliver the same benefit.

Zana Buçinca, Maja Barbara Malaya and Krzysztof Gajos tested designs that push people to think before accepting an AI answer, such as asking for their own judgment first. In an experiment with 199 participants, these designs reduced overreliance more than standard explanations, but participants rated the most effective designs least favorably.

That trade-off shapes how we design. A customer who commits to an answer before seeing the AI's suggestion is less likely to follow a wrong one and less likely to enjoy the product, so we reserve that kind of friction for high-stakes moments such as moving money or acting on a health risk, and use lighter uncertainty signals elsewhere. In Mila's COVI app, graded risk guidance replaced blunt exposed-or-not alerts, so each message carried its level of concern with it.

4. Design for the first visible AI error

The first error a customer notices is a predictable risk, and it is worth designing for before launch. We recommend setting expectations about accuracy before customers rely on the system, making errors easy to report and correct inside the flow, showing what changed after a correction, and giving a clear route to a person when the stakes are high.

Whether telling customers about accuracy up front helps them keep using a system after its first visible mistake is a hypothesis we test directly. One group is briefed on accuracy before first use, and we compare how many customers in each group keep using the feature after they notice an error.

Measure what customers do as well as what they say

Survey questions about trust measure attitude. We pair them with behavioral measures, because stated trust and actual use do not always move together, and a design change can shift one without the other.

Test acceptance of correct and incorrect AI outputs

The behavioral measure we use is calibrated reliance: the share of correct AI outputs customers accept, compared with the share of incorrect outputs they accept. When customers accept nearly all correct outputs and few incorrect ones, reliance is well calibrated. When they accept both at similar rates, reliance is poorly calibrated, however high overall acceptance looks. We use this as an operational design measure in testing, and it is not an industry-standard score. Measuring it requires cases where the right answer is known, either from real outcomes or from errors seeded deliberately in testing.

A controlled experiment comparing design versions gives the strongest evidence, but it is not always necessary or practical. Earlier in development, simulation and quick behavioral tests can catch many design failures at a fraction of the cost.

• For a global consumer electronics company, we built a behavioral design process for its digital health team and proved it on a mindfulness service inside a health app used by 60 million people. Every target behavior was tested through experimentation before any production code was written, and design time fell by three quarters even though the new process added research steps.
• For a digital wellness platform serving 25 million members, we ran the behavioral engine behind its recommender against synthetic members in a validation harness before it touched the product, so design failures surfaced in an afternoon instead of six months into production.
• For Mila, we ran weekly surveys of Canadians throughout the pandemic, measuring trust in tracing apps and privacy concerns, and fed the results into the app's design.

We advise against fully automated customer-facing AI when an error would be costly and customers have no practical way to correct it or reach a person. The same applies when a team cannot yet measure how often customers accept wrong outputs, since overtrust stays invisible until it causes harm. In those cases, we recommend keeping the AI behind a person, for example by having it draft responses that a staff member reviews and sends, until testing shows customers rely on it appropriately.

1. Financial services. AI budgeting tools, fraud alerts, suitability checks and investment guidance all ask customers to act on a machine's judgment about their money, and overtrust costs them directly. Our investor protection work with the Ontario Securities Commission shapes how we approach these products.
2. Digital health and wellness. Symptom checkers and recommenders have to produce advice people act on, and a clinically correct recommendation can still go nowhere. We built the per-member decision logic behind a wellness platform's recommender for exactly this reason.
3. Mental health services. Chat and coaching tools used by people in distress need escalation to a human that works every time. We led the redesign of Kids Help Phone's digital support gateway from research to launch, and we study how AI is changing youth mental health journeys.
4. Customer service. Chatbots that answer questions about policies and prices speak for the company, and the Air Canada ruling shows the company owns every answer they give.

If you are building a customer-facing AI product, our team can test whether customers rely on it appropriately before launch. You can talk to us about behavioral testing for AI products or see our behavioral science consulting services.

How do you build customer trust in AI products?

Aim for appropriate reliance: customers follow the AI when it is likely to be right and check or override it when it may be wrong. In practice that means plain disclosure, real control over outputs, uncertainty signals at high-stakes moments and a planned response to the first visible error, tested with real customers before launch.

What is appropriate reliance on AI?

Appropriate reliance means people follow an AI system's output when it is likely to be right and check or override it when it is likely to be wrong. A product can show high acceptance rates and still fail customers if they accept wrong answers as readily as right ones, so acceptance alone says little about it.

Does telling customers they are using AI reduce trust?

It often lowers trust a little at first. Research by Oliver Schilke and Martin Reimann found that disclosing AI use lowered trust, but hidden AI use discovered later lowered it most. In the EU, chatbots must now tell users they are talking to AI unless it is obvious, so the practical question is how to disclose well.

Do explanations help customers trust AI appropriately?

Explanations alone usually do not. In experiments, AI explanations often made people more likely to accept recommendations whether or not they were correct. Confidence scores, and designs that ask people to form their own view before seeing the AI's answer, did more to reduce overreliance, though people found the second kind less pleasant to use.

How do you measure customer trust in an AI product?

Pair a survey of stated trust with a behavioral measure. We compare the share of correct AI outputs customers accept with the share of incorrect outputs they accept. A wide gap means reliance is calibrated. Where the stakes justify it, we test design changes against a comparison group before rolling them out.

Why do customers stop using AI after one mistake?

Experiments on algorithm aversion found that people lose confidence in an algorithm faster than in a human after seeing both make the same error. Follow-up experiments by the same researchers found that letting people adjust an imperfect algorithm's output, even slightly, made them more willing to keep using it.