Last updated October 23, 2026
A Gartner survey of 5,728 customers, run in December 2023, found that about 3 in 4 of them used self-service at some point while trying to fix a service problem. Only about 1 in 7 of their issues were fully resolved there, and for issues the customers themselves described as very simple, the share was just over 1 in 3.
Behavioral science improves customer self-service by designing for the decisions customers make inside it: whether to try it at all, whether the answer they find fits their problem, whether to trust that answer enough to act on it, and whether they can reach a person when it falls short. At The Decision Lab, we work on those decisions inside service journeys, from the agent conversations of a top Canadian insurer to the connected service model that BDC, Canada's bank for entrepreneurs, is now implementing. In our experience, the self-service journeys that fail most often are organized around the company's departments and products, while customers arrive thinking only about their own problem.
The stakes have risen as companies put AI chatbots at the front of support. In 2023, the US Consumer Financial Protection Bureau warned that poorly deployed chatbots can leave customers in repetitive loops with no route to a human. Our guide to customer trust in AI covers how customers decide whether to rely on an AI system. This guide covers the wider set of help centers, knowledge bases, chatbots and escalation paths those decisions happen in.
Customers avoid help centers and chatbots when they expect to waste their time, cannot see their own problem in the content, or doubt they will be able to reach a person. Support contacts fall without a cost to satisfaction when self-service resolves more of those problems and the route to a person stays visible, so customers who do call are the ones who need to.
The four customer decisions that determine whether self-service works
Every self-service journey asks the customer to make four decisions in sequence, and a journey that fails at any one of them ends in a contact or an abandoned problem. The six design principles further down are how we work on each of these decisions in practice.
| Decision | What the customer is judging | Behavioral mechanism | Principles that address it |
|---|---|---|---|
| Start | Whether self-service is worth trying at all | Expected effort, learned from past attempts | 3. Keep a person in sight; 5. Set honest expectations for chatbots |
| Find | Whether the content describes their situation | Recognition of their own words and circumstances | 1. Write for the words customers use; 6. Notice struggle and step in |
| Act | Whether to trust the answer enough to follow it | Uncertainty about what happens next | 2. Answer the next question before it is asked; 5. Set honest expectations for chatbots |
| Escalate | Whether they can reach a person without starting over | Perceived risk of being trapped, and the effort of repeating themselves | 3. Keep a person in sight; 4. Hand off without making customers start over |
Why customers avoid help centers and chatbots
Customers avoid help centers and chatbots when they expect wasted effort, cannot recognize their issue in the content, need reassurance a page cannot give, or do not trust that they can reach a person. Each is a reasonable judgment from the customer's side, and each points to a different design fix.
Help centers and chatbots tend to fail in different places. Help centers usually fail on findability and fit: the answer exists, but under a name the customer would not search for, or written for a situation slightly different from theirs. Chatbots more often fail inside the conversation, by repeating the same suggestions or promising more than they can do, and at the exit, when the route to a person is hidden or starts the conversation over.
When either fails, customers call straight away, search elsewhere for an answer, come back later through another channel, or drop the problem altogether. The four reasons below explain why customers expect that outcome before they start.
1. They expect to end up on the phone anyway
In a study of more than 75,000 customers who had used phone or self-service channels, published in Harvard Business Review by Matthew Dixon, Karen Freeman and Nicholas Toman, 57% of inbound calls, a little over half, came from customers who had gone to the company's website first. A customer who has tried a help center and still had to call has learned that the phone is the reliable route, and skips the website the next time.
2. The content is written in the company's language
In the same Gartner survey cited above, the most common reason self-service failed, in about 2 in 5 cases, was that customers could not find content relevant to their issue. Almost half of the customers who started in self-service said the company did not understand what they were trying to do. In our experience, help centers tend to be organized by product and policy name, while customers search by symptom, such as “charged twice” or “can't log in”.
3. The problem feels too risky to handle alone
Michelle Kinch and Ryan Buell's research on customer anxiety, published in Management Science, starts from a long-standing finding in social psychology: anxious people seek advice from others. Financial services and healthcare are full of anxious moments, and companies in both are increasingly moving customers to self-service. A customer worried about a loan or a diagnosis wants a kind of reassurance that a knowledge base article cannot give.
4. They fear being trapped
The Consumer Financial Protection Bureau's 2023 issue spotlight described customers caught in repetitive loops of unhelpful answers with no way to reach a human, and warned that chatbots can fail to recognize when a customer is invoking their legal rights. Gartner's December 2023 survey found the same worry among customers: the concern they cited most often about AI in customer service was that it would make it harder to reach a person, a finding we cover in our guide to designing for customer trust in AI. Once a customer believes the exit is hidden, every minute spent in self-service feels like a minute lost.
Six behavioral design principles for self-service customers will actually use
Each principle below works through a specific behavioral mechanism, named at the start of the principle, and each comes with the evidence behind it.
1. Write for the words customers use
Mechanism: recognition. Customers scan for the words already in their heads, and they skip content written in other words even when it holds the answer.
Customers find answers when the help center is organized around their symptoms instead of the company's product names. When Travelocity rewrote its help section to remove jargon and simplify the layout, use of its “top searches” feature doubled and calls fell by 5%, according to the same Harvard Business Review study. That figure is company-reported rather than measured against a control group, but the change is cheap enough for any team to test.
The raw material for this work already exists in most support teams: search logs, chatbot transcripts, call recordings and agent notes. At a top Canadian insurer, we analyzed more than 170,000 hours of agent sales calls, using process mapping to find the critical moments in each conversation and machine learning to cluster where calls went wrong. Applied to support conversations, the same analysis shows which words customers use for each problem and where they lose the thread.
2. Answer the next question before it is asked
Mechanism: uncertainty. A customer who does not know what will happen after following an answer is likely to call to check.
A good answer resolves the current issue and the one that usually follows it. In the Harvard Business Review study, about 1 in 5 repeat calls (22%) involved a downstream issue connected to the original problem, even when that problem had been solved. Bell Canada mined its contact data for issues that tend to arrive together and trained its agents to address the likely follow-up in the first call, and it reported 16% fewer calls per service event and 6% less customer churn.
In self-service, this means ending each article or chatbot answer with the next step most customers in that situation will need, such as how to read the first bill after a plan change.
3. Keep a person in sight
Mechanism: perceived risk. Knowing that help is available makes it safer to try on your own first.
A visible route to a person makes customers more willing to stay in self-service, and few of them actually take it. In Kinch and Buell's field experiment inside a credit union's online loan-approval process, customers who received an invitation to connect with a human loan agent accepted their approved loans at a rate 24% higher, about a quarter higher, than customers who received the same messages without the invitation. Very few of the invited customers chose to contact a person.
In practice, this means showing the contact option and the expected wait on the same screen as the self-service content, and never requiring a customer to fail a set number of times before the option appears. After we helped coordinate a redesign of Wellness Together Canada, the government-funded mental health support platform, we led research and product design for its companion app PocketWell, and the app gives mobile users quick access to self-assessment tools alongside phone counseling and crisis support services.
4. Hand off without making customers start over
Mechanism: effort and repetition. Every time a customer has to tell their story again, the work of getting help goes up, and so does the sense that the company is not keeping track of them.
A handoff works when the person who picks up already knows what the customer tried. In the Harvard Business Review study, more than half of customers (56%) reported having to re-explain their issue, and a similar share reported being transferred.
A chatbot or help-center handoff should pass the conversation, the articles viewed and the customer's own description of the problem to the agent, and the agent's first sentence should show they have read it. In our work with BDC, Canada's bank for entrepreneurs, the service model we helped define starts from the same idea: the entrepreneur explains their ambition once, and account managers remain the point of trust while AI coordinates the work behind them.
5. Set honest expectations for chatbots
Mechanism: expectation violation. Disappointment grows with the gap between what a chatbot seemed to promise and what it delivered.
A chatbot that presents itself as more capable than it is turns a customer's frustration into anger at the company. In five studies, including an analysis of real chatbot conversations at an international telecommunications company, Cammy Crolic, Felipe Thomaz, Rhonda Hadi and Andrew Stephen of the University of Oxford's Saïd Business School found that a humanlike chatbot lowered satisfaction among customers who arrived angry, along with their evaluation of the company and their intention to buy again. The effect disappeared when the chatbot lowered expectations of what it could do at the start of the chat.
Showing the work helps too. Ryan Buell and Michael Norton's experiments on the labor illusion found that people valued a website more when it showed the work it was doing, even when it returned the same results more slowly than a site that answered instantly. A chatbot that says which order it checked or which policy it applied gives the customer a reason to believe the answer, as long as the work it shows is real.
6. Notice struggle and step in
Mechanism: timing. Help offered while a customer is still trying arrives before they have decided to give up.
Customers signal that self-service is failing long before they call: they search again with new words, rephrase the same question to the chatbot, move back and forth between articles, or type “agent” into the chat window. EarthLink assigned a team to offer a live chat to customers who had spent more than 90 seconds in its knowledge center or clicked its contact link, and it reported 8% fewer calls.
For American Financial Solutions, a nonprofit debt management program, we built a model that traced dropout risk to elements of the first onboarding call, months before clients quit, turned it into an early-warning process for clients at risk, and redesigned the calls, texts, emails and letters around it. The program's dropout rate fell by half after the redesigned touchpoints were in place. Support teams can use the same logic: identify the signals that come before a failed self-service attempt, and reach the customer while they are still trying.
How to reduce customer-support contacts without hurting satisfaction
Reduce support contacts by making self-service solve more problems and by spotting struggling customers early, while keeping the route to a person easy to find for those who need it. Hiding the phone number or adding steps before the contact form also cuts contacts, but the unsolved problems move into churn and repeat visits.
Ryan Buell, Dennis Campbell and Frances Frei's study of 26,924 customers of a nationwide retail bank shows why contact volume alone can mislead. Customers who did a larger share of their banking through self-service were no more satisfied than customers served face to face, and in some channels they were less satisfied. Where heavy self-service users stayed with the bank, the data pointed to switching costs, and heavy users of channels that were easy to leave defected more often than other customers. We suspect many contact-reduction programs are counting customers who have given up or are locked in as successes.
Averages can also hide who pays for a lower contact rate. An overall drop in contacts can sit alongside worse outcomes for first-time customers, customers in financial hardship, customers with disputes or others at higher risk, so we break every measure down by customer group before calling a change a success.
What to measure instead of deflection alone
We recommend one primary outcome: repeat contact across every channel within 7 to 14 days, because it shows whether the problem was actually solved. The Customer Effort Score and the share of handoffs that reach an agent with context sit beside it as supporting measures.
The repeat-contact window and the Customer Effort Score both come from the Harvard Business Review study, where the effort score predicted customers' intention to buy again and spend more better than satisfaction scores did.
| Metric | What it measures | Failure mode to watch for |
|---|---|---|
| Repeat contact within 7 to 14 days (primary) | Whether the issue was solved, across every channel | Looks worse than first-contact resolution scores, because it catches more |
| Customer Effort Score | How much effort the customer says it took, from 1 (very low) to 5 (very high) | Asked only after calls, so self-service effort goes unseen |
| Handoffs with context | Share of escalations where the agent receives the conversation so far | Agents receive the history but reply without reading it |
| Return to self-service | Share of customers who start in self-service again on their next issue | Counting returns driven by a hidden contact option |
| Deflection or containment rate | Share of self-service sessions that end without a contact | Counts customers who gave up as successes |
Test changes against a comparison group
Most self-service changes can be tested the way product teams test features: some customers see a new article or chatbot opening, and the rest see the current version. At the top Canadian insurer mentioned above, we piloted new call scripts and agent training against a control group, and sales in the first pilot rose 12% compared with the control group. That was a sales project, but support teams can run the same kind of pilot with repeat contacts and customer effort as the outcomes. Where a randomized test is not practical, comparing matched groups of customers before and after a change still tells a team far more than the deflection rate alone.
When self-service is the wrong answer
Some service moments should start with a person, or offer one immediately. These include disputes where a customer may be invoking legal rights, which the Consumer Financial Protection Bureau warned chatbots can fail to recognize, and moments of high anxiety such as a declined payment. They also include customers who arrive angry, the group Crolic and colleagues found humanlike chatbots hurt most, and customers who have already failed in self-service on the same issue. For these moments, good self-service design means a fast, clearly signposted route to someone who can help, with the customer's history attached.
How The Decision Lab designs and tests customer self-service
Our self-service work draws on our behavioral product design practice and follows four steps, each of which leaves the client's team with something it keeps.
- Map the decisions. We trace customer journeys through every self-service and assisted channel, using search logs, transcripts, journey data and interviews with agents, and mark the points where customers decide to keep going or reach for the phone.
- Diagnose the barrier for each group of customers. The same failed journey can have different causes for a first-time customer and an anxious one. For the Canadian Cancer Society's text-message quit-smoking service, our analysis found four groups of smokers with different main barriers, and we built a separate conversation flow for each.
- Design and test. We redesign the content and the handoffs using behavioral design thinking, then test the changes against a comparison group before they scale.
- Build and hand over. We stay involved until the winning version is running, and the client's team owns the methods and assets afterward.
Related guides on customer behavior and AI
- Behavioral science for customer trust in AI covers how to design transparency and control so customers rely on AI products appropriately.
- When to use behavioral economics instead of customer research sets out six signals that a problem needs behavioral experiments rather than more interviews.
Frequently asked questions
What is behavioral science in customer service?
Behavioral science in customer service applies research on how people actually make decisions to the design of support journeys. It explains why customers start or skip self-service, and why they give up and call. At The Decision Lab, we use it to redesign help content and handoffs, then test the changes against a comparison group before they scale.
Why do customers prefer talking to a person?
Customers turn to a person when a problem feels risky or emotional, because anxious people seek advice from others. Many have also learned from earlier attempts that self-service ends in a call anyway. Offering a person inside the self-service journey can improve outcomes even when few customers use it, as Kinch and Buell's credit union field experiment showed.
Does adding a “talk to a person” option increase call volume?
It does not have to. In Kinch and Buell's field experiment, very few customers who were invited to contact a loan agent did so, while the share accepting their approved loans rose. Results vary by setting, so test the option with a share of customers and compare contacts and outcomes against customers who do not see it.
How do you measure whether self-service is working?
Track repeat contacts across every channel within 7 to 14 days, and ask a Customer Effort Score question after self-service sessions as well as after calls. A deflection rate on its own counts customers who gave up as successes, so pair it with at least one measure of whether the problem was actually solved.
Should customer service chatbots seem human?
For routine questions, a humanlike chatbot did no measurable harm in the Oxford research. For complaints and other moments when customers tend to arrive angry, humanlike chatbots lowered satisfaction because they raised expectations the chatbot could not meet. In the same research, a chatbot that stated its limits at the start of the conversation removed that negative effect.
