How to Implement Disaster-Free AI in Service: Ease of Escalation is the Key!

Almost every organization I speak with is grappling with how to implement AI. Management assumes it will be in place within months with great cost savings. As usual, this is a grossly oversimplified view of the multiple issues under the AI/bot label.

The keys to AI success in service are to (1) assure quality information and (2) understand when to exit or to allow the customer to escalate. Recently, I needed a specialized quote from Geico and called the phone number in the app. I barged into the IVR menu, asked for a representative, and stated my need. The CSR acknowledged that she could not help me and transferred me to someone who could. Despite trepidation, I was connected to the right person within 60 seconds. Geico produced a delighted customer by clearly defining the self-service divide and mastering the escalation process. This also has fostered positive word-of-mouth (WOM) recommendations.

To place AI in its proper context, there are four steps.

1. Categorize your service workload into simple, predictably difficult (with a limited set of solutions), and complex or emotionally messy.
2. For the difficult and messy, map their complexities, and the usual causes of dissatisfaction and escalation.
3. Equip AI with high-quality knowledge and protocols to handle difficult issues. These will include mechanisms to identify dissatisfaction with the reply and immediately offer escalation as well as processes to create delight.
4. Aggressively measure outcomes of response to difficult and messy issues. Flag where AI fails leading to escalation as well as when it seems to be successful, allowing you to expand its domain.

I’ll first review the cost of not following these steps and then describe factors that lead to the need for escalation.

Why Care About Ease of Customer Problem Escalation?

When customers are not satisfied with a response, they either give up or try to escalate. If no easy channel is clearly available, 50-90% of customers give up. If they attempt to escalate and encounter barriers (described below), more damage is done and anger increases. Anger then often causes irrationality (due to blood draining from the brain to the muscles) resulting in further demands and difficulty gaining a positive resolution.

This figure shows customer behavior when encountering a service frustration.

The majority of both consumers and business customers abandon their inquiry or complaint with varying levels of damage to their loyalty, price sensitivity, and WOM behaviors. If the customer does attempt to escalate to another channel and struggles, the damage to the relationship is usually double compared to when they abandon. The struggle becomes more memorable and often creates a second problem — anger or rage. Addressing the second problem can add significantly to the cost of servicing the customer.

On the other hand, easy escalation is nowadays unexpected and therefore a pleasant surprise or delight, resulting in increased loyalty and positive WOM.

The challenge is to get management and operations to understand that, given the revenue value of most customers, easy escalation is a value-add that results in a higher long-term value of the customer and much less price sensitivity.

Market Damage of Escalation Frustration

Frustration with escalating a problem to a human is salt in the wound. This is what causes 67% of Americans to feel rage as described in the latest National Rage Study.

• Loyalty Damage of Abandonment. Across all industries, when a customer has a problem that is not resolved, there is a 20% decrease in loyalty. This means that for every five customers with an issue, one out of five will be at risk.
• Damage of struggle to escalate. If the customer complains and gets a poor response, damage grows to at least 30%. One out of three customers is at risk. A failed escalation results in 50% damage — one out of two customers is lost. Not only did the company have a bad product, but the struggle showed it didn’t care.
   
  When Finance objects to escalation to a human costing $25, place that within the context of a customer worth $500 to thousands of dollars. Help Finance to do the math. Is it worth saving $25 while losing $250 (50% X $500)? At a minimum, the ROI of easy escalation is 200%, and in this case, $83/$25=330%, assuming a 33% gross margin.
• Poor response followed by barriers to escalation creates irritation and another problem. The single largest cause of consumer rage according to the latest CCMC 2023 National Rage Study is difficulty reaching a human. You can encourage the use of smart bots as a quick-access intermediate step. However, if the customer rejects it, escalate them immediately to a human channel, with both estimated time in the queue and a virtual queue option for call-back.
• Higher service cost. Angry customers take much more time and make additional, often unreasonable demands. When you get upset, the blood drains from your brain to your muscles and you truly are irrational. At Toyota, we found that frustration added an extra 90 seconds to talk time.
• Damage to staff morale resulting in 20% higher turnover. I recently was talking to a car rental CSR at the Preferred Desk late at night and she could neither connect me to a supervisor nor communicate with the airport manager where I was calling from. She agreed that she felt helpless and could not escalate to the management needed to mitigate the mess that I was in. That call had wasted her time and my time. Work with HR to arrive at a number to add in when you talk with Finance about enhancing ease of escalation.
• Regulatory Intervention. In July 2024, The Consumer Financial Protection Board (CFPB) floated the idea of a regulation requiring easy access to a human, ideally via one click. The NY Times article noted that the FTC and FCC were watching closely to consider similar actions. Such a draconian rule would add dramatically to service costs, especially as one explores what “access to a human” would mean.
• Pool of Negative WOM and social commentary. Negative WOM can kill a marketing program. On the other hand, positive word of mouth (PWOM) is the most powerful marketing dynamic. Companies like USAA, Chick-fil-A, and Harley Davidson acquire more than 70% of their new customers via WOM. Easy escalation is a powerful PWOM generator. Escalation struggles can kill a product.

Drivers of successful escalation

If a company accepts that it is better to escalate than to lose the customer, management must break down the barriers to escalation. There are three approaches.

• Message inviting escalation. There are two necessary approaches to the invitation. First, a bot or CSR must sense when a customer is unhappy with the response and, after trying one more time, so state and offer escalation with a message such as “I’m not giving you an acceptable answer, can I refer you to a better helper?” The second approach is to overtly invite escalation with a message like TELUS uses: “Escalate to our Management Team. If your issue remains unresolved after contacting customer service, please complete the form below.”
• Availability when needed. The invitation must be visible or offered exactly when the customer expresses dissatisfaction, whether via text or voice. If the customer asks the same question using slightly different terminology, the AI needs to see the same issue is still at hand.
• Immediate visibility of escalation channel. The offer to escalate must be physically in front of the customer and visible or offered by voice to the customer. Having an offer at the bottom of the webpage is not helpful.

Two other requirements for easy escalation besides the invitation and channel are:

• Ease of submission with one click or request. Data carried along from an electronic or voice escalation must be accompanied by a full history of the customer’s previous interactions.
• Proper expectations set for response time. If there is a wait to connect with the escalation channel, a wait-time range should be explicitly communicated.

Clearly Define the Self-Service / Human Divide

First, categorize the workload into:

• Simple. Straightforward transactions such as password resets, change of address, and simple orders. There should not be a need for escalation unless the customer is clueless.
• Complex. New orders, questions about specifications, size, variable insurance coverage, explanation of charges. In many cases, AI can respond to these if provided with a set of flexible rules, called flexible solution spaces (FSS). FSSs can be used if a half dozen FSSs can address at least two-thirds of the situations. Then the issue can be handled by AI provided that easy, immediate escalation is available for the other one-third which fail.
• Complex/Emotional Requiring A Human. These are issues that are so complex and emotional that an FSS addressing the six most standard situations will not be sufficient to handle a majority of issues or the need for empathy is a critical factor. A good example would be a call concerning receipt of a diagnosis of cancer. Keith Farley, SVP of Aflac, reported in CMSWire that their contact center has a separate split for policyholders who have received such a diagnosis.

For the complex and human-needed, map the complexities. What information is needed, where does it come from and what decisions need to be made to create a complete response to the customer? 

Most critically, what are the probable causes of escalation? As much as possible design them out of the process or anticipate them. For example, if a Gold-level traveler wants to cancel or change a reservation that would normally incur a service charge, the FSS can look at the charge and the value of the customer and waive the charge while educating the customer that this is a one-time courtesy. This can be built into the FSS with no need for a human or supervisor.

The following are the usual reasons for escalation. All are predictable and can be identified in advance:

• Delay in response or access. If proper wait time expectations are not set, customers become impatient. In all cases, immediate acknowledgment is required. Reasonable response wait times (unless otherwise set) for chatbots is 60-90 seconds and for email is two hours for B2B and four hours for consumers. Social posts vary from four hours to one day.
• No response. Once the expected response time has passed, the customer has a second problem in addition to their original – the company has ignored their plea. Escalation is attempted.
• Confusing process. For example, an endless loop or confusing IVR or multiple demands for the same information
• Response does not address all customer issues . The response may have been partial or did not resolve the main issue. It is critical to view responses from the customer’s perspective, not the company’s. To say, “This request is against our corporate policy” is NOT a responsive answer. At a minimum, the response must justify why the response is logical and has treated the customer fairly, regardless of policy.
• Lack of authority by the current level of staff. This is an extension of the “It’s not our policy” response above. Lack of authority is never viewed by customers as a legitimate reason for an employee not to do what is requested. “Let me speak with your supervisor” is the logical next request.

You can shift this divide by “empowering” AI to do more — if you trust it. You’ll have to live with its decisions. You can also reduce escalations to a supervisor by empowering the front line. You can also reduce escalations to a supervisor by empowering the front line. We’ve rarely seen a problem with further empowerment and the benefits are happier customers and staff and less supervisor burden. Finally, you can reduce the need for escalation by setting better expectations upfront. Transparent, simplified marketing can reduce escalations by 30%.

Encouraging Appropriate Escalation

Remember that if the customer is unhappy, you want them to escalate rather than go away angry and disloyal. This is the hardest concept to accept.

Once you agree with the value of escalation, there are four components to create.

• Detection of dissatisfaction. Either a human or AI are adept at detecting dissatisfaction – certain word patterns show the customer is unhappy. Another approach is using a one-question survey asking, “Did I answer your question?” With three possible responses, yes, somewhat, and no.
• Visibility of the offer of escalation. The offer must be in front of the customer just when needed. This can be next to the chatbot response panel or presented by the chatbot itself in a message like that described below.
• The escalation message. The message should convey: “You do not seem happy with my answer, would you like to escalate to a specialist, supervisor, or manager?” This is where most managers cringe, fearing that the customer will say “yes.” But, remember, escalation is profitable at least 90% of the time.
• The channel including expectations for accessibility Provide the channel including setting appropriate time in queue. If it is more than five minutes, offer a virtual queue where the customer will receive a call-back.

Measurement, Target Setting, and Quantification of Payoff

The CCMC 2023 National Rage Study and the CCMC 2024 CPG Complaint Studies both suggest that ease of escalation (EOE) of problems is now the key metric for success. Overall satisfaction scores have become meaningless as we are seeing high levels of both delight and dissatisfaction/rage across both digital and telephone workloads. The fact that most issues are well handled is not as important as what happens with the exceptions and at the murky boundary between simple/complex/emotional and personal/digital.

Therefore, a single number, which is an average of promoters, passives, and detractors provides next to no actionable data. What is very actionable is whether a customer had to escalate and if so, their ease of escalation rating.

Most of the world focuses on overall averages — “We got 92% of issues well handled!” Yippee Do!!! That 90% is nowassumed and expected by both consumer and B2B customers. What makes or breaks you is what happens with the 8% of atypical, exceptions. Machine learning can enhance success by 8% to 75-90% depending on the industry. A good target for the resolution of the murky half of the 8%. Your important metrics are:

• The percentage of escalations that are successfully escalated to the right place. A target is 85% of escalations, NOT of overall contacts.
• The percentage of escalations that are handled to successful completion on first contact. A reasonable target is 66%, with 85% as a stretch goal. This means that 1.2% of overall contacts (.15 X .08 % exceptions) must be researched off-line. There would then be sub-analyses of different escalation channels such as escalation to supervisors or a specialist such as the underwriting department.
• A third metric that is rarely examined is failed website searches. Failed searches are one form of self-service that is usually handled by the digital/IT staff and, in 80% of companies, are not even reported to CX or customer service. About 40% of failed searches result in abandonment rather than going to FAQs. Many others turn into phone calls.
• Quantification requires analysis of the number of customers escalating by each type of issue and their value vs. the cost of handling the escalation. To quantify the long-term payoff, send a survey by email to all those customers two weeks after the case is closed. While you’re skeptical, escalation is a high-involvement transaction where most customers will take the time to give feedback. You’ll be amazed at the level of loyalty and WOM of those who have successfully escalated.

Implementation of Disaster-Free AI in Service

Five steps are simple but not easy to implement:

1. Educate executives, especially in marketing, that escalation should be easy and flags expensive failures, lets off steam, and builds trust. Less than 10% of escalations are not justified. TELUS, the Canadian phone company overtly invites customers to escalate to management – which projects supreme confidence in the service system – and flags where issues need help. Lou Gerstner, the former CEO of American Express, provided his personal email in newsletters
2. Log every escalation and abandonment including the question or underlying issue. Use this data to analyze processes to be automated. Also, note the reason for escalation; no/slow response, incomplete response, lack of authority, failure to accept or understand the answer, or escalation to human without trying self-service due to lack of confidence in self-service.
3. Divide the workload into three categories: simple, complex and messy, and emotional. Apply backed by flexible solutions spaces to the complex middle category as well as the simple issues. Be conservative in your testing and measure every outcome until you learn which FSSs work well and which need tweaking.
4. Survey satisfaction with ease of escalation and identify the issues where escalation is difficult or fails to happen (e.g. customer abandons or gives up) – These are customers with Rage who will spread negative word of mouth and social posts.
5. Quantify your success in terms of customer retention. Correlate ease of escalation with overall satisfaction, NPS, and customer effort. You will find that EOE is a highly effective predictor of success on difficult, memorable issues.

Summary

The vast majority of customer transactions can be self-service. However, when a human is needed, the ease of escalation will determine whether the overall experience is memorably positive or negative.

• Inadequate AI and self-service answers cause customer rage and the need to escalate
• Encouraging escalation almost always has an ROI of more than 200%, although management often fears the cost
• The key metric is the ease of escalation of dissatisfied complainers
• Management must be educated on 1) the cost of inadequate escalation processes and 2) the huge payoff of empowered self-service and service reps to prevent the need for escalation