Generative AI will change customer service forever. Here's how we get there

Companies that want to exploit generative AI must focus on how the technology benefits their customers.
Written by Vala Afshar, Contributing Writer
people walking on a yellow background toward a door in the sky
Andriy Onufriyenko/Getty Images

Successfully injecting generative artificial intelligence (AI) into customer service requires a systematic approach. 

Improving the customer experience using generative AI solutions is about re-inventing legacy processes and optimizing how data is accessed, analyzed, and applied to make faster, more informed, and better decisions that serve all stakeholders -- employees, customers, business partners, and our communities. 

Also: Generative AI on its own will not improve the customer experience

The point we want to bring home is that processes must be checked and double-checked before organizations apply generative AI. Mistakes can come back to embarrass the company that builds the generative AI system. In a recent example at Chevrolet, an AI-enabled bot told customers false information that resulted in financial loss. A North American airline, meanwhile, attempted to make the legal argument that its bot was a separate entity whose answers were not the company's responsibility. 

In these situations, it's crucial to remember the machine cannot discern between good and bad data or good and bad processes. The machine only sees data and follows the processes and rules created for it.

Organizations should also ensure the data consumed by generative AI is accurate, and that the customer service processes they design can be successful using generative AI. Start small, start where there is measurable value and benefit, start simple, start with clean data, and invite a small number of loyal customers into your design process to ensure you are on a path to success. 

Also: AI's latest trick: Better KPIs for measuring business success

During our last post, we said generative AI alone will not improve the customer experience. We provided three key points of advice: data must be trusted, accurate, and available; the customer process must be accurate, intuitive, and fair; and there needs to be measurable value. Now we are shifting gears with a conversation about how organizations can accelerate the adoption of AI technologies, and we will look at some areas where generative AI will have the greatest impact on customer service.

The rise of ChatGPT shone the spotlight on customer engagement by empowering chatbots with generative AI technology. Chatbots have been a part of the self-service toolkit for decades. They failed to revolutionize customer service because they were not intuitive enough to answer questions with a rate of success that satisfied most customers. Now we have bots based on generative and predictive AI, and the expectations are high that these technologies will radically change self-service.

To better understand how generative AI adoption aims to improve customer experience and customer service, I interviewed two of the world's leading experts on customer relationship management (CRM), customer experience (CX), and customer service. 

Michael Maoz is senior vice president of innovation strategy at Salesforce. Before joining Salesforce, Maoz was research vice president and distinguished analyst at Gartner, serving as the research leader for the customer service and support strategies area. 

Ed Thompson is senior vice president of market strategy at Salesforce. Before joining Salesforce, Thompson was research vice president and distinguished analyst at Gartner, covering CX and CRM strategy and implementation. Maoz and Thompson shared their points of view on what businesses need to consider and implement before applying generative AI solutions to their customer service applications and processes. 

What key mistakes are businesses making? 

Ed Thompson - At the moment, companies fall into two schools of thought, or two approaches, as regards intelligent bots: they either start by thinking about enhancing the productivity of their agents, or they have as an aim the use of generative AI to reduce the number of customer service agents.

Managers of customer service departments are often skeptical about the promises of generative AI-powered bots as a tool to help with agent productivity. Over the years, promises were made by software suppliers that failed to materialize. By comparison, heads of digital, COOs, and finance heads are not as focused on agent productivity. They tend to be excited about reducing costs using the new generation of generative AI-powered bots.

Also: How renaissance technologists are connecting the dots between AI and business

Where organizations can go wrong is when they take the new generation of intelligent bots, only to make a lot of the same mistakes from the past when it comes to their implementation. Three of those mistakes are:

  1. To build or deploy standalone bots that are not part of an omnichannel solution, so the bots run separately to voice, email, WhatsApp, SMS, portal, and other service channels.
  2. To ignore the fact that the underlying data being used to provide an answer to a question is a mess, with the result that the answer to the customer is incorrect or unhelpful.
  3. Sending the customer into repeat loops they can't escape -- an approach that drives down satisfaction scores and drives up escalations.

Generative AI is bringing new concerns such as bias, hallucinations, and toxic answers. But the old concerns remain and may be more serious. Perhaps generative AI will provide the impetus to solve unintegrated and siloed data sources, out-of-date data, poor quality data, incorrect data, and bad service processes and workflow design. My recommendation is to avoid the mistakes of previous bot implementations.

What should you consider before starting an AI initiative? 

Michael Maoz - First, fix the knowledge base and clean up the customer data. When we add generative AI to a bad customer self-service process, we have made things worse, only at a greater scale and with more severe consequences than before. A chatbot that gives customers incorrect information without generative AI does so because the underlying knowledge base is incorrect. Once generative AI is added, the company will find that the incorrect information is now given to the customer with perfect confidence by the bot, and not because it is hallucinating. 

As a result, the company has more eloquent false information disseminated in a new form. Recently, we visited the website of an NBA basketball team. We asked the bot questions about the coach and we received accurate information about the coach. There was one problem -- he was not the current coach, but the former coach who had subsequently moved to coach a different NBA team. The bot did not have information about the current team coach because the knowledge base had not been updated. The result was an entertaining hallucination.

Ed Thompson - There is another alternative to moving forward with generative AI. If most of the knowledge base is old or unreliable, it might be easier and better to scrap it and start over from scratch with reliable content. We estimate that upwards of 40% of businesses will start again from scratch and generate knowledge articles from customer interactions (calls, chats, emails, etc.) using generative AI. For these businesses, it will be quicker and result in a dramatically lowered risk of hallucination from old and contradictory sources of knowledge.

Another 30% of service organizations will have either: good enough knowledge; knowledge that is good enough that they can trust it; knowledge that is good enough that they can fix it; or knowledge they can easily separate good from bad (its age, its location, some form of quality score). In these examples, it is not worth starting from scratch.

Also: The future of generative AI: Here's what technology analysts are saying

The final 30% will find the data is a mess but due to regulations or their industry or politics, they'll be forced to use it and either grind through cleaning it up over many years or give up as the task is too difficult.

Recommendation: Consider rebuilding the knowledge base from scratch.

Michael Maoz - With simpler service transactions, there are already great examples of generative AI handling up to 75% of customer requests. This can be raised to 95% over time. The reason the current results are less than future results revolves around the required education that a large part of the customer base needs to undergo before they trust the technology and are confident enough to give it a try. Not all customers are comfortable with self-service, and many chafe when interacting with software rather than a human. Businesses need to begin educating the customer about the efficacy of AI and the safeguards built into the technology, as well as communicating with the customers on how to use the technology and its benefits to them.

Recommendation: Educate the customer to build trust in the new generative AI technology

Ed Thompson - There is a decade's worth of research to show that, on average, the younger and the wealthier the customer, the more they prefer digital channels over human. Why? It is not because they're technically superior -- that's a fallacy. It's because they grew up with them without having to switch, and found that when they tried the human channels they were less useful (i.e. harder to reach, more variable in quality, and slow). 

Also: 4 ways to overcome your biggest worries about generative AI

We'll see more of a shift to digital as the older consumers and software users are outnumbered by Gen Z and Millennials over time. In the Popper and Kuhn philosophy of science debates of the 1960s, it was argued that the new theory of science would win out eventually over the old through rational debate and argument. To me, the new theories win because the old theory's supporters eventually lose influence or power or retire or die. But there are examples where older, less wealthy, and non-tech-savvy consumers adopt digital channels in the same proportion as those who are younger and wealthier. One example is WhatsApp. Yes, it is a simple tool and benefits from network effects, but it is also true that younger adopters taught their parents and grandparents how to use it.

Recommendation: Target your education at those least likely to adopt self-service.

Michael Maoz - That is a great point about generational divides and channel preference. Now, let's look at the process design of the digital channels. Given the capabilities of generative AI, early adopters are discovering a hard lesson -- unless the process is designed such that the customer can successfully and intuitively solve their issue, the use of AI without the backstop of reaching a human can undercut the customer experience. To accelerate success in onboarding generative AI for customer service, leading companies are bringing the customer into the process design cycle to gain insights and test their responses.

Ed Thompson - I'm continuously amazed when attending forums and conferences and hearing the discussion about channel shift or generative AI or deflection of customers -- and there's no discussion about what the customer wants. The discussion is not informed by what the companies' customers want. There is not even much discussion about what devices the customers are using. I'd say 75% of the discussion is inside-out -- what does the company want, with only the occasional reference to the impact on the customer that shows that the company sat and watched and listened to the customer.

What is the role of generative AI in the future of work? 

Ed Thompson - The future promise of customer service is beyond point solutions that answer simple questions, and lies in generative AI utilized in the flow of processes. That approach entails running the entire sales process, marketing process, or service process, whether in a reactive mode in response to a customer request or proactively as an autonomous agent reaching out to the customer to assist with an emerging issue. The long-term goal is for customer service to disappear as a separate department and instead morph into a personal assistant that is invoked sometimes without the customer asking for assistance and at other times in response to a customer request. 

Think of 'customer' including devices and objects such as vehicles, machinery, equipment, and infrastructure that connect through APIs and sensors to a company network. In these scenarios, an elevator, a bridge, or an engine can 'call' for assistance without the knowledge or intervention of a human, and the request is translated into the dispatch of a human agent or software agent or a combination to respond to the request.

Also: Generative AI advancements will force companies to think big and move fast

As an aside, we forget that the customer engagement center, known popularly as the contact center, evolved from just a center that took phone calls -- the call center. Only 60 years ago, there wasn't even a call center. It is amazing to think about and, perhaps, we can learn about the future from the past. How did businesses survive in those days? One way may have been through less specialization perhaps. That would mean that folk who sold goods and services also serviced the product as a part of their role.

Michael Maoz - That is a great question. Businesses weren't intentionally focused on customer experience as a differentiator even 40 years ago. They saw product quality and price as critical, and later they saw that the customer experience played an integral role in company growth. The companies sub-optimized what they did, meaning customer service was cut off overall from marketing and sales, and engineering and product development. That strategy worked until competitors latched onto the need to work back from the customer experience -- and reengineer products and services to meet the shifting needs of customers.

Ed Thompson - So, the specialization may have come in later, and is only a part of the story. Since post-World War Two, there has been explosive economic growth and the introduction of tens of thousands of new products never before imagined.

Michael Maoz - I guess some historical background helps. In the US in the 1940s, you picked up a telephone receiver and didn't dial. Instead, you waited for a voice to say, "How can I connect you?" No company had its own call center until about the 1960s, and that change represents an enormous shift to scale in addition to what you call specialization.

Also: How Adobe is leveraging generative AI in customer experience upgrades

On the other hand, in the United States, the Sears catalog was a printed book with thousands of items for sale and delivered by mail anywhere in the country. You could buy a pre-fab home, sewing machine, or anything for a kitchen, but it was only through a catalog and there wasn't a concept of after-sales support. You mentioned World War Two. Post-World War Two saw the rise of the suburbs and interstate highways, franchise businesses, and retailers that sold an infinite range of products across the country from local shops. National GDP tripled and consumers purchased goods and services faster than the manufacturers could service them. That process led to new post-sales service businesses and units forming.

Ed Thompson - Returning to my original point, we are where we are, with all of the intermediation and all the complexity of supporting the customer. I agree that the important breakthrough in business came during the 1990s with the work of visionaries such as Martha Rodgers and Don Peppers around the future of one-to-one marketing. That approach fits with your point about 'customer service' disappearing as a separate department, and further out seeing the possibility of each consumer or customer having a personal assistant that understands their needs. We might also see a time in the future when there is an end of mass marketing. Campaign management and mass marketing of emails will likely go away over the next five years. Instead, marketing will use the same idea of the personal assistant to curate the right offer at the right time across the right channel.

How should you approach generative AI? 

Michael Maoz - These are non-trivial projects, and we can expect IT and the lines of business (marketing, sales, customer service, billing, logistics) to find new collaborative opportunities. The business unit that wants to deploy generative AI solutions to handle customer service requests must make friends with the generative AI leader to improve the customer experience. 

Also: The 5 best AI art generators

The best companies understand that the core business processes need to be designed from the outside-in, which means from the customer perspective. This approach is hampered by the organizational design where core IT roles are not integrated into customer processes. The IT skills of integrating systems, building automation, migrating data, and creating data visualizations are far from the customer-facing roles of marketing, digital commerce, and sales and service. IT teams have generative AI skills, and marketing, sales, commerce, and customer service have deep insights into how customer-facing processes work best. Creating cross-departmental teams with shared accountability and shared success metrics helps focus generative AI initiatives on what works best for the customer, and on prioritizing the order of generative AI solutions.

Ed Thompson - We could summarize the insights for companies that will win with generative AI as:

  1. There are few to no generative AI skills for 95% of companies to hire. Don't think you can go and get talent using the traditional methods, and you cannot afford to wait for the capability to evolve.
  2. Visionary businesses will decide to upskill their employees whenever possible.
  3. An important place to start, after business process design, will be to put the focus on prompt design and data governance.
  4. Great businesses will ensure that the technical skills are not trapped inside IT.

What is the future of voice in generative AI?

Voice and text are handled differently by the brain, and each has an advantage in customer service. Human fluency in typing or texting is only one-third as fast as voice or image processing. There are reasons to prefer text in many situations, ranging from privacy to precision. There is also the advantage that humans are far less precise when they speak than when they type. For example, a person might speak to a human by saying they want to buy two tickets on the first base line at a baseball game: "I'd like two tickets located on the first base line. Yeah, two tickets, whatever good seats you have, and if not, I'll take third base." They would never write, type, or text as they speak. They would think first and then type once or use visual cues such as a digital map of the stadium that included availability and pricing.

Also: These 3 AI tools made my two-minute how-to video way more fun and engaging

For a voice AI interface, the voice bot would need to figure out that the second mention of two tickets is redundant, and that the alternative location refers to the third base side of the stadium and what constitutes 'good'. The differences are even more dramatic between text and images. Humans process image data thousands of times faster than text data. The current systems for customer service, field service, and commerce -- along with many other disciplines -- will need generative AI solutions that understand human intent unless the request is simple and unambiguous. Generative AI in complex scenarios might complement standard workflow and predictive AI. I would take this further. The best uses of generative AI that will be truly differentiating, rather than the merely essential/commodity use cases, will be those that mix generative AI with predictive or mix both with more traditional workflow.

Generative AI is powerful but rules apply

Generative AI is an amazing addition to the tools for customer service, both for the human agents and technicians who will be enriched by the knowledge and advice it brings, and also for the customers who will be served quickly and accurately. Businesses focused on customer service and getting it right are those that remember the basics:

  • The customer should feel no stress.
  • They should leave the interaction satisfied.
  • The interaction reinforces that the business understands the customer.
  • The dialogue is consistent with previous and subsequent interactions -- not disconnected from the rest of the relationship.
  • There is a way to reach a human as necessary.
  • The customer service experience does not feel like a siloed department, but a reflection of the business overall.
  • There is a bit of novelty and 'wow' that impresses the customer with the thoughtful process design that makes the technology and engineering all but invisible.

Ed Thompson - "We should think about the timeline within which all this will play out. We are likely looking at a decade-long planning horizon. At the same time, organizations will need to be very flexible with their generative AI planning, as opportunities, priorities, and capabilities will shift as the technology and the organizational competency matures."

Yes, we'll want quick payback on initial pilots and early projects. We'd advise not to plan such that all at once you deploy 20 use cases of generative AI in customer service in the next two years. At Salesforce, we are working with 20 use cases across customer service, and dozens more in each of our other clouds. We can all expect that there will be 50 use cases in customer service in two years.

Getting started

The bottom line is generative AI and the road to autonomous service agents are programs for the long haul. You can expect wins all along the way, and the fun part is that you can bet your career on it.

Generative AI represents a fundamental change in how businesses think about software. We must keep in mind that generative AI goes beyond predictive AI. Rather than deliver an output based on strictly defined rules, generative AI creates new data and new content. Creating these results safely and reliably will depend on clean data and access to that data in near-real time. Integration and reliable APIs will be crucial, as will a dose of trust in data and data governance.

Also: 7 ways to make sure your data is ready for generative AI

In summary, here are a few pieces of advice when building a generative AI competency:

  1. Don't boil the ocean by chasing 100 use cases in parallel. Be laser-focused and try things, but focus on taking three proof of concepts into production with a proven financial payback and not just metrics around 'time saved'.
  2. Don't use generative AI as a hammer for every nail in the many situations where predictive AI will deliver the same result and is already out there and proven.
  3. Don't go recruiting skills. Whenever possible, upskill your people. It is a terrific way to recognize your talent. Look for people who are interested in learning. Make room for them to upskill. Focus on prompt design skills and prompt action/workflow skills for those outside IT and data skills for people inside IT.
  4. Focus on the flow of work. Generative AI is most impactful in the flow of a sales process, a customer support process, a marketing process, an ecommerce process, and a field service process.
  5. Beware of copilot proliferation. One copilot per application that needs to be maintained and supported -- imagine 1,000 apps with 1,000 copilots.
  6. Unless you're a mega bank or tech company, focus your model-tuning efforts on two or three industry-specific use cases and use out-of-the-box tech from your existing suppliers for the rest.
  7. Most importantly, focus on data quality, access, and governance. These are the fuel that AI needs, so it requires investment -- and likely more investment than the AI investment itself.
  8. Write to us. We love fielding questions and sharing what we know: mmaoz@salesforce.com for Michael and ed.thompson@salesforce.com for Ed 

This article was co-authored by Michael Maoz, senior vice president of innovation strategy, Salesforce, and Ed Thompson, senior vice president of market strategy, Salesforce. 

Editorial standards