Artificial Intelligence in Customer Service: Will It Live Up to the Hype?

Experts say 80% of companies will use artificial intelligence in their customer service operations by 2020. But are they right?

In this post – our first Peak Support Deep Dive – we will provide an overview of the current state of artificial intelligence in customer service and an assessment of how quickly—and how far—artificial intelligence will spread in this industry. In summary, we believe AI will play a growing role in customer service. Most large corporations will incorporate AI into their customer service operations by 2025, if not 2020.

However, in the near-term, the smartest companies will use artificial intelligence to supplement – not replace – human agents. The technology is not good enough to manage complex interactions with customers without human oversight.

The Rise of AI (in Customer Service and Everywhere Else)

Artificial intelligence has permeated society to the point where many people do not even recognize its uses in their daily lives. Most people have an artificial intelligence device right in their pockets—one that will answer basic questions for them and lead them to complicated destinations with ease.

And artificial intelligence is beginning to permeate the business world. Large corporations are rapidly adopting or at least experimenting with artificial intelligence in a variety of areas. Businesses spent as much as $40 billion developing new artificial intelligence capabilities in 2016, according to McKinsey.

Customer service is no exception. Some experts claim that 80% of companies will use some form of artificial intelligence in their call centers by 2020. According to a survey by Oracle, 37% of companies are already using AI to improve “customer experience.” Indeed, there is no doubt that artificial intelligence is playing a growing role in call centers and in customer support as a whole.

But how much can AI do today, and how quickly will its capabilities increase? It’s possible that 80% of large companies will use AI in their customer service operations by 2020. The question is, what will AI be able to do? Will it simply be a back-end tool that provides modest efficiency improvements – or will it already be advanced enough to replace significant numbers of human agents?

If I don’t run a call center, why should I care about this?

The answer has massive implications for nearly 5 million customer support workers around the globe (approximately 75 of whom work for Peak Support). The business process outsourcing (BPO) industry, as it is known, is large and diverse. It includes individuals providing customer support by phone, email, chat, or social media; phone sales representatives; and workers who complete myriad other back-office tasks.

The U.S., the Philippines, and India each employ more than 1 million people in the industry, according to the Economist, but sources vary widely, with some estimating the U.S. has more than 2.5 million contact center workers. In developing countries, a shrinking call center industry could reduce overall economic growth. In the U.S., states with the highest number of call center jobs – Texas, Florida, and Arizona – will feel the pain if technology makes some of those jobs obsolete.

What can artificial intelligence tools for customer service do today? 

Today, dozens of companies are competing to replace customer service agents with AI. These companies range from huge firms like IBM to emerging startups (see the Appendix at the end of this post for a comparison of some of the main AI tools on the market). Some offer tools for chat specifically, while others can handle queries via email, chat, or social media.

Chat appears to be the most common use case for AI in customer service – but do not confuse artificial intelligence with basic chatbots. Most chatbots are rule-based – you can program them to answer specific questions, but they don’t have the ability to learn as they interact with customers. If you want them to answer new questions, you have to program it in. AI chatbots, by contrast, get smarter over time just by handling more and more customer tickets.

Most AI tools for customer service have some combination of the following three capabilities:

  1. Answering basic queries: The AI software answers basic customer queries that do not require complex critical thinking skills
  2. Routing: If the software cannot answer the ticket, it routs queries to the appropriate agent.
  3. Suggesting answers for agents: It then offers the agent a pre-drafted reply (or a choice of several pre-drafted replies). Human agents can then send the reply as written or amend it as needed.

Today, AI bots still require significant human assistance

What percentage of tickets can AI software typically answer on its own, without human assistance? Depending on who you talk to, the answer ranges from 6% to 85%. And yes, that is a pretty large range.

A spokesperson for Zendesk told Peak Support that its AI-powered Answer Bot can, on average, answer 6% of queries on its own, without human intervention.

Another player, Agent AI, says its tools can answer 30% to 50% of “routine customer inquiries,” but does not specify what percentage of inquiries can be categorized as “routine.”

Helpscout interviewed experts in the field and says 10% to 35% of tickets can be resolved “without a human touch.” And software built by DigitalGenius, another AI startup, touches 50% of customer tickets, according to Juan Ageitos, a marketing manager. He did not specify how many of these tickets were answered entirely without human intervention. Digital Genius has raised over $25 million and has customers including KLM and Unilever.

Some companies say they can answer a much higher percentage of tickets on their own. Bob Morgen, director of sales at True AI, a London-based startup, says its technology, once fully trained, can answer 85% of tickets accurately without human intervention.

However, Morgen says True AI is designed to supplement – rather than replace – human agents. When a query comes in, True AI offers three suggested answers. A human agent can choose one and send it, choose one and edit it, or ignore the suggestions altogether.

“This is too important to leave out human agents,” Morgen said in an interview. “Your relationship with your customer can be the difference between keeping and losing that customer. We don’t envision a world where 100% automation is the way to go.”

Why are humans so important?

The reality is that AI tools still do not have the empathy or critical thinking skills of a human agent.

Let’s say you’re a customer of a telecom company, and you complain that your service is out. After a brief conversation, the issue is resolved, and your final message says: “Service is back, thanks. This has happened twice in the last month. Maybe I should explore other carriers.”

86% of customers expect chatbots to have the option to switch to a human agent; however, many companies don’t offer that option.

An AI tool might completely miss the overall tenor of the ticket, and respond with something tone-deaf like, “I’m so happy to hear your service is back. Thank you for your business!”

Given these challenges, it’s no surprise that 86% of customers expect chatbots to have the option to switch to a human agent; however, many companies don’t offer that option.

BTW: Emails are much more complicated to automate than chat

Emails are much more difficult to automate than chat, Morgen says. A chat is a conversation, and each interaction typically includes one question, with a single-part answer. An email, by contrast, tends to have a lot of content covering multiple topics. The customer may provide background information, then ask a question, then provide more background information, then ask another question.

Agents need to understand the whole email in order to craft a response. “It’s an order of magnitude more complex,” Morgen says.

So chat is the best medium for AI. But 92% of all customer service contacts still happen through email or phone (see IBM’s Webinar, “Tomorrow’s Technology, Today”). That means many companies must shift to chat before they can realistically implement AI.

… And AI isn’t always an improvement over “dumb” technology that already exists

Companies might not invest in AI if it can’t offer a significant improvement over existing technology. And many companies have already figured out ways – without AI – to automate customer service.

For example, some AI bots provide customers with relevant self-help articles. But most companies already provide self-help articles on their websites in an easily searchable form. If a customer wants a size chart, she’ll Google the company’s name and “size chart.”

If she’s sending an email to the customer service department, it’s probably because the size chart didn’t answer her question. And her question will be something like, “I’m a size 8.5 in Saucony and a 9 in Adidas. What size of your shoes should I buy?”

Delivering this customer a size chart may just enrage her, without answering her question.

Other AI bots suggest pre-drafted replies that human agents can customize. But customer service software platforms like Zendesk already enable users to create “macros” – pre-drafted replies that agents can use for routine inquiries. AI companies didn’t invent macros – they just make them better.

Granted, they may make them a lot better; DigitalGenius, for example, says one of its clients increased macro usage by 215%.

Even if AI bots could answer most tickets on their own, they would still face several significant challenges to large-scale adoption

Here are four more challenges that will slow the growth of AI.

Ticket Volume

AI tools require a significant volume of data in order to “get smart.” This is not an issue for large companies, but it means many chatbots are simply unusable for startups and even mid-sized companies.

According to, its tool needs to review about 5,000 queries before it can suggest relevant answers to agents. But that’s just the minimum, and it’s not enough for the bot to start answering queries on its own. Furthermore, at small companies, the answers to customers’ questions are often still evolving. So by the time a company reaches 5,000 tickets, the first 2,000 may be useless.

Accuracy is still lower than we would like … You can build a neural network with 10 tickets but it just won’t be robust. -Bob Morgen, True AI

Morgen says True AI requires about 50,000 tickets to reach a 20-40% level of accuracy, and “hundreds of thousands” to get to 85% accuracy. In the long term, the company hopes to lower that number.

“True AI’s long-term vision is to make automated customer service accessible to smaller & medium sized enterprises,” Morgen says. “We have been able to obtain accuracy of about 60% on a client dataset of only 5,000 conversations.”

He adds, “This accuracy is still lower than we would like … You can build a neural network with 10 tickets but it just won’t be robust.”

Reputational Risk

Allowing bots to operate without human oversight carries a huge reputational risk. Microsoft’s Tay is an example of an AI bot gone haywire. Microsoft launched this chatbot on Twitter in March of 2016 and it began tweeting offensive and racist statements within 16 hours of its launch. Microsoft immediately took Tay offline for “adjustments” and re-released it a few days later, but it had to be taken down within hours once again.

Many people regard Tay as the “worst of AI.” This may be an extreme example. But customers who get bad service from a bot will complain about it – loudly. Many companies will be very reluctant to take that risk.

IT System Complexity

Customer service agents often need to navigate multiple IT systems to solve customer problems, particularly at small companies. An e-commerce company, for example, might use Zendesk for its customer service ticketing system, Shopify for orders and inventory, and Loop for returns management. Its fulfillment center would operate on its own IT platform. And shipping companies like UPS and the U.S. Postal Service have their own platforms as well, of course.

Many queries involve accessing and navigating more than one of these systems, which often aren’t integrated. This is fairly easy for a human and very difficult for an AI bot. No AI bot we know of will call the U.S.P.S. and wait on hold for two hours to find out what happened to a package – and, yes, that is something customer service agents sometimes have to do.

Large corporations often have custom-built, fully-integrated IT systems, so it will be easier for them to deploy AI bots with advanced capabilities.

Return on Investment

There is no doubt that AI offers potential for cost savings. Zendesk’s Answer Bot can cost as little as $0.70 per interaction, and IBM says it can lower customer service costs by 40%.

But bots today are only answering the easiest queries – which are already the cheapest ones. Companies will probably need to train agents on those tickets anyway, in order to give them a base of knowledge that will enable them to solve more complex problems.

Furthermore, deploying an AI system requires a significant up-front investment of time and money. AI can’t just be cost-effective; it must save enough money that it rises above other potential investments the company could be making. Many companies will decide not to make the investment while the capabilities of AI are still fairly immature – and the reputational risks are high.

So when will artificial intelligence replace human agents?

Of course, AI capabilities will continue to improve, and the customer service bots of tomorrow will be much more powerful than the ones on the market today. How quickly will they get to the point where they can displace significant numbers of human agents?

Of course, we can’t say for sure, but Amazon’s Alexa is an interesting case in point. In just over one year, Alexa went from being able to interpret 1,000 voice commands to understanding over 10,000.

But that progress didn’t come cheap. Amazon has about 5,000 employees working solely on Alexa and related products. Plus, hundreds of independent developers are creating apps that expand Alexa’s capabilities.

IBM’s Watson is the only player in the customer service space that can match Amazon’s resources, which may indicate that it’s a player to watch in this race. Indeed, “Transform your call center” is the top “use case” presented on Watson’s website.

TL;DR: Sum this all up for me, please!

Artificial intelligence is perhaps the most-hyped technology of our time. It has the potential to transform everything from customer service to the treatment of cancer. But that transformation may not happen as quickly – or be as complete – as many people believe.

In the short term, only large companies are likely to adopt AI, and the smart ones will deploy it to supplement, rather than replace, human agents. For now, job growth in the BPO industry may slow as a result of AI. But large-scale AI-driven layoffs are unlikely.

Once the technology gets good enough to answer complex queries – not just by chat, but by email and phone – significant job losses could occur.

Customer service outsourcing companies need to prepare by incorporating AI and other advanced technologies into their offerings where appropriate. Furthermore, they need to invest in the skills of their team members.

Agents will still be needed to handle the most complex queries, and to perform higher-level tasks, particularly in sales, marketing, and operations. The firms – and agents – who invest in developing these capabilities will be most equipped to weather the AI storm.

Appendix: A Sample of Companies Offering AI for Customer Service

  • Funding: $2.7M
  • Customers: Intel,
  • Integrations: Apple, Android, ZenDesk, Sales Force Desk, Facebook Messenger, WeChat
  • Pricing: $1 per AI-assisted conversation for “standard” package, custom pricing for enterprises. also offers free CRM.

Answer Bot

  • Funding: Unknown; owned by Zendesk
  • Integrations: Zendesk
  • Pricing: Price ranges from $1 per resolution, for 50 resolutions per month, to $.70 per resolution, for up to 7,000


  • Funding: $26M
  • Customers: KLM, BMW, Unilever, Magoosh, TravelBird, HSCB, Panasonic, GoFundMe
  • Integrations: Salesforce, ZenDesk; other integrations available as well
  • Pricing: Custom


  • Customers: Beachbody, P90X, Tough Mudder
  • Integrations: Lowest package integrates with Shopify; second tier with Shopify and ZenDesk; highest tier will develop an integration for any CRM.
  • Pricing: Price tiers include $55/month for up to 300 tickets and $275/month for up to 15,000 tickets; custom pricing available for larger enterprises


  • Funding: Unknown; owned by Intercom
  • Customers: Intercom is a customer-messaging platform that many small and large companies use, but it’s not clear how many use Operator specifically. Companies such as Microsoft, Spotify, Stripe, and Shopify use Intercom
  • Integrations: Salesforce, Intercom
  • Pricing: Free with Intercom

True AI

  • Funding: $735K
  • Customers: Piloting with government and businesses of varying size
    Integrations: Salesforce, ZenDesk, Intercom; able to develop integration with other CRMs
  • Pricing: Unreported


  • Funding: Unknown; owned by IBM
  • Customers: Macy’s, H&R Block, Staples, 1-800-Flowers, Chevrolet, North Face, TD Ameritrade
  • Integrations: Watson Virtual Agent
  • Pricing: $265/mo.+ per subscription for up to 10 agents for “standard” package; custom pricing for “premium” package.

Writing and reporting support from Kaci Barker. Additional research support from Aidz Estrella.


Leave a Reply

%d bloggers like this: