customer support AI

Will Customer Support AI Replace Human Agents?

Everyone’s talking about the AI revolution and wondering how AI will change the world, but will AI replace human knowledge work entirely? Will we even need human customer support agents in the future or will they be replaced by customer support AI?

The answer to this is not very simple. Although customers don’t prefer chatbots in customer service, companies are increasingly relying on them to reduce caseloads. So it only stands to reason that companies are interested in next-generation chatbots powered by technologies like GPT. Goldman Sachs predicts that 300 million jobs will be impacted by the advent of AI.

Even the experts don’t agree on a definitive answer. 

We recently asked Jon Stokes, co-founder of Ars Technica and an expert on AI, if it will replace customer service jobs. “The short answer is yes and the longer answer is yes. It’s a trivial yes. It’s a really easy yes,” Stokes says. 

Colin Crowley, a customer experience advisor at Freshworks, recently appeared on the MYC’D UP with Tech Leaders podcast, and he gave a very different response to a similar question:

“No, I don’t think that’s ever going to take place because there are just too many unique characteristics that human beings bring to every conversation, especially when it comes to emotional intelligence and empathy and so forth. And we’re in a world where more and more customers have higher expectations when it comes to quality. So the need for that is increasing, not decreasing.”

Realistically, the need for customer support agents isn’t likely to fully disappear in the very near future due to the constraints of current technology and systems. 

Problems with current customer support AI

HappyFox, which sells chatbot software, openly admits that chatbots aren’t a one-size-fits-all solution and offers seven reasons why chatbots fail:

  1. They don’t identify the customer’s use case. In plain English, the chatbot doesn’t understand what your customer needs.
  2. Bots don’t understand customer emotions and intent. Empathy is key in customer service; machines simply don’t have it.
  3. You’re not transparent about using a chatbot. Fortunately, this is easy to fix: make it clear to your customers that they’re talking to a bot, and make it easy to talk to a human for more complex needs.
  4. 54% of customers in the United States prefer talking to humans.
  5. Chatbots can’t address personalized customer issues. Fixing this requires more investment on your end: “This would require regular upkeep and human intervention to identify and understand what the users need, leading to happier and satisfied customers.”
  6. They lack data collection and analysis. This requires regular check-ins to gauge performance.
  7. You’ve deployed the chatbot in a way that doesn’t fit your brand. This takes work on the part of your team to choose a chatbot and craft scripts that align with your company’s brand and values.

As you can see, chatbots aren’t simply a “set it and forget it” solution. While they can reduce pressure on your customer support reps, they take work and maintenance to be effective, and even then, the majority of your customers may find them unsatisfying.

Crowley relayed a chatbot issue he had at a previous job:

“We noticed that the Chatbot wasn’t returning the appropriate articles, even though we had all these great articles. And it ended up just being a phrasing issue because customers were referring to these deliveries as boxes. So they were asking: “Well, where’s my box?”, “where’s my order?”, “where’s my delivery?”. So we were using internal lingo, and that small thing prevented the chatbot from really understanding the customer intent.”

With the current crop of chatbots, it takes careful and intense human training to ensure that chatbots work as expected, and even then, Crowley emphasizes only using them for simple tasks.

Advanced large language models like GPT may improve chatbot technology, but drawbacks exist.

GPT is unpredictable

There are very few chatbot companies talking about GPT. One of the few exceptions is Ada, which has been working with OpenAI for the past year to integrate GPT technology. The company is exceptionally open about the challenges in implementing generative AI for customer service.

Businesses and customers value consistency, which GPT doesn’t excel at. “LLMs have no understanding of factuality, and therefore can generate inconsistent answers to the same question,” Ada CEO Mike Murchison writes.

That’s a very nice way of saying that GPT has a tendency to make things up. The technical term for this is hallucinations. AIs like GPT will confidently invent facts to fill gaps in their knowledge because they have no way of knowing what is true and false.

What makes hallucinations particularly egregious is that they can often sound like something that could be true or express misleading half-truths.

“Oftentimes, generations seem correct, which can make inaccuracies particularly nefarious. While this might be useful in a creative context for consumers, in a business context, if your store hours or insurance policy is misreported, you’re in trouble,” Murchison writes.

Even when GPT provides accurate responses, the output could differ even when the same prompt is used. When testing TextExpander for ChatGPT prompt engineering, we often received different answers to the same question.

Speaking of prompt engineering, another issue with generative AIs is knowing how to phrase your query to get the expected results. Prompt engineering is a skill and art that you can’t expect your customers to be well-versed in.

“The problem with AI in terms of support is the customer has to ask the right questions to get the right answer,” says TextExpander support team lead Vince Crighton.

As anyone who has worked in customer support knows, customers often struggle to phrase their issues, and it takes skill and intuition on behalf of the rep to extract the necessary information from the customer. “The problem is that those AI chatbots don’t make those interjections. They just have that singular answer,” Crighton says.

To ensure consistency and accurate information, you need a “human in the loop,” especially for complex queries. A Snippet tool like TextExpander can make your human customer service agents drastically more efficient and ensure consistent responses to customer questions.

See how TextExpander works

Thank you for your email! I’m happy to look into this for you.

I’m happy to walk you through our active-user-based billing system. TextExpander works on pre-purchased credits, where 1 user needs 1 credit a month, or 12 credits per year. We only deduct credits for users that expand a Snippet within a billing period.

There are two payment plans available:

  • Annual Payment: You purchase 12 months’ worth of credits per user on your billing date at a 20% discount.
  • Monthly Payment: You pay for one credit per user at the beginning of each month

Once you use up your credits, you are charged again at the start of the next billing period, either for another year on the annual plan or for the next month on the monthly plan. There are also additional billing resources at this link.

Does this clarify?

Thanks for reaching out to TextExpander. I’ve escalated your request to our tier 2 support team, who should get back to you shortly with a solution.

We value the positive feedback you provided about your customer experience! If you have a moment, we’d greatly appreciate it if you could take a moment to share your experience on social media. Our handles are below for your convenience.

Have a wonderful day!

Thanks so much for reaching out and I’m sorry to hear that you’re having trouble. I’ve included a link here that should explain the proper way to troubleshoot this issue. Can you please follow the steps and let us know if the problem has been resolved?

What is good customer service worth?

Generative AI like ChatGPT is the biggest tech revolution since the iPhone, and Stokes likens its potential impact to the printing press. AI’s potential to change the world can’t be understated, but at the same time, we have to be realistic about its limitations and its potential costs.

Oftentimes, customer support is the only in-person interaction customers have with your company, and that one interaction can make or break that relationship. And with inflation driving up costs, customers have higher expectations than ever. Freshworks surveyed 2,000 Americans and found that 58% expect better service now that everything is more expensive.

Freshworks also found that customers expect their problems to be resolved quickly—in five minutes or less and that increases with age: 39% of Gen Z, 49% of Millenials, and 66% of Baby Boomers.

Not only do customers expect more, but they’re also less forgiving. Verint surveyed 2,000 shoppers and found that 88% are likely to make a repeat purchase if they experience good customer service, but 66% are less likely to return if their customer service issue isn’t resolved on the first attempt.

That doesn’t just apply to individuals. Fusion Connect and Gartner surveyed 300 executives and directors in information security and information technology and found that 91% of them are willing to pay more for better customer service.


  • 52% would pay 6-15% more for better customer service.
  • 24% are willing to pay whatever it takes to get top-notch service.
  • 21% said they would cut ties with a supplier after one bad interaction.
  • 15% said they would reduce spending with a supply after one bad interaction.

The stakes with customer service have never been higher. While customer service is often considered a cost center, it may be the most important department in your organization, which is why customer service training and preventing customer service burnout should be emphasized.

Chatbots can save your customers and team time with simple requests, but when double-digit revenue is on the line, you want to arm your team with the tools they need to be in control. With TextExpander, your top talent is always in control. Try it for yourself.