What happens when customers meet chatbots? Discussing our latest research on AI in customer service

ChatGPT has ushered in a new era of AI applications, especially a new generation of chatbots for customer service. But while there have been plenty of discussions about how it can transform your business, we often end up overlooking the perspective of the end users. Ultimately, they are the ones who will be interacting with these chatbots. So, what are their thoughts on the matter?

Now, we’re no strangers to the world of bots – we’ve been debating and developing chatbots and versions of AI bots for years. So the notion of engaging with AI bots and deploying them in your business feels like second nature to us. That’s why, as we were building our newest AI chatbot, Fin, we felt the need to conduct comprehensive user research. We wanted to avoid bias or preconceived notions about bots, so we got out to the general public and asked them what they felt about them and whether their perceptions were changing with the latest developments in AI technology.

The truth is, traditional chatbots people have dealt with up to now are quite unpopular. Ranging from helpful to plain enraging, these bots have often been seen as mere obstacles on the path to reaching a human agent who can actually solve their issue. But could the tide be turning? Could interacting with a chatbot actually become a positive experience for customers? That’s precisely what we set out to explore.

In today’s episode, we’re joined by our own Senior Product Researchers, Cormac O’Dwyer and Sophie Woods, to find out if and how AI is changing people’s feelings toward chatbots.

Here are some of the key takeaways:

• Perception of traditional chatbots is mixed – their limited ability to understand and help customers leads to frustrating encounters that have soured people on the experience.
• Likely due to ChatGPT, customers perceive AI bots as superior to traditional ones, anticipating better understanding and communication – and an overall greater experience.
• Customers have expressed concerns about not having the option to escalate issues to a human since bots can’t understand and empathize with more complex or charged situations.
• The option to speak to a human builds trust between the business and the customer by acknowledging the bot’s limitations and recognizing the value of human touch when needed.
• It’s still unclear how patterns of interacting with chatbots will evolve over the next few years, not to mention how market dynamics will play on people’s expectations of AI bots.

If you enjoy our discussion, check out more episodes of our podcast. You can follow on Apple Podcasts, Spotify, YouTube or grab the RSS feed in your player of choice. What follows is a lightly edited transcript of the episode.

To bot or not

Liam Geraghty: What prompted you to conduct user research on customer attitudes toward AI chatbots in the first place?

Cormac O’Dwyer: I guess we spend a lot of time talking to Intercom’s customers, but from the point of view of chatbots, they’re not the main users. Customer service teams are the ones that buy them, set them up, and report on them, but the majority of the interactions are actually end users. Our customers’ customers.

We know that our customers care a lot about their customers. They buy a tool like Intercom because they want to provide a great end-user experience. So, the more you can understand that or close that gap, the better. Since ChatGPT happened, a lot is being talked about AI in general, but from what we’re seeing, it’s almost all from an industry or business perspective. And as our customers’ customers are the ones who are going to be interacting with this stuff, we really wanted to get their take too.

“If we don’t get out there and speak to end users about their perceptions of AI bots, we risk seeing it through this very unrealistic modern world that we live in”

The other angle on this is that the distinction between those two groups is not that clear. Everyone who works in customer service is also a customer service user. Part of evaluating Fin, for example, is just interacting with it as you would as an end user. There’s a lot of overlap there, and if customer service folks are out interacting with other chatbots, they’re forming their own impressions and expectations based on that. So, they’re going to bring that to how they think about Intercom too.

Sophie Woods: I think it’s really easy for us to get caught up in a small corner of the internet where we and our customers live. We’ve been talking about versions of AI bots and chatbots for years, and while ChatGPT has recently catapulted the technology, the notion of AI bots feels kind of normal to us and some of our more progressive, forward-thinking customers. If we don’t get out there and speak to end users (which is essentially the general public) about their perceptions of AI bots, we risk seeing it through this very unrealistic modern world that we live in, which, as researchers, is what we try to avoid.

Liam: How do customers perceive traditional chatbots, and what are the main reasons for their negative sentiment?

“One participant described it as getting past a gatekeeper, which really stuck with me”

Cormac: Well, the perception is not great, as you mentioned. Maybe it’s more accurate to say it’s very, very mixed. If a traditional chatbot works for you and gets you the help you need, that’s fine. If it works well, it can be even better than waiting for a human. It can be faster, simpler, and for some folks, it takes social anxiety out of the equation. The problem is that a lot of the time, they don’t work very well. There are a lot of terrible ones out there. Probably all of us have had a really frustrating interaction with a chatbot, and our participants have had similar experiences.

The way folks described traditional keyword-based bots was like trying to crack a code, almost. You’re trying to communicate with this thing, but it only understands things in a certain way, you don’t know the rules, and there’s no rule book that you can reference, so you have to just, through trial and error, hope for the best, and it always feels a bit like a shot in the dark. That can be exasperating. The other aspect is you might not necessarily associate a chatbot with helping you. Sometimes, there’s this sense that the chatbot’s there to help the business. On the back end, that would look like a triage bot to help you get routed to the right person – it helps the business, but it doesn’t necessarily help you as a customer. That’s not necessarily frustrating, but it’s not fun either. One participant described it as getting past a gatekeeper, which really stuck with me. So, it can be frustrating, and it can also just be something I have to get through but doesn’t necessarily add value to me.

“This is a two-way conversation, but all the work is with the end user. The bot just goes, ‘No, I didn’t get it.’”

Liam: The social anxiety point is really interesting. I’m the kind of person who’s still happy to chat with people on the phone, but so many people aren’t these days. And then, as you said, it’s almost like a puzzle. Anytime I’ve interacted with a chatbot, you’re just trying to come up with these prompts that hopefully will get you over the line.

Sophie: You’re trying to figure out their language and speak it, which is hard to do when you don’t know what their language is, you know?

Cormac: Absolutely. And it can also feel a bit like you are the one doing all the work. It should be a normal conversation. This is a two-way conversation, but all the work is with the end user. The bot just goes, “No, I didn’t get it.”

Liam: Computer says no.

Are customers ready for AI chatbots?

Liam: Could you tell me some of the insights your research presented about how customers feel about AI chatbots?

Sophie: In short, the one-sentence soundbite is that they’re seen as better than non-AI or more traditional chatbots. AI is helping to change perceptions of chatbots, and that seems to be happening quite quickly. That’s for a bunch of reasons. When asking customers, “What do you expect it would be like to interact with an AI chatbot or a GPT-powered chatbot?” there are things that make them think the experience would be more positive. One of them is that they expect that the AI bot would better understand them compared to traditional bots. And if the bot is able to understand you better, that has a few positive knock-on effects like, “Okay, well, I’m going to be able to communicate with them more efficiently and effectively.” Coming back to what Cormac was saying about having to figure out how to speak to the bot, that disappears a little when we think about AI bots. And if you’re able to communicate more effectively with the bot, you are less likely to have this deeply frustrating experience where you’re banging your head against a brick wall and not getting anywhere.

“AI has intelligence in the name, so there is the expectation it will be more capable than the previous version of a bot”

Another expectation we’ve seen is that customers expect an AI-powered bot to be fast. There’s a real sense of, “Okay, this is automation, so why would it be anything other than immediate? It doesn’t have to do human thinking to help me out.” AI or GPT-powered bots are also expected to have a higher likelihood of returning a helpful answer, and I think this is because GPT has proven to be pretty good, the people we’ve spoken to about this are familiar with the concept of ChatGPT, and they realize it’s a leap in how AI can help us. Their experiences with interacting with it have been good, and that translates to how they would expect a chatbot to be able to use that technology. AI has intelligence in the name, so there is the expectation it will be more capable than the previous version of a bot.

The way customers spoke about this is that it seems like they’re more willing to engage with an AI bot compared to a traditional bot because ultimately, they think they’ll have more success with it. So, I suppose what we’re hearing is that, on the whole, people aren’t super freaked out by this. It’s not like, “Whoa, keep the chatbots away from me. I don’t want it to read my mind and take over the world.” Instead, they’re actually leaning into it and feeling more optimistic about how it’s going to positively impact their customer service experience.

“Being able to convey emotions and frustrations is a really core part of that experience. Ultimately, people want to feel heard”

Liam: When you use a traditional chatbot, you’ve given it all your information, it can’t help you, and you’re passed on to a human that asks you all the same questions again, it’s like, “What is going on?” It’s like Groundhog Day. Were there any concerns raised by customers about AI chatbots? And on that, how do they feel about the availability of humans?

Sophie: The availability of humans is one of the biggest concerns that we heard coming from people dealing with AI bots. At the moment, as Cormac mentioned, end users are familiar with the limitations of traditional bots, meaning they’re aware that bots are usually the first step in some longer process that eventually puts them in touch with a human. The bot is there to triage, whether it’s successful or not, but ultimately, they get to access a human, and that’s the expectation. In contrast, they’re less familiar with the intended role of an AI-powered bot, and there’s concern that they would get stuck trying to resolve an issue with a bot that can’t understand or empathize with them, and there’s no other outlet.

“Okay, the AI bot speaks like a human, and it thinks like a human – has it replaced the option to actually speak to a human? That’s something that came up a lot”

When people speak with customer service, or when we speak to people about their customer service experiences, being able to convey emotions and frustrations is a really core part of that experience. Ultimately, people want to feel heard. And it’s even better if they feel that because of the frustration they were able to convey, the human on the other end gives them some flexibility with a policy. Like, “Because I can feel this has been really crap for you here, have 10% off.” We’ve all experienced that, and it’s like, “Okay, you’re listening to the fact that I suffered.”

In this new world, there’s a fear that AI won’t let us do that. What if it doesn’t let me speak to a human? Okay, the AI bot speaks like a human, and it thinks like a human – has it replaced the option to actually speak to a human? That’s something that came up a lot. Just knowing that there is the option to speak to a human and literally having the button, “Forward me to a human,” puts people at ease. That’s a really powerful thing you can do to help people trust an interaction with an AI bot.

From sci-fi nightmares to real-life assistance

Liam: Our exposure to AI has been in pop culture, movies, and TV, you know what I mean? And AI has never been a good thing in movies. I’m thinking specifically of 2001: A Space Odyssey, when the astronaut is stuck outside the ship asking the AI bot to open the pod bay doors, and it’s just like, “I’m afraid I can’t do that.” We’re like, “Oh yeah, that’s really poor customer service there.”

“There’s nothing more frustrating than getting stuck in that doom loop where you can’t progress”

You mentioned trust. Could you explain the role of trust and confidence in customer interactions with AI chatbots and how the option to escalate to a human agent can affect customer trust?

Sophie: I think there are a few things going on. A lot of it relates to customers’ past experiences talking to bots and how there’s nothing more frustrating than getting stuck in that doom loop where you can’t progress. Maybe there is a future where AI bots have become so good that end users have forgotten about that world where bots were rubbish and don’t feel they need the reassurance of a human waiting in the wings at all times to sweep in.

But for now, that isn’t the case. The mental models are that bots can be really difficult. The option to speak to a human builds trust in two ways – not only in the bot itself, but it communicates that the bot knows its own limits, and the bot knows that at some point, it may not be able to help you anymore.

“Yes, they’ve seen the movies with the terrible dystopian AI presence, but in their day-to-day lives, these are helpful tools they’re using that are helping them”

It also helps a customer understand that the company they’re dealing with is trustworthy because the company is saying, “We understand that some issues just need a human touch, and a bot can’t do that,” whether that is because your issue is just too complex and nuanced for a bot to help you with or because it requires empathy and understanding. I think there are two angles of trust there. One is, “I need to trust that I can speak to a human because the bot might not be able to solve my problem,” and the other is, “I need to trust that the company understands this about my experience.”

Cormac: Sophie, you mentioned this earlier, but it struck me that a bunch of the people that we spoke to were already using things like ChatGPT in their daily lives. So yes, they’ve seen the movies with the terrible dystopian AI presence in them, but in their day-to-day lives, these are helpful tools they’re using that are helping them. They’re not locking them out of spaceships – they’re helping them rewrite an email, answer questions at work, or whatever it is. So I guess that’s helping build some trust as well.

Liam: 100%. Sophie, I’ve never heard the phrase “doom loop” before, but I’m going to remember that.

Evolving mental models

Liam: How do you see customer expectations evolving as they become more accustomed to interacting with these AI chatbots? Are there any future research areas you’re planning to explore around this field?

Cormac: Yeah, there are definitely a bunch of questions that are still on our minds. One thing I’m finding really interesting is that the people that we’re talking to are not AI experts. They’re just ordinary folks. And it’s interesting to see how their mental models of AI are updating unevenly, in a way. Yes, there’s this expectation that an AI-based chatbot is more intelligent, helpful, and probably more capable to understand my language better, but in practice, we saw a minority of folks who were still stuck in a keyword-based interaction pattern. A small number of folks were inputting one or two keywords – “billing,” or “account update” – and hoping for the best. It’s like they’re porting that over from their historical interaction pattern. I’m curious to see how quickly that will change when we get to the nuts and bolts.

“We’re undoubtedly going to see a lot of AI bots introduced to the market in the very near future, and we can be quite sure that the quality is going to be hit-and-miss”

Will there be folks that continue with the older way of interacting, or, as this becomes a more common interaction pattern across contexts, are people going to learn this as a skill? How quickly will people develop a sense of the limits of AI? I guess we’ll see how that goes.

Sophie: Another thing at play that we’re really interested to see in terms of future research and how it evolves is what’s going on in the market. We’re really curious about how that will impact the general public’s expectations. We’re undoubtedly going to see a lot of AI bots introduced to the market in the very near future, and we can be quite sure that the quality is going to be hit-and-miss. Some will be super impressive, and some will be super frustrating.

Will people have experiences with poor-quality AI bots, and will that damage their expectations going forward, or will we see that their expectations become really high because people end up having these really great interactions with great AI bots, and that continues to increase the number of use cases of AI bots? We don’t know where the market is going to go, and perhaps it depends on the volume of good versus bad AI bots that are released into the market. Time will tell.

Liam: Well, Sophie, Cormac, thanks so much for joining me today.

Cormac: Thanks, Liam.

Sophie: It was a pleasure. Thanks for having us.