Intercom on Product: The AI revolution and the future of customer service
Recent advances in GPT technology are rapidly transforming the world around us. What does it all mean for customer service teams?
When a promising new tech hits the market, it’s always interesting to look back at prior generations of similar tech to understand how it may play out. Phones evolved from landline telephones with rotary dials to mobile phones and smartphones, following a somewhat predictable trajectory based on technological advances. They became smaller, more affordable, and widespread, with better battery life, better cameras, and more advanced capabilities.
But ChatGPT is another thing entirely. Since its launch in November of last year, it has swiftly begun reshaping entire industries, including, undoubtedly, customer service. There are now bots, like our own GPT-powered Fin, capable of quickly solving a big chunk of customer issues and with minimal setup. And that has groundbreaking implications for customer service. What does this mean for CS agents? Are people going to lose their jobs? And how can we take advantage of this technology?
In today’s episode of Intercom on Product, I sat down with Paul Adams, our Chief Product Officer, to talk about the transformative impact of AI, evolving landscape of customer service, and how humans and AI will work together in the future.
Short on time? Here are a few key takeaways:
- Rather than significantly shrinking the workforce, AI will drive changes in the way customer service teams organize – from their size and shape to their priorities and areas of focus.
- Much like it happened in the past with previous technology, the use of AI can raise the bar for customer expectations of quality.
- With the increasing affordability of AI, brands can differentiate themselves by delivering world-class experiences. Then, customer service will truly become a profit center.
- AI may excel at many customer service tasks, but humans will still be the best for solving complex, urgent, and emotionally charged conversations.
- Support leaders in the near future will focus more on knowledge and bot management – from their tone to the rules to the content they feed off.
- As traditional metrics like response time or handling time become irrelevant, there will be a need to develop new metrics to measure the impact of AI on the customer experience
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.
Redefining customer service
Paul Adams: Hey, everybody. Welcome to Intercom on Product. As always, joined by Des Traynor.
Des Traynor: Hey, Paul.
Paul: We’re going to talk about the future of customer service and specifically about how automation and humans will work together in the future, how AI is changing things, and even things like how we believe the majority of questions in the future may indeed be answered by AI and not by people. Let’s start, Des. Could you kind of outline where we’re at today?
Des: On November 30th, ChatGPT launched. A few weeks back, GPT-4 launched. There’s been like a step change in AI, full stop. How has that applied to the customer service industry? There are now bots – our own Fin is one example – that are capable of resolving a huge percentage of customer issues instantly with zero training. And that has groundbreaking implications for what customer service is about.
The zero training piece means you just point it where it needs to read. The instancy means it’s much quicker than a human – it doesn’t need to type; it’s a bot. And the percentage can vary. It depends on how complex the business is, et cetera. But we’ve seen anything from 10 to 60 or 70% of total resolutions, and that changes the workload. It also changes the areas of investment and how you should think about what the field of customer service is about. That’s where we are. We can try and pretend that stuff isn’t true, but that’s the reality of the CS industry today.
“Does tech tend to stop at arbitrary limits? Will it stop at 10% end resolutions, or will it stop at the door of product management and not go inside? No, tech never stops”
Paul: You mentioned the ChatGPT launch back last autumn, and I think that, to me, personally, was the big one. Before that, of course, we’ve been building and designing bots for many years. 2016, 2017, I think, was our first foray into launching bots, and we’ve seen them used for all sorts of different types of things, in sales as well as customer service. But until now, there’s just a lot of friction to get them started. You have to set them up yourself, all sorts of different things. And it’s interesting for us to reflect on even just how we run the company, how we build products, how we work on our support team – all sorts of things have changed. It really feels like a transformation. Do you think there’s some inevitability here? Is AI taking over?
Des: Yeah. I mean, I do. It’s not always comfortable to talk about that because it’s easy to talk about the industries that aren’t yours, but I believe AI could come for the product management job. I believe it can come for the engineering job. Generally, if you ask yourself, “Does tech tend to stop at arbitrary limits? Will it stop at 10% end resolutions, or will it stop at the door of product management and not go inside?” No, tech never stops. Will it get faster or slower? Faster. Will it get more accessible or less accessible? More accessible. It will get cheaper. There’s a pretty good precedent for what’s going to happen here. I do think there’s an inevitability about how it’ll upend every industry, including customer service. Obviously, I don’t believe it’ll displace everything and every human. We firmly believe in the idea that it’ll be automation plus humans. But I think it will be difficult to overstate how significant this change is.
Paul: I think so, too. It reminds me of other types of automation in the past – cruise control, setting the speed limit, and taking your foot off the accelerator. And now, with assisted steering, there are self-driving cars. That has felt like a pretty gradual evolution. You can track it as it’s going along. Each new car seems to be better than the last one, in kind of small but also significant ways. This is different for me. This is a rapid transformation. And I worry a lot of people will assume it will be slower than it’s going to be. But they do have genuine reasons to, for want of a better phrase, fear the technology and hope progress will be slower.
Navigating org changes
Paul: From talking to our customers, what have we heard as the biggest fears people have?
Des: Yeah, I think your point about speed is quite relevant. We’ve always seen those charts of how YouTube got more popular than home television, and I’m always like, “Yeah, because clicking a link is a lot easier than going to a store and buying a television. It’s a lot cheaper, too.” And I think there’s a version of that. People replace their car probably every seven years or something like that. So, even if a groundbreaking piece of car tech is released today, it won’t be commonplace for at least a decade. And then specifically, cars really only affect the motorways and the roads. I think the difference with AI is it’s going to be distributed very quickly across the entire internet, and it’ll affect every single industry. It won’t wait outside to be invited. There’ll be a tectonic shock to the world when this is finished.
“Your org chart changes – maybe the size of your team, and certainly the shape of your team, where your team focuses its efforts, and how you measure and manage them”
What are the actual genuine scary pieces? I think that the biggest one people come back to is, “Well, what about the humans or the jobs? What about employment in the sector?” We’ve seen figures estimating that AI could displace 300 million jobs within a year, and I think they’re valid concerns. It sounds like a very big number, but, as I said, we’re at a time when people can access this tech immediately. I think ChatGPT has a hundred million users already. That does not mean a hundred million people have somehow automated away a load of jobs – it means this technology is augmenting the lives of a hundred million people already.
I think the nature of employment within the CS world changes. The org chart changes, and the areas of investment change. In a post-Fin world, things like knowledge management are really important relative to frontline support. The types of issues they deal with will be more complicated because the simple stuff gets resolved now. So, I think your org chart changes – maybe the size of your team, and certainly the shape of your team, where your team focuses its efforts, and how you measure and manage them. All of that will change. Just to be clear, fear is often warranted because it’s messy to go through all that change. It’s hard. Think about any single simple reorg. You can have a bit of anxiety around we land this. Now think about the reorg you have to do for something like this. It’s gargantuan in comparison.
The other area where I think fear/skepticism is justified is that the customer experience could get worse. I can imagine circumstances where you’re dealing with a pretty high-taxing, high-emotion situation, and you get a chirpy Botty the Bot bot jumping onto the screen to try and help you, like, “Hey, you want to buy a coffin?” It doesn’t work. On the flip side, I could say, “Hey, what if you really need something done immediately, and this thing replies in 0.2 seconds, and you would’ve been waiting 45 seconds?” That’s a far better experience. I think there’s a lot of justifiable anxiety, or what you used to refer to as anxiet-ment, to say it can be exciting, but it can also be pretty scary. And that’s totally natural. There will be versions of this for every industry. CS is just the one we’ve chosen to deploy software.
Paul: The quality thing is a really interesting one for me. I’ve been thinking of this a lot. I’m a big believer in looking back at prior generations of similar tech to understand how this one might play out. And again, with automation, if you go back to the 1950s, for example, when a lot of white goods started to become commonplace, mass-manufactured, and suddenly everyone had a vacuum cleaner, a fridge, all these things. There were a lot of people employed to do that work – society looked quite different in terms of the makeup of the workforce. There’s a fear that this would put people out of employment and that didn’t quite happen. It did happen to some degree, but people also redistributed themselves into different areas, as you were saying.
Another thing that happened was that the expectations of quality went up. Cleanliness expectations went up. If you have a vacuum cleaner and have all these mods, there’s no excuse for your house not to be immaculately clean. And so, these expectations of quality went up, and a lot of the work required didn’t go down. I think we might see the same in customer service. Most customer service teams that we speak to are under pressure. There’s a lot of volume. The internet creates this increasing volume of support questions, especially when you’re using chat and messenger – it’s super easy to get talking. Volumes are up, teams are under pressure, and what this might do, rather than take huge swaths of jobs away – I think teams will get smaller –, is actually increase the quality of the experience and the expectations that go along with that.
“If you’d taken a photo of an office floor in the 1960s, you’d see 300 people with pens and paper and all that. Today, that’s probably one spreadsheet. And yet, those floors are still full”
Des: Absolutely. And what that would look like is it would be unacceptable for brands to do anything other than jump on a video call and do a bit of screen sharing or co-browsing to try and help people out there. Because if you’ve got through to a human, that’s the bare minimum you expect of them. The assumption is they can do that because all the low-hanging fruit support queries have been resolved automatically. And a lot of that comes back to the lump of labor fallacy. This idea that, yes, that job doesn’t exist anymore, but now the job downstream of it is expected. So what used to be white-glove service is now the common state.
There are other things that I think would follow too. If you’d taken a photo of an office floor in the 1960s, you’d see 300 people with pens and paper and all that. Today, that’s probably one spreadsheet. Technology changes things. And yet, those office floors are still full. It’s just that everyone’s doing a higher order of work. I think we’ll see versions of that.
The other side is that Japanese paradox – as things get more affordable, you tend to do more of them. It’s the same basic idea, but if you can now deliver world-class customer service, that becomes a way to genuinely distinguish your brand. And you can do that because the cost of it has come down because you can automate so much of the undifferentiated heavy lifting of it, so a lot more brands will go to battle on that and say, “Hey, we’re going to win, we’re always going to convert a customer because we’re going to lay out the right carpet and the white glove and every bit of experience you could want.”
And in that world, you start investing more in CS because it actually starts to pay back better. And you might finally see this realization of CS as a profit center, not a cost center. We don’t know. We have to be honest about that. We don’t know. The least likely thing, because it just never happens this way, is that this technology lands, the general average is like 20% or 50% inbound displacement, CS jobs worldwide decrease by that exact percentage, and we all just carry on about our business.
The human touch
Paul: You talked about the different types of ways in which people might be redeployed in teams, and their roles might change. We’ve obviously been speaking to lots of our customers about how they think about AI and products like our new Fin bot. We’ve lots of customers on the beta trying it out. We’re both learning a lot about what does and doesn’t work. But as you said, we’ve had some amazing results so far.
“And then, as I said, high-urgency, high-drama, high-emotion – those are the types of things where humans specialize.”
Almost always, there’s still a role for the human agents to play too. I was fascinated when some of our customers said, “We know that Fin, or an AI bot, can deliver a response faster, and arguably more accurately than a person can. But we actually want people. We will differentiate our customer service on people.” And maybe that’s just for VIPs. You mentioned white glove – maybe there’s another level of white glove that goes above and beyond and it becomes the new norm.
Des: Proper concierge might be a thing, and onboarding effectively an in-browser customer success person dedicated entirely to you, perpetually available. But there’s also a class of conversation that is high-urgency, high-emotion, high-bandwidth, very messy and requires a lot of disambiguation. The unknown unknowns.
For example, there’s no way Fin, or any bot, could resolve a query that says, “I don’t know if this thing is doing the thing it’s supposed to do.” You’re like, “What thing?” And I’m sure some extremely talented product person is going, “Well, you could infer from the URL, blah, blah, blah.” What I mean is the customer doesn’t really know what they want. They don’t know if they’re in the right part of the product, and the best thing to do there is to actually say, “Hey, talk to me about your business. What are you trying to achieve?” Could a bot do all that? Yes. Is it as good an experience, or as smooth as a human? No. And then, as I said, high urgency, high drama, high emotion. Those are the types of things where humans specialize. I think they’re going to persist.
There’s a lot of that classic ticketing, as in, “Hey, this is actually a longitudinal issue. We’re going to need multiple people.” They need human approval, verification, and authentication, or they need to change something about how it works, and then they need to get back to the customer. There’s going to be more of a dynamic. Maybe there are more people on the customer side. I never say never in AI anymore because I’ve been caught too many times, but I’d say we’re probably grand for the ticketing use case for a few more years. I think there’ll be a lot of stuff we’re used to in the CS world that will still exist long after the bots have had their way.
Paul: Yeah. The ticketing one is fascinating to me because oftentimes, with technologies, you could argue that lots of ticketing means you sometimes need people in other departments, so the ticket must get passed along. Like, “Hey, I’ve done my bit. I passed to you.” Des, you do your bit, pass it over to someone else, it comes back to me, and then I could get back to the customer. You could argue that a bot can do that, a bot can do the pass-along thing. But life’s messier than that.
“There’s a spectrum of the ROI of automation, in a sense. The high end for me, in the support world, is when they remove the need for someone to do something”
You look at the workplace and think that Slack, email, and these things all work fine, and video calls worked fine during covid when we couldn’t meet in person. But then, of course, people came back to the office and did the type of communication that was only possible when you bump into someone in the hallway or grab someone to a meeting room or a whiteboard, and that’s kind of the fastest, more efficient, and sometimes the only way to get that type of work done. I think this messiness that is human nature will also persist.
Des: I also think there’s a spectrum of the ROI of automation, in a sense. The high end for me, in the support world, where automation or AI are at their most valuable is when they remove the need for someone to do something. That’s, say, end-to-end resolution on an inbound query. A step before that would be, say, what we released in January, the inbox ChatGPT features that massively speed up something a human has to do, but the human still has to do it. A couple steps down from that might be things like you’re alluding to, where the bot can infer by a rule when to reassign a ticket. I wouldn’t call that transformative. It’s probably just a good best practice for a platform to know how to reassign something.
As long as there’s a unit of work, we can either remove the unit of work, which is what Fin does, or optimize the variables around it. But when we’re optimizing variables around it, such as the average handling time or the time it takes for Des to pass it to Paul, they’re all second-order things. They’re not as big as getting rid of the thing entirely. So I think that stuff, whilst it’s going to be hugely beneficial, it’s not transformative in the same way.
Leading the charge
Paul: We’ve talked about the facts, and we’ve talked about the inevitability we feel. Let’s wrap up by talking about the coexistence of bots, AI, and people. We’ve talked about ticketing and the fact that it’s messy and so on, but we do see other types of coexistence.
Des: Yeah. I think that the new role of support leaders of the future, and by the future, I mean-
Paul: Next year.
Des: … mid-2023, will be managing the intelligence, managing the bots, and managing the knowledge management – where do the bots actually get their information from, in a sense, and making sure that information’s up-to-date. How do we measure the performance of a bot? If we want to run two bots off against each other, how do we do that? How do we measure the customer experience when the majority of it is automated? What is a productive human when you’re not measuring them on tickets per day, but rather on the surprise and delight they’ve delivered to the VIPs they’re managing?
“There’s an entire customer journey and experience to be designed in a world that is assumed to be automation-first”
I think there’ll be a lot of work around managing the bots. Everything from the tone and manners to the rules, when they fire, when they shut up. It’s important stuff. The knowledge they consume, how we manage that, and keep that really crisp is hugely important. Should the bots read your API docs so they can answer API questions? Will they answer it accurately? How do we know? That’s a whole new type of work. Today there are tools for doing support QA where you can review 20 conversations by Paul and see if you get, on average, five stars, and, when you get it wrong, if it is usually in a certain area. There’ll be a version of that for bots, too. That’s like administration, management of the knowledge.
Then, the actual delivering proper automation-first customer service, which includes the proactive nature of bots. When should bots choose to step up, interrupt a person, and be like, “Hey, you’re doing this wrong.” There’s an entire customer journey and experience to be designed in a world that is assumed to be automation-first, where humans are only dealing with complex issues, and the bots can, for the most part, resolve anything that’s a common issue and proactively address things.
That’s not going to be easy. There will be new jobs created, and a lot of people who are best equipped to take those jobs will be people who do jobs today. We firmly believe that humans plus bots is the way. I think there’ll be as much change for humans with new roles, measures, reports, and roles managing automation. In software, every company would have its own server room with servers in it. And now we entrust all that to the Amazons or the Azures or GCPs. But we still have people who own that, and they make sure we have a good setup. They’re secure, reliable, have fast performance, blah, blah, blah. All of that’s going to have to happen in this world too.
Paul: Just double down on the content. When I was at Google, we were designing things like Wikis and these knowledge-based type products. And the problem was always managing content. You mentioned that keeping content up to date is going to be critically important. Fin is amazing at taking in content in seconds and answering questions on the content. It’s really, really good at it. However, lots of the contents are out of date. We’ve got to update our knowledge base and keep our articles up to date. Content’s going to be a massive area of investment. And as you said, redeploying people from other types of roles could be pretty common.
“First response time will be zero seconds. That’s not an interesting metric anymore”
How about reporting? That’s a whole other area with very standard metrics where people tend to report upwards, and the performance of support leaders is sometimes measured in a pretty cold, calculated way. How do you think that might change with bots?
Des: I think all the metrics have to change. First response time will be zero seconds. That’s not an interesting metric anymore. And likewise, the way in which a human will be evaluated will not be like, “Did you do your 60 cases a day?” It will shift to being, “Did you dramatically improve the quality of life for one of the customers you engaged with? Are they more likely to be loyal, stick around, engage more, and use the product more?” A lot of the standard bot metrics will have to be invented from here, honestly, because the response time and handling time aren’t things anymore. The thing might be accuracy or specificity.
The other thing that we notice even with Fin is that once people realize they’re talking to a bot, they tend to not treat it like a human for obvious reasons. They don’t necessarily always say, “Yes, that answered my question. Thank you very much. I have no further follow-up questions.” So, you might need to do a lot of inferred metrics. If somebody asked, “How do I add a project?” your definition of success of the answer might be that the number of projects that person had increased as opposed to being like, “Well, they didn’t ask any follow-up questions.” You might have to infer how good your answers are because you can’t necessarily rely on humans to be like, “Good job, Mr. Bot.” They’re probably not going to bother.
There will be a lot of stuff where I think we will need to do deeper dives to work out if this is actually a strong customer experience. A lot of that might be just spot-checking in the early days. But I think that, at scale, we’ll need to look into aggregating that into meaningful metrics. Did the customers end up doing the thing they’re trying to do? If so, we can determine that the bot helped, hindered, or did nothing.
Paul: Yeah. It’s human nature to be change-averse. It’s scary. Any advice to people who run support teams today who are in that kind of mode?
Des: I appeal to a lot of resilience and open-mindedness. Your world is going to change. I’d struggle to come up with any answer that says, “No, your world’s not going to change.” It is. So what I would say is learn what’s possible. Look at the tooling. Play with ChatGPT and ask it some basic questions. If you work in a hotel or an airline or whatever, play with ChatGPT, ask it the questions you get asked, and see how decent its answers are. Do that in order to allow yourself to accept changes on the way. Investigate the tools out there.
“I can try and pretend this isn’t happening, or I can face reality and say, ‘This is happening – let’s use it to our best advantage’”
Think about what you would like to happen to your business if someone told you, “Hey, we can augment your team by reducing the most common difficult issues by removing many undifferentiated steps.” Is that possible? How would you make that possible? You get on the front foot. And no matter the bot, it can only work off the information it has. If you’re the sort of org that relies on a lot of tacit knowledge or people learn through osmosis by sitting beside other people and hearing them on the phone or whatever, the bot will struggle in that environment.
I’m sure somebody is going to say to me something like, “Well, OpenAI has this whisper technology. They can do voice text.” I get all that. But I would say the better documented everything about your business is, the better automation and AI can help. And the better you can get ahead of this and see the areas where you really want to extract the value, the better position you’ll be in. Scott Galloway said, “AI probably won’t take a job, but somebody who’s really good at using the AI probably will.” I think that’s the right type of paranoia to have. I can try and pretend this isn’t happening, or I can face reality and say, “This is happening – let’s use it to our best advantage.” I think that’s the best perspective you can have.
Paul: Good advice there, Des. Thank you for that practical stuff. To everyone else, thank you for listening, and we’ll see you next time.
