The arrival of ChatGPT just eight weeks ago was a watershed moment.
Suddenly it seemed that generative AI might transform industries from education to marketing. And Intercom’s area of focus – customer service – is among those most poised to benefit.
However, getting from technological breakthrough to useful applications is far from straightforward. The trickiest thing about large language models (LLMs) is that they’re great at appearing plausible, even when they’re wrong. ChatGPT suffers from “hallucinations” – confidently providing incorrect information. This means we need to think carefully about how to apply the tech in practice to make it truly useful. It has also made it hard to tell what potential applications will ultimately be transformative, or merely end up being toys.
We always believe the best way to cut through the hype is by putting real features in the hands of our customers.
We’ve built some exciting new features into Intercom using this new GPT technology and have given access to several hundred customers – and the response has been astounding. We’re incredibly excited to reveal these beta features today, and discuss how we’re thinking about the application of AI technology in customer service.
Introducing our AI-powered beta features
The first feature we built is a conversation summarization tool.
There is a lot of important context sprinkled throughout customer support conversations, and support reps often have to write summaries before handing conversations over to teammates.
“A support rep can simply click the ‘Summarize’ button to generate a detailed TL;DR of an entire customer conversation”
Enter our new AI-powered Summarization feature. Large language models are fantastic at reformatting or reprocessing text that’s already written, so they’re perfectly suited to condensing text. Now, a support rep can simply click the “Summarize” button to generate a detailed TL;DR of an entire customer conversation.
Our beta customers have been particularly excited about the Summarize feature over the last few weeks:
“The AI Inbox summarization feature is great for catching up on conversations my colleagues have had with users, and also for quickly entering a summary into our bug tracking system. Great stuff!”
Robin Salimans, CTO at Luna
“I say this about every new Intercom feature, but this time, I’m telling the truth – this could be my favorite! As a manager, being able to instantly summarize a conversation between my team and the customer is amazing. It saves me having to read through blocks and blocks of emails or chats, which is not only a huge time saver for me, but also helps to highlight areas for improvement within the team. It’s also great if someone is out sick and the team needs to pick up their tickets. They can quickly summarize them to understand what’s going on. All in all, I love this. It’s really going to be life-changing for us – 10/10 on this one.”
Dean Kahn, Customer Experience Manager at RateMyAgent
It’s not always perfect – sometimes a teammate does have to edit a summary before handover – but even then it saves a lot of time, and we’re working to refine it further.
Composer AI features
Our customers spend almost half of their Inbox time in the composer – the text field where they write messages to their customers.
Even before the recent GPT models came out, we were investing substantially in AI to make this process more efficient.
Over the past month, we wanted to see if we could use GPT to reduce friction and give our customers a more delightful experience, so we started with some simple but magical tools – some of which delivered even more value than we expected.
Adjusting the tone
It’s common for support reps to use different tones of voice with different customers – depending on the industry they work in, or the type of query they’re answering. Editing text for tone can be tiring, but it’s something that GPT excels at. We’ve added toolbar buttons to our Inbox composer so you can simply click to make your response friendlier, or more formal.
The writing effort required for customer service is often underestimated. With the sheer volume of queries arriving in the Inbox, it can be difficult to find the right words. If your message gets the meaning across, but you aren’t happy with the wording, you can now press a button to rephrase it.
Both the tone adjustments and rephrasing seem to work fairly reliably for people, and provide small but real delight.
“Our team has been using the new AI Inbox features a lot. We operate a q-commerce grocery delivery company, and have found the tools useful for generating new responses to customer issues, adding variety to our usual canned responses. Overall, we’re really impressed with the capabilities and the intuitive user experience!”
Richard Moyles, Head of Customer Relations at buymie
We’ve also gotten positive feedback on the general idea of “paint” style editing, where small adjustments to text – like tone and phrasing – can be achieved at the click of a button, instead of having to do a rewrite. We think this might be an emerging trend – that we will soon be able to adjust tone similarly to how we make text “bold” today.
A more significant feature we’ve been working on, and which is still early in its journey, is Expand. The idea is that you can just write short notes or bullet points, and use the Expand tool to elaborate and turn them into a fully fledged response.
A number of our customers love this feature and have had great success with it so far:
“Intercom’s new Expand feature is a game changer for me and my team at Kala Burdo Consulting. It allows us to quickly jot down notes during customer calls, which can then be expanded to full replies that we send as follow-up messages after the call. This saves us a ton of time and enables us to assist even more clients. It’s an essential tool for any business looking to streamline their communication and improve efficiency.”
Kala Burdo, Owner at Kala Burdo Consulting
However, suitability does vary per customer. Sometimes the expansions aren’t quite what the teammate wanted, and they end up using Rephrase or even manual editing to reshape the text.
The number one piece of feedback we’ve gotten from our beta is that people want Expand to be more contextually aware of their conversation, and how they normally talk.
We’re working on a prototype which pulls in your previous replies and even relevant macros. We’re very excited about that, although it will take a little more tuning to reach our quality bar. We think this could unlock a lot of the value generative AI promises, saving teammates minutes finding relevant content and rewriting it each time. We’ll be sharing progress on that as we go, and you can sign up to stay informed about it.
Generate a help center article from shorthand
This last feature is similar to Expand in the composer, but works even better because help desk articles are long-form. Something many companies struggle with is coverage – writing enough articles to cover their customer questions. This feature lowers the friction of writing help documentation. We hope this not only speeds up our customers, but encourages them to write more help content.
While it does need to be double-checked for accuracy, it can be really assist in getting to a first draft – often the hardest part, as anyone who has written a lot of help desk articles knows! This feature is one of the most recent we’ve shipped, and so we’re still learning about its potential – but we suspect AI text completion will quickly become a table-stakes feature for most long-form content editors.
“Within just three minutes of trying the new AI features, I discovered how helpful they are for my job. One thing I was super impressed by was how I could use the new functionality to help me write help center articles. The fact that it can generate coherent content about complex technical issues will save me so much time.”
Andrew Dell, Chatbot Developer and Technical Specialist at Super Dispatch
Will ChatGPT be able to answer all customer queries?
We’ve talked about the features we have in beta that are designed to help teammates work faster. But one of the most common questions we’ve been asked on this topic is “can ChatGPT just answer our customer queries?”
We don’t think it can today – or at least not “out of the box”!
“Our testing shows that hallucinations – where it confidently gives wrong information – are still too big a problem”
First, models like ChatGPT do not know your specific business questions and answers by default; and second, our testing shows that hallucinations – where it confidently gives wrong information – are still too big a problem.
Intercom’s existing Resolution Bot uses large neural networks to answer questions. These are great at answering known questions – even questions phrased unusually – and never hallucinate. But they aren’t as good at understanding free-flowing or multi-sentence dialog as ChatGPT is.
There are techniques emerging to reduce hallucinations in large language models, and we’re investing heavily to see if it’s possible to get the best of both worlds: a system with the conversational understanding of modern generative models, but which also has the accuracy our customers need to trust it. We’ll have more to say about this soon.
What’s the bigger picture for customer service?
So we’ve talked about the features we have in beta and some of the things we’ve learned from our customers. But what’s the bigger picture for customer service?
The reality is that we’re just at the start of figuring out the innovations this technology will unlock. Based on what we’ve learned so far, here’s where we have conviction:
- We believe recent and upcoming developments in AI are the most disruptive technology to hit the customer service industry in decades. The reaction of our customers to what we’ve built gives us confidence that the value here is real.
- We’ve long believed that almost all companies will eventually be able to automatically resolve most of their customer questions; our recent experience with this technology has convinced us this timeline has moved forward.
- AI will not be a point solution that injects bot dialog, but instead a technology that permeates the platform. Yes, it will help bots answer customer questions; but equally it will route conversations to the right team, speed up teammate response times, create help center content, and enable teams to understand the trends in their conversations.
- The current technology constraints are real. Specifically, the latest LLM models “hallucinate,” meaning they just make things up. And they do this frequently, and convincingly. The challenge now is for companies to build a product around the constraints of the core tech, shape it to the actual needs of customer service teams, and keep adapting as the technology advances.
- The future is automation-first, but it is not automation-only. There will be a durable and sizeable need for humans to handle more complex questions, solve problems for customers, and provide a genuine human connection. The future of customer service platforms is a well-integrated combination of the automated and the human.
We’re extremely excited about how recent AI breakthroughs will change customer service. We’re investing heavily in this area, and we’ll be sharing our progress as we go.