Written by 
The next generation of customer service has truly arrived – Fin 2 is here and it’s fundamentally changing what’s possible in customer support.
Fin 2 far surpasses any AI agent you’ve seen before – achieving an unprecedented 51% average resolution rate straight out the box and a 99.9% accuracy rate, it’s ready to take on all frontline support tasks, allowing your human team to focus on more impactful work.
We’re lifting the curtain on exactly what it’s taken to get to this point – from the journey we’ve been on with generative AI, to everything we learnt from building Fin’s V1, and feedback we got along the way from our amazing customers.
We’ll also show you how Fin 2’s powerful new capabilities work and present real data from real customers using the technology to prove that it’s real, it works, and it’s transcending the gap between humans and bots.
You can watch the keynote announcing Fin 2 above, or read a lightly edited transcript below.
Transcript
Hello. Good morning. Thank you all so much for being here, and thank you to many, many people watching online. It’s my honor and privilege to announce to you, Fin 2.
Today, there’s a lot of hype about AI in the world. And some of it is real, some of it is, to be honest, a bit of bullshit. It’s hard to see the signal from the noise. There are many products that claim to do things that they can’t, or companies who have big numbers – and these numbers are cherry-picked numbers – like 80% resolution rate – that might be a one-off customer or after 4 or 5 months of deep integration partnership.
What we want to do today with you is just get real. We want to show you our real numbers. All the numbers we’re showing you today are not cherry-picked. They’re numbers across our entire customer base. We’re going to send you the numbers, and share the numbers of every single customer. So, you can have confidence that Fin really works.
These numbers are quite remarkable to us. They are going to change our industry. We’re at a point where we’re seeing really, really profound change, and many of you here today are customers who are the pioneers leading this. If you’re in the audience and you’re not quite sure yet, or you’re here to learn how it works, I would really encourage you to find the customers of ours who have Fin live and are seeing these transformational results.
It really is a pivotal moment in the customer service industry. It’s a pivotal moment that we will look back on and say this is where things started to change – and really change for the better.
Fin 2: Revolutionizing customer service with real results
So, they’re big words, but it’s really clear in our data. Fin 2 is the best agent on the market. Fin 2 has a 51% average resolution rate. 51%. That’s the average across our entire customer base. This isn’t the number that people spend weeks and months trying to get. It’s the number from the very, very start. It’s a really, really transformational number.
Over half of customer questions that Fin gets asked can be answered accurately and resolved. That’s going to change how the industry works. We’ve also worked really hard to make sure that Fin is accurate, and has a very, very, very low hallucination rate. Lots of our customers, if 51% is the average, of course, are getting much higher. We’ve peaked numbers around 86%. Many customers in the eighties. Again, if Fin is answering 80% of customer questions, it’s going to free your team up to do all sorts of other things.
Resolution rate isn’t everything, of course. We’ve also been working on this big powerful feature set so Fin 2 can do all your frontline support. We have customers here today where Fin 2 is doing all their frontline support, and that means their team is free to do all sorts of new things that they couldn’t do before.
And, the reason for that is the new generation AI Agent. Literally a new generation. To help you think through that, we are going to go back in time a little. We’ve been doing this type of thing for a long time, over many, many years.
From clunky rules to intelligent conversations
We launched Custom Bots in 2018. It was our very first bot product for customer service.
Custom Bots was deterministic, which means it was rules-based with decision trees and canned answers. It worked well if you set it up deeply and well, but that takes a lot of work, a lot of resources, and a lot of people. Many customer service teams are under pressure and they just don’t have the resources to do that. So, it worked well, if you set it up well, but not all of us could do that, and we ended up with bad bots – “shit bots,” not chatbots.
And, bots kind of got a bad reputation. Honestly, if you’re going to blame anyone for giving bots a bad reputation, you should probably blame us because we invented the thing that started off the bad bots. But, we’re working on something better.
We’ve had a machine learning team for nearly a decade, and in 2020, we launched Resolution Bot. Resolution Bot was a new type of product. It worked in different ways to Custom Bots. If you ask Resolution Bot a question, machine learning works out the best answer. For some customers, they invested a lot in it and it worked excellently. But it was still a manual setup. You needed a bank of answers, so you put the work in to build a bank of answers, and then Resolution Bot, using machine learning, could pull out a really great answer, but it was still manual.
The LLM breakthrough was a game-changer for customer support
Up to 2022, we were working on Resolution Bot, trying to make it better, trying to figure out ways we could make that kind of curation set up step way easier. And honestly, suddenly, everything changed overnight.
ChatGPT launched and showed what was possible. Over the course of a weekend, we saw that this was one of those very rare “before-after” moments in technology – one of those one-way doors that once you go through, you just don’t go back.
Technology only moves in one direction, and we knew this was gonna change everything. It was the dawn of the age of Large Language Models (LLMs) and LLMs solved curation problems for customer service. It meant that they could generate answers in the moment. It meant that a whole new approach was possible, and this was a thing that started to change our industry.
At Intercom, we acted immediately. Over the course of that weekend, we changed our strategy and our road map. We knew this was going to lead to a very different world. We acted really fast because we saw this mapping of things customer service reps typically do, and the things that Large Language Models can do. There was just this huge overlap.
Straightaway we could see there were profound implications because if AI can do many of the things that frontline support reps do, what will they do? What could they do? They could work on higher impact things. They could work on new types of things. They could spend more time with customers. Some customers need a lot of time and empathy, and if people were freer, they could do that type of thing. We realized if you change the fundamentals you change the industry. And that’s what happened.
The journey with Fin V1
Within one month, we shipped our first AI features. And within four months, we shipped Fin into beta. We’re going back now 18 months. Fin was the world’s first LLM-powered customer service bot.
Fin is very different to Custom Bots and Resolution Bot. No rules. No trees. No setup. You feed it content, and it basically works out of the box and generates answers in real time.
Fin’s performance surprised us. We could kind of guess the possibilities, but from V1, we had 23% of resolutions out of the box. It was a really, really high number. That’s nearly a quarter of customer support questions. But, Fin V1 couldn’t answer a lot of the time. You have complex questions, you have troubleshooting. Sometimes Fin gave the wrong answers, and that’s not acceptable.
It wasn’t acceptable to us. Humans make mistakes all the time; we give wrong answers all the time too. But, we needed Fin to be better than that and give accurate answers as often as possible. So, we invested really deeply in improving Fin’s performance. That’s been the obsession over these last 18 months.
It was huge work, it was hard work. It was very, very technical work. We ran over 120 A/B tests over the course of this period. Each point on this chart was really hard won, and we learned – and it’s kind of obvious in hindsight – that the harder questions as you go up this chart, get harder. The harder questions are more complicated. The harder questions mean that you’ll get them wrong more often with higher error rates.
And so, oftentimes, when we did these 120 experiments, we’d get the resolution rate up but the hallucination rate would increase as well. So, Fin could answer more questions accurately, but it would also make more mistakes. That wasn’t good enough for us. We had to keep the hallucination rate neutral or improve it.
We worked and worked and worked and worked because we wanted to deliver a bot that can both answer questions accurately as often as possible, and not make many mistakes. We got to the average that we have now of 51% out of the box pretty much. In the first few days or weeks of people going live with Fin, they get 51% on average. Many get 70s and 80s, and I’ll show you that in a second.
These are real numbers. This is our entire customer base. There’s no cherry-picking here. These are our real numbers. These numbers change industries. And you just don’t go back once you get these types of numbers, you don’t go back.
We also heard a lot of feedback from customers – a lot of you. Thank you. The more Fin customers we got, the more feedback we got. We knew that resolution rate wasn’t everything. There were a ton of other things that the product needed to do. You needed ways to test it, ways to try it out to actually see if it works well in certain contexts versus others. You needed Fin to use new content types, like different types of internal content, external content, content sources, PDFs, Wikis, all sorts of stuff.
Once you get all that knowledge into Fin, it’s obviously hard to manage, so you needed ways to manage the knowledge. You needed reporting to see if this thing actually works. You needed to understand how it works. You needed dashboards. And then channels. We’ve built all of this over the course of this period: WhatsApp, social channels, SMS, and most recently, email. So, Fin works just as well in email as it does in chat.
This is important because when a lot of people think about AI and customer service, they think about chat. AI in customer service is not about chat. That is the first channel you see, it’s the primary and dominant channel today. But, this technology works just as well in email. I’m going to show you that in a second. So, AI and Fin is not about chat – it’s truly, truly omnichannel. We have some customers using Fin to answer tickets. So, this thing is really, really, a much bigger deal than chat.
We went early and we built a big lead, and invested in this resolution rate and all these features. It’s important to me to share this with you because of the hype out there. People are claiming big product changes and big numbers. I’m giving you the real numbers. These are just the facts. I can’t tell you there’s bullshit out there and then bullshit you. That’s not credible. So, these are the real facts.
We are one year ahead, and I think other companies that have recently launched their first versions of these products are going to find that to get something like 51% average resolution rate is going to take a long time. It’s a hard and technical journey and I’m really proud of us for going on that journey and building a product that’s truly good. We’re improving the fastest.
Behind the scenes of Fin 2’s AI engine
But, how did we get to this industry changing number? We’ve built this deep, sophisticated AI engine. I’m not going to share all the details of it, but our AI team is here in the room if you can find them later on. They’ll be more than happy to take all of your questions. I’ll explain in a brief way.
The first thing Fin does is a phase called “refine the query.” When you look at customer service questions, a lot of them are “I’m stuck.” Or, “it’s broken.” And you can’t really do much with that. And so, Fin refines the query. It works to better understand what is the actual problem, because the better the query you feed the system, the better the chances you can find the right answer to that question.
So, we have a “refine the query” phase. We’ve built a RAG system specifically for customer service. It’s tuned to customer service and that RAG system generates the response. But a big thing with Fin is it doesn’t guess. Fin wants to try and answer the question as accurately as possible.
We have a third step that involves accuracy and safety checks, where we check the answer that Fin wants to give. Is it the right answer? Is there a better answer? How can we deduce those types of things and make sure that the answer that is delivered to the customer is actually a really good answer and actually does resolve the query? This happens in milliseconds. Customers don’t experience any of this. They just get immediate answers.
The system also has two other parts. One is an optimization engine. This improves the efficiency of the system that we’ve built. And then lastly – and maybe most importantly – there’s a safety and security system within the Fin AI engine. This makes sure that Fin is compliant and adhering to regulations, legal things and third party data usage. Obviously, it’s critical that Fin doesn’t accidentally say something it shouldn’t or share data externally that it shouldn’t. So, it’s really important, and we’ve invested a lot of time in that too.
Powering Fin 2 with Anthropic’s Claude
The next thing about the AI engine that we’re announcing today is very, very exciting news. We’re announcing today that Fin is now powered by Anthropic’s Claude.
Anthropic are here in the audience today. They’re an incredible company and have built this incredible Large Language Model. Claude is just brilliant for Fin. We have run so many tests on all models. Every time a new model comes out from all different providers, we test it. We run millions of conversations through these different models to see which one works the best. Claude works the best. It’s honestly that simple. Claude performs the best. So, Fin runs on Claude and we’re announcing that today.
Anthropic also use Intercom for customer service, which we’re very proud of. We’re excited to talk about that as well. Again, an incredible company, and we are very proud that one of the best AI companies in the world have chosen Fin to run their customer service. It’s just great validation for us personally that Fin is real and it really truly works.
A look at our customers’ real resolution rates
This chart is maybe the most important thing for me to share. When I saw this, it gave me the grounds for what I’m talking to you about. This is the distribution of resolution rates across our entire customer base.
So, you can see the kind of big numbers I shared earlier on the left and the right. But, in the middle, is the distribution. You can see that across our customer base, these are the resolution rates that they’re getting. This is real data. This is all of our Fin customers, no one excluded. This is a no bullshit chart. You have to pause and look at it because when you see it, you realize that these are industry changing numbers.
Once you have a distribution like this, where many customers are getting 20, 30, 40, 50, 60, or more percent of the resolutions answered by AI and by Fin, it totally changes what you can do. It totally changes what your people can do. What’s also really important about this is this is worst Fin will ever be. This is the worst the technology will ever be. Technology only goes in one direction.
All of these numbers are going to get higher, and the distribution chart is going to go from left to right. Those in the 20s and 30s today will go into the 40s and 50s. If you are in the 40s and 50s today, you will get into the 60s and 70s, and so on. This is going to get better and better and better. It’s really important to internalize this.
Fin is a win-win-win for our customers
You may think, “Paul, you work at Intercom. Of course, you’re gonna say this.” I’m selling you a great product. So you don’t have to listen to me. You can listen to our customers. We have many customers using Fin, and many of them are here today. And again, if you not, we’d love you to talk to them.
This is what we hear from them. We hear that it’s basically a win-win-win. Fin customers have happier customers, because they are getting immediate answers over half of the time. They have happier human agents. Because the human agents are working on more interesting things. They have more time spent with customers, there’s more time spent with customers who need more time. They’re working on new things, like I said earlier.
We also have happier leaders. Many of you are leaders in the room. You can scale your support without scaling your team. I often talk to people about this, and we all admit to each other that no one really likes hiring. We love our team, but we didn’t like the part where we had to hire all the people. And for customer service, this can be a treadmill sometimes. A lot of people leave the job, especially when the job is kind of repetitive. Then you have to hire more people and it can feel like a bit of a treadmill.
We’ve tons and tons of Fin customers now who’ve basically paused hiring. Fin is doing all of the new things, all the FAQs and repetitive things, and their team is doing all new things, which means the team is much more interested. The job is more interesting. The people are doing things like managing the knowledge, designing the system, designing how Fin works – all sorts of cool stuff like that.
There are also happier operations leaders, because the team can do way more with the same dollars. So, it really is a win-win-win.
An AI Agent that delivers human-quality support
We also hear more and more feedback all the time. And over the last few months, the feedback we have heard most often breaks into two themes: One is the quality of experience matters. People say, “we’re not sure if Fin delivers a great customer experience.” Again, bad bots – we’re sorry about that. Does Fin provide a good experience? Because I know my team gives a good experience because I’ve taught them. I’ve trained them. I’ve coached them. They give a good experience. Does Fin? Does AI give a good experience? That’s one big theme of the feedback. How can I tell if it’s a good experience?
The other one is that AI is a black box. It’s very hard to know how it works. People ask how they can control it, and how they can do things they do with their team, like training, coaching, mentoring. They say, “I don’t know how to mentor Fin. How can I work with it? I don’t really know how it works.”
You are unsure how AI works, you’re unsure if you can replace the team.
Fin 2 can do it. That’s what we’ve been working on. Fin 2 is the first AI agent that delivers human-quality service. That has been our mission, and it can do it. We’ve built it in partnership with all of you. Again, thank you for all the feedback. You can monitor, train, and coach Fin just the same as you monitor, train, and coach your human team. That’s how Fin 2 works.
Fin 2’s new feature categories
There are four big categories, and they map directly to how you think about a great human rep experience. Fin 2 is like your best reps in these four big ways.
- Knowledge: With your human team, you keep them up to date and your best reps are proactive. They keep themselves up to date. Teach themselves new things. They learn the latest info by themselves, so Fin does that.
- Behavior: Your best reps follow your guidance. They follow your policies. They apply good judgment with these things. Fin 2 does that too.
- Actions: Your best reps don’t just give everyone the same answer and close the conversation. They give personal answers. And use personal data to give personalized responses. They take action. Sometimes queries can’t be resolved without an action, so they take action.
- Insights: You have to trust Fin the same way you trust your team. We’ve built a whole bunch of reporting and different things that I’m going to walk you through in a second. This is so that you can trust Fin and feel like you can turn it on for lots of your customer base and it will work really well. We want you to hire Fin for all your frontline support. We want you to do it. We have customers doing it today.
Let’s look at all the ways in which Fin 2 features work in these four categories.
Knowledge
The first is knowledge. You can teach Fin and increase its knowledge. Obviously, the more knowledge Fin has, the more questions Fin can answer. You can connect all sources. We’ve designed this to be easy to use. You do not need dozens of people spending months trying to do all this stuff. No code, no engineering background, no technical background. That’s our goal. Anyone can do it.
Once you connect all these sources, it’s really easy. We’ve built a much better upload system, so new content gets uploaded 10 times faster and Fin gets the up-to-date information immediately. We’ve put all of this knowledge into the central Knowledge Hub, because the more content you bring into Fin, the harder it is to manage and know what’s going on. So, everything’s in one central location, which makes it way easier to keep up to date.
We’ve also built a different type of way for Fin to answer. Most AI agents that are out there pull pieces from the most relevant source. They look up and find out what it is, find the most relevant source and then pull pieces from that. Fin uses multiple content sources.
Fin will look at all these different types of content sources and pull different pieces of the potential answer together and construct a brand new answer, which means it’s more comprehensive. It might be longer sometimes, but it’s a more comprehensive answer and that more comprehensive answer delivers higher resolution rates. So it’s really powerful. It really helps drive up the performance. You only train Fin once. Once you connect to these sources and have them syncing, you update the source and Fin gets updated. It’s really powerful.
Behavior
The second thing is behavior. The first part of this is a new feature called “Tone of Voice.” I think this is our most requested feature of all time in the decade plus of Intercom. You want Fin to speak like you do. You can teach your reps to speak like you do, and now, you can teach Fin to speak like you do. So, you can change the tone of voice of Fin and you can change the answer length of Fin.
Here’s an example of this. This is exactly the same answer by Fin, but for two different companies who have a different tone of voice. On the left, they’ve chosen a short and formal tone of voice, On the right, they’ve chosen a longer, more fun tone of voice. Different companies have different brands and different brand attributes. It’s the same answer, but in different tones of voice. It’s really cool.
The next feature is also really cool. It’s one of my favorites. Real time translation. The challenges here are really obvious. You have many different customers who speak many different languages. And the way to solve that in the past was to hire people who can speak their languages. It’s time consuming and it’s hard and there’s a big overhead in it. You don’t need to do that anymore. Fin can do it automatically. It can automatically translate anything into the customer’s language.
Here’s an example of this. This is an article written in English, which only exists in English. An English speaking person wrote this article.
On the left, Fin is delivering an answer from this article to a customer who speaks Spanish, and, on the right, to a customer who speaks Japanese. Fin is talking to them, giving them the right answer in their own language. The article was never written in those languages. The article only exists in English.
So, you no longer need to localize content. Again, this frees up your team. Think about all the resources you spend localizing content. You can do other things with that resource now. You can have people do all sorts of other things.
Earlier, I mentioned that AI in Fin is not a chat tool. It’s an omnichannel tool. This all works in email as well.
The next thing is behavior controls. What do we mean by that? It’s really critical that reps follow policies. And different companies have different types of policies like how they do refunds, cancellations, billing, all sorts of things like that.
You need to control that. You need to teach people how to respond if someone is asking for a refund. Lots of companies, if they blindly give refunds to every single refund request wouldn’t be around for very long. So, you need to control these types of things and properly understand what the customer is asking for and where the refund request is coming from so you can actually help them. With Fin, you can control this.
We have a new feature called “AI category detection.” That’s the first step here. Every single conversation has an automatically AI generated topic and sentiment. It’s really easy to set up. It’s natural language. Again, we’re trying to make this as accessible as possible to as many people as possible. You just type in the type of guidance that you want to give to Fin, using natural language and Fin will start automatically doing that in the future.
You can set all this up in our Workflows product. Our Workflows product is really easy to use, drag and drop, no code automation builder. You can give Fin hard rules in Workflows. You can give Fin guidance. For example, if a customer is angry, you can pass them to the special team that deals with all the angry people. If it’s a cancellation, you can pass it to the team that deals with cancellation requests. If it’s a feature request, you can tell Fin, “hey, can you file a ticket in Jira?” If it’s a bug, “can you file it in GitHub, but inform the right team about it?” You can set all these rules up. Any imaginable attribute, event, segment, region, channel, whatever. You can set it all up in Workflows.
This actually is the best of both worlds. The deterministic, rules-based world we had before generative AI wasn’t universally bad. It took a lot of time and manual setup to get it to work well, but it wasn’t bad. Rules are there to be followed sometimes, and so with this, you have the best of both worlds. You can combine rules with generative AI answers.
You can tell Fin to do very specific things, and Fin will converse and answer. You could say, “hey Fin, follow these rules. But when you’re talking to the customer, converse with them, understand what they’re trying to do, troubleshoot with them. If you see certain things happen, follow this other rule.” So you can build both of these things together. You can now train Fin the exact same way you train your human team. It’s exactly the same. All the ways you’re familiar with training your human team, you can now do it with Fin.
Actions
Next is actions. We have a really great research team at Intercom. They are called the RAD team, which stands for “Research and Data Science.” They’re very proud of their name. It’s a great name. They do a lot of research into customer service, obviously, for us. We do a lot of research into customer questions because for Fin to answer more and more questions, we need to understand what the questions are and what shape they take.
We see these three huge buckets. You can see the percentages there. Informational questions means that the answer is the same for everyone. Do you ship to the UK? How much is your basic plan? The answer is the same for everyone. With Fin 1, our big focus was informational questions and making sure that it works for real. Getting these real, real, real average resolution rate numbers.
Over the last six to nine months, we’ve started looking at these other categories. With personalized questions where the question might be common – is my flight delayed? – we have lots of people potentially asking that, but the answer is different depending on what flight you’re on. It needs a personalized answer specific to the customer. And an action to do something about it. If your flight is delayed, maybe you need Fin to book you a new flight.
Sometimes, to get full resolutions, there’s a task involved and Fin has to go into an action. Other companies say they have actions. It takes months and many people to set it up. That’s not what we’re trying to do. We’re trying to make this as easy as possible for anyone to do in minutes. That’s our goal.
The first part here is connecting common sources like Stripe and Shopify and Statuspage. Depending on your business, you’ll use different types of services like these. In the cases where you use these very common services, it’s super simple to do. Here, you click Stripe. And then you get this model and you can choose what to turn on. This is automatically pulling from Stripe and we can see all the things that are connected. You can just turn each one on. That’s all you have to do to start having Fin use Stripe data and take actions in Stripe. You just set it up and turn it on.
In figuring out what you want to turn on or how you want it to work, you just set up a workflow. Obviously, you can turn all these things on, but you might want to trigger them in only very specific cases. You can just manage that in workflows. Again, workflows are easy to use – drag and drop, no code. We’re trying to make this as easy as possible for as many people as possible.
Here’s a nice simple example of this. The customer is asking a question about their payment failing. Fin has connected to Stripe. Fin has pulled the data from Stripe. Fin has worked out that their card expired. That’s why payment has failed. Fin is saying, “hey, the card you used is an expired card. You should try a different card.” This happens in milliseconds. A human rep would have to look up the data, look up Stripe, find some other dashboard, so this really, really frees your team.
When Fin can be doing this type of stuff, you don’t want your human team that you’ve trained, hired, mentored to be looking up this information because AI can do it way faster. You want them to be doing way more interesting things than that. Way more valuable things than that.
This is actions. Same data source, Stripe. You just do a one time connection to Stripe. Fin does the actions. Fin can do all sorts of stuff, like handling payment issues, amending an order, changing the address, pausing the subscription. So, really, really powerful, and really, really flexible.
Almost all businesses will need custom actions. For the very specific things you’re trying to do, you can set those up too. Again, it’s really easy to do. You just connect, configure, and use natural language to set this up. Give Fin the guidance that you want it to follow, and Fin can work out what it’s supposed to do. We’ve lots of customers using this in beta, and it’s working really well.
An example here is loyalty points. A customer has a loyalty points program. Lots of people ask, “how many points do I have?” “How many points am I from the next tier?” Again, you can have human reps answer these questions, but they can do way more interesting and valuable things because Fin can do these things now. It’s really easy to set up, and it has huge capabilities.
Insights
The last section is insights. The last question you probably have is, “how do you measure the experience?” Thanks, Paul, for the features, but does it actually work?
The truth about all of this is that trust is built over time. Whether it’s humans or AI, we build trust over time. We usually don’t automatically trust someone. You don’t hire a new human support person, and say “here’s our whole entire customer base of every single imaginable question. Off you go. Have fun.” You know, it’s not how it works. You have to teach them, train them, and build trust over time.
We do this in two ways. Typically, on the quantitative side, we use CSAT, the industry standard metric. On the qualitative side, we manually QA things to see if they’re good.
CSAT is the metric we love to hate. CSAT has no feelings. You can admit it – it’s just not a great metric. There’s two big problems with CSAT:
First, it has really poor coverage. Only about 16% of customers asked to fill in a CSAT survey fill it in, which is very low. That means 84% of your customers are the silent majority. You don’t know what their experience is like. You don’t know if it’s good or bad and you haven’t measured it. Really poor coverage is one big problem.
The second is that it’s unreliable. When you look at CSAT scores, almost everyone answers either 1 or 5. 1 is just all the angry, upset people. That’s pretty obvious. And the fives are complicated. Sometimes the fives are people who got a great answer, but sometimes the fives are people who like the human rep, and they are like, “I don’t know where this is going. Maybe it’s going to their boss. I still have the problem, but I don’t want them to get in trouble. I’ll give them a 5.” So, we’re flying blind. It’s not the best metric.
We wanted to solve this. Imagine if you had 100% coverage of whether the question was resolved, what the sentiment of that conversation was, and how accurate it was. We’ve built this. How do 84% of customers feel? We’re going to find out.
We have an AI product that measures every single conversation, 100% resolution and sentiment. It’s very imaginatively titled, “AI generated CSAT.” It analyzes the question and measures the satisfaction. It measures the sentiment. You get CSAT across 100% of conversations. Honestly, we think this is a huge step forward. This is again a one way door. When you have 100% coverage, you’re not going back to the old way of doing CSAT. You’re just not.
You can compare your survey CSAT with the AI generated CSAT in the product. You can look for patterns there. Think this is industry changing. And ultimately, it’s going to lead to way better service. We think there are so many people in that silent majority who are just getting a terrible experience, but no one knows yet. No one knows. If you don’t know, you can’t fix it. So we’ve built this product.
We think it’s extremely powerful, but it’s not enough. This tells you something might be wrong or bad. But you need to know what to do, so we have another product that we’re announcing today called the “Conversation Quality Report.”
This tells you where to improve. It surfaces the lower performing areas, so now you have 100% coverage and the lower performing areas highlighted. Straightaway, you can dig in and say, “we need better content there, we need better training there.” Maybe it’s human training, maybe it’s Fin training. It’s all there. You can start to dig in and expose what you need. Again, it’s AI-powered, really powerful.
The third thing we have to help is the “Holistic Overview Report.” Fin 2 and AI agents are the future of customer service because they will answer the vast majority of questions. AI is going to do all frontline support. But the future is humans and AI working together. Humans will answer questions for many, many weeks, months, years to come, and they’ll still be better than AI at answering questions oftentimes. There are also humans managing this system. Many of you are going to have to orchestrate all this and work out how it works. So, the future is humans and AI working together. And this report gives you the full picture, the holistic picture of the overall health of the customer service operation you’re running.
In summary, we’ve got an AI-generated CSAT that is going to be industry changing. And we have a conversation quality report that lets you dig in, zoom all the way out and see how humans and AI are performing and how they are working together.
That was a lot of stuff, so I’m going to give you a quick recap and then we’ll finish up. Fin 1’s results have been incredible to us. Really, truly incredible. We went deep to make this thing real, to not have it be an isolated number of customers who got these like results that we can’t replicate. That’s no good to anyone. It’s no good to Intercom as a business, and it’s no good to the customer service industry if these are cherry-picked marketing stories. We had to make it good for everyone. That was our goal. We drove it up over weeks, months, and many, many hard technical experiments.
Today, we’re announcing Fin 2. We believe this is a big step up in capability. Not only do we have the big resolution right now, we have a great customer experience, and AI is no longer a black box. We’re going to keep opening up that black box, so you can get in, control Fin, change its tone of voice, give it guidance, ask it to follow policies – all the things you do with human reps. We truly believe Fin can do all your frontline support.
Controlling the Fin experience
We have one more feature that didn’t really fit into the other categories, and that’s called “Controlled Rollout.” A third big theme we heard from you, our customers, is that you need to test it to see if it works. You wanted to check all sorts of different things like channels, different segments, and regions. So we’ve built this powerful feature called “Controlled Rollout.” You can see how it works here. It’s very easy to use.
People are using this in all sorts of cool ways. For example, they’re turning Fin on for surges. A lot of customer service teams face real problems with surging volume because something happened. So you can set up rules, like “if this team is at capacity, turn on Fin for this segment.” You have full flexibility and control. People use it for out of office things, all sorts of stuff like that. So again, with workflows and controlled rollouts, you have tons of flexibility and control. You can set it up whatever way you want your customer service operation to run.
More capabilities, same price
We want you to think about Fin as your fully capable frontline customer service team. Fin only costs 99¢ per resolution. That’s what it’s always been. We only charge for Fin if it works. We only charge if there’s a resolution. We invented the outcome metric – you only pay if the outcome is achieved. Why did we do this? We could probably make more money as a business in the short-term using seats. We did this because we want people to experience Fin. We want you to try it. We want to make it as cheap as possible, because we think it’s industry changing. We want as many people as possible to use it. We want as many people as possible to free their team and see all the ways in which you can create a way better customer experience.
If you have Fin today, everything is free. We are not charging anything extra for anything I’ve shown today. Fin is exactly the same at 99¢ per resolution, and if Fin can’t answer, you don’t get charged.
Get help from our new professional service team
We’re also announcing a professional service team today. Intercom, as I said, is designed to be easy to use, but obviously AI is a new world, and sometimes you need help. Our professional services team is here too. If you find them, you should ask them for advice. We want to help you. A big theme we hear from lots of customers is, “I’m sold, but I need help. Teach me how to use it, teach me what you’ve learned.” What we’re trying to do today is share all the things that we’ve been learning over the last 12 months. And we now have professional services to help.
First in class
In summary, Fin is the best performing product on the market. We’re really intentional about that. It’s got the richest and most powerful feature set. All of it is real. There is no smoke and mirrors. It’s got a really good customer experience based on the results we’re seeing. We think these are transformational results.
With Fin, customer experience is better. It’s immediate and fast for almost all customers, CSATs are better, and you get off the hiring treadmill. You can invest in your team, free up your team, get your team doing all sorts of new things that you have wanted them to do for years.
Talk to us about these themes, talk to each other if you’re in the room, we have tons of. Find customers using Fin, listen to the other presentations – there’s lots of real evidence. We want to share all this real evidence with you.
And that’s it for me. I just want to finish up by saying a very sincere and deep thank you to everyone who’s talked to us. So many of you in this room talk to me and give feature requests. You talk to our teams, our Sales team, our Product team, our Engineering team, our AI team. Please keep telling us what works about Fin. We are trying to do this in a very open, honest way. Real results are what we’re about. So thank you so much. We couldn’t do this without you.
The future is AI-first. And there’s so much hype out there. It’s real today. I hope I’ve persuaded you of that. Fin works, and Fin works now. It’s delivering these very real results. I honestly, deeply believe Fin or not Fin, Intercom or not Intercom, our industry is changing in profound ways right in front of our eyes.
You shouldn’t fall behind wondering. Earlier, I said you shouldn’t die wondering, but that would take too long. You shouldn’t wait that long. Don’t fall behind wondering. Try it. Try it. It works. And then you can start to see all these advantages on the human team.
With that, thank you so much.
