The following is a lightly edited version of my keynote from last week’s Pioneer AI summit.
I’ve been in this game for a long time, from the early internet through mobile and social, and I have never seen a new technology as disruptive as AI. It’s already transformed customer service in ways we could have only dreamed of a few years ago.
We’re living in a remarkable time, moving rapidly away from a world where customer service leaders are trapped in a never-ending battle against volume, headcount, and staffing. Now, we can deliver instant, accurate, and reliable service, 24/7, in almost any language. It’s a total transformation.
Since the beginning of Intercom, our mission has been to make internet business personal. We do that by building tools that let you deliver a concierge-level of service to every single customer, every single time.
Now, we’re taking a huge step forward by announcing Fin 3.
From immediate utility to high quality
Just two and a half years ago, in March 2023, we launched the first version of Fin. It was built on GPT-4 (released the very same day) and was the very first AI agent for customer service, answering informational queries from a knowledge base.
The impact was immediate, but we learned very quickly that this wasn’t a small product. The way our customers used Fin showed us it had big, deep, and broad applications. So we relentlessly shipped product updates, and we saw something remarkable happen.
Over the course of its first year, Fin’s average resolution rate across our entire customer base grew from 23% to 52%. That’s not a cherry-picked number, it is the average number across our thousands of Fin customers.
A year ago, we launched Fin 2, which showcased our relentless focus on quality. We had an AI agent that could truly deliver human-quality service, with the right brand voice, connecting to important internal and third-party systems.
We also introduced the Fin Flywheel, a continuous improvement loop of training, testing, deploying, and analyzing:
Charting our improvement
Since the launch of Fin 2, the average resolution rate climbed from 52% to 66%.
If you look at the chart, it’s just this consistent, continuous line of improvement, gaining about 1% every single month without fail.
That’s across over 6,000 customers and millions of real conversations. A fifth of our customers are already seeing resolution rates over 80%. I don’t think any other company has a chart like this.
So, how are we doing it? ITo build the highest-performing product, you have to invest deeply in three distinct layers.
First is the app layer, this is the Fin Flywheel, and the many features therein that our customers use every day.
Below the app layer, is the AI layer. This is our RAG system, which we believe is the best in the world for customer service thanks to hundreds of A/B tests to experiment and improve performance.
Below that, there’s the model layer. This is the layer that the RAG system depends on.
General models like GPT and Claude are incredible, but they aren’t specialized for the nuances and complexity of customer service.
To optimize the performance of Fin, we started building our own custom models, trained on our huge dataset of historical conversations. We now have 5 custom models in production that outperform general models.
We’re making these huge, often invisible, investments in the AI layer, and model layer, because we believe that in the long run, the highest-performing product will win.
The one thing every service leader cares about most is delivering the best customer experience, and to do that, you need the highest performing AI Agent .
That brings us to Fin 3.
The current frontier is complex queries
Today,our average resolution rate is 66%, but our goal is for Fin to do more and more work for our customers, and deliver 100%.
In developing Fin 3, we focused a lot on query type. Not all resolutions are equal. A 30-second FAQ is not the same as a 30-minute billing dispute over the phone. If Fin can resolve more complex queries, then Fin is doing more of the really complex, time-consuming, and expensive work that your human team do today
Fin 3 can resolve all types of complex queries, across all customer channels, thanks to big updates to all parts of the Fin Flywheel:
Sophisticated training with Procedures.
More powerful testing with Simulations.
Across more channels with Voice, Slack, and Discord.
Richer insights with improvements to Insights.
Procedures
Natural language instructions with Procedures
Procedures lets you train Fin to resolve your most complex, multi-step queries that once required human judgment, coordination across teams, and careful manual steps. These are the kinds of cases where accuracy really matters: billing disputes, cancellations, refunds, anything with a high cost of getting it wrong.
First, you can describe what you want Fin to do in plain natural language. For example, check a customer’s subscription status, confirm the type of change they want, determine eligibility for a refund, and so on.
From there, layer in precision where you need it most. Deterministic controls let you hard-code logic into the flow using data connectors, branching conditions, or even raw code. It’s the best of both worlds: the speed and flexibility of natural language paired with the certainty of strict rules.
Then, Fin approaches every step agently, reasoning about what to do next rather than marching through a fixed path. If the customer interrupts, changes the subject or needs clarification, Fin can adapt, pull the right data, and keep the conversation feeling natural.
Last, we’ve built an AI assistant to help you configure all of this. You can describe the process you want, attach any SOP docs you already have, and the assistant will draft a first version for you, pulling context from your content and past conversations.
The result is a simpler, smarter way to automate the hardest parts of customer service — replacing complex manual workflows with something intuitive and resilient, that always feels human to talk to.
Simulations
Complete conversation simulation
When you give Fin the power to handle your most complex workflows, the next challenge is trusting that everything will work exactly as you intend before it ever reaches a customer. That’s where Simulations come in, an automated testing capability that’s a sister product to Procedures.
It runs a fully simulated customer conversation from start to finish:
You define the user, their query, and the success criteria, and watch the simulation run.
If it fails, you can see exactly where it went wrong, debug it, and even get suggestions from an AI assistant on how to fix it.
Simulations are a way to safely test Fin’s behavior end to end, before you deploy. They’re a full-fidelity rehearsal space where you can run automated conversations and watch Fin think through them step by step.
If something fails, you can see precisely where and why. Each step of the simulation is logged and visualized so you can debug quickly — adjusting your procedures, your branching logic, or even the data sources that Fin depends on.
There’s an AI assistant here too. You can chat with it, and if a simulation fails it will suggest changes.
You can also save test cases in your own library of simulations. Run them any time you update a product, change a policy, or roll out a new integration, and instantly confirm that Fin’s still performing as expected.
Simulations gives you software engineering-level stability, with continuous testing, fast feedback, and automation. It means every Fin interaction can be rehearsed, refined, and proven long before it reaches a real customer.
Channels
Fin already works across email, messaging, and voice — and now we’ve added Slack and Discord so you can meet people in the places they actually use. In Slack, Fin behaves exactly like you’d expect: threaded replies, emoji reactions, teammate names and avatars. It just fits into the team workflow. Discord works the same way for communities and audiences that live there.
The Fin Voice Demo as it happened
Fin Voice has been the biggest lift — and the biggest payoff. The default phone experience today is still “voice jail”: long holds, endless menus, and shouting to reach a person. That shouldn’t be normal. We launched Fin over voice and then poured work into configurability, testing, transcripts, summaries, and integrations with all major telephony providers. So when I called Fin onstage it did everything a human rep would, and it did it fast: confirmed my identity from the phone number, pulled up two subscriptions, tolerated my interruption, clarified which subscription I meant, offered the business-friendly pause instead of a cancel, and sent an email confirmation — all inside roughly 90 seconds. There’s a bit of latency while Fin looks up systems; that’s just the system doing real work. This is the real, unedited experience, and it’s already far better than waiting on hold for twenty minutes.
The point is simple: omnichannel matters. The more channels you deploy Fin on, the more work Fin can take off your people, and the better the customer experience becomes. We built these channels to be easy to adopt and to behave like native experiences — because friction kills adoption. Try voice, try Slack, try Discord. You’ll be surprised how much of the old pain disappears.
Insights
If you’re running Fin at scale you need to know how it’s doing and where to improve it. That’s the Analyze step of the flywheel, and we’ve vastly upgraded Insights to make it actionable.
CX Score gives you full coverage across conversations — I think it’s a far better metric than relying on CSAT alone — and now CX Score comes with reasons, so you can see whether a low score was due to product quality, policy, answer quality, or something else. Those reasons are first-class attributes you can filter and segment by.
Topic Trends automatically surfaces spikes and emerging issues that used to require manual digging. If resolution rates drop for a topic, or a new complaint is spiking this week, you’ll see it early and act fast. Because you asked for it, we gave you full curation control: rename, merge, or create topics so your reporting matches how your business actually talks about things.
AI-powered suggestions are smarter too.
They spot duplicated or contradictory content, recommend missing knowledge, suggest data or actions to add, and learn from the edits you accept or reject so suggestions get better over time. And we don’t care where your content lives: Fin works with Zendesk, Salesforce, and the major platforms — accept a suggestion and we’ll update the source for you if you want. The goal is the same whether it’s a Zendesk doc or an internal SOP: make the customer experience as good as possible. AI suggestions are possibly the biggest leap forward for helping you answer more and more complex queries.
This is all designed to be self-service. I want teams to be able to experiment, iterate, and own their customer experience without waiting for vendor engineering cycles. The flywheel only works when you can train, test, deploy, and analyze quickly — and when the tooling puts control in your hands. That’s how you turn automation into better experiences rather than brittle processes: measure, learn, and improve.
Here’s a closer look at each feature.
Building for you
One last thing that’s really important to us is how we build all of this. We believe that customer experience is something every company should own for itself. You should be in total control of it. That’s why everything in Fin is built to be self-service. Other companies create a dependency, forcing you to contact them for changes, which is slow and turns them into consulting-ware. We think that’s the wrong approach.
It’s much harder for us to design these powerful tools to be easy for anyone on your team to use, but we believe it’s the right way to do it. You should be able to experiment, change things, and control your own destiny.
So that’s Fin 3. With Procedures for complex queries, Simulations for automated testing, major upgrades to Voice, new channels like Slack and Discord, and smarter Insights, Fin can now do more work for you than ever before. This frees up your team for higher-value work and continuously improves the experience you can deliver.