Intercom's decisive bet on AI
Co-founder Des Traynor takes us behind the scenes
👋 Hi, it’s Kyle and I’m back with a new Growth Unhinged, my newsletter that explores the unexpected behind the fastest-growing startups. Subscribe to join 71,685 readers who get Growth Unhinged delivered to their inbox every Wednesday morning.
Intercom is one of those companies that’s become a household name within software.
Founded in 2011, Intercom has raised nearly $250 million in funding. They’ve been named to the Forbes Cloud 100 for eight (!) consecutive years. The business generates “hundreds of millions” in annual revenue. You can see the product out in the wild at just about every up-and-coming startup.
Even still, Intercom was arguably the first major SaaS company to go all-in on AI along with a disruptive outcome-based pricing model, which they launched in early 2023. That was more than a year ahead of most competitors. Now even Salesforce is pivoting their AI pricing, positioning it as “a consumption-based model that aligns cost with business outcomes.”
Today, Intercom’s Fin AI agent is doing “tens of millions” in revenue and has quadrupled in size over the past year. Fin now autonomously resolves 56% of conversations for the average customer, more than double where it started (~25% resolution).
I’ve been a fan for a while and knew this was a story that needed to be told in Growth Unhinged. Thankfully, co-founder Des Traynor agreed. Des has been instrumental in shaping Intercom over the past 14 years and I asked him to unpack Intercom’s decisive bet on AI.
My top takeaways from our conversation:
It took a ‘war time’ CEO to be the first to go all-in on AI.
Each 1% improvement in Intercom’s resolution rate was hard fought with constant A/B testing. Most of the good AI tools are not simply thin ChatGPT wrappers.
Competition will no longer be about who has the most features. It’ll be about how well those features work.
The winners will be the ones who write the RFP for customers, teaching them how to evaluate and buy AI products.
If AI agents are doing the work for you, the pricing model needs to match. But prepare for potential trade-offs around attribution and predictability.
Users don’t necessarily see whether an AI agent is getting better. Educate them about improvements and empower internal AI champions.
On being first to go all-in on AI
Rewinding back the clock to October 2022, Intercom co-founder Eoghan McCabe returned to lead the company as CEO. Eoghan made it clear that he was taking Intercom back to its roots with a singular vision: become the dominant platform in the customer support space, and redefine that space.
Less than two months later, on November 30th, OpenAI released ChatGPT for public use. By Monday Eoghan made a decisive call that Intercom would be all-in on AI to automate work. Their first modern AI release was on January 31, 2023. Fin came out (in beta) only six weeks after that.
I asked Des how such an established company could possibly turn the ship so fast? He attributed it to two things: (a) Eoghan being ready to be a ‘war time’ CEO and (b) having an established relationship with OpenAI prior to ChatGPT (via a product called Resolution Bot), which made Intercom well positioned to act quickly.
“With ChatGPT’s first release, all the chat immediately was about hallucinations,” Des recalled. “As a result, we decided to make our first release agent-facing and not user-facing.” These features, including things like conversation summaries and inbox improvements, proved to be extremely popular. The team immediately deprioritized other projects to push out AI inbox features with a focus on taking something that customers were already doing and automating it.
Then Intercom got access to GPT-4. “It was very clear that something was here, but also that you need to build a lot of software around GPT,” Des told me. This became the catalyst to go even bigger.
Why most of the good AI tools are not thin ChatGPT wrappers
There’s a lot of dismissive online chatter about AI products being thin wrappers around ChatGPT. The implication is that they aren’t durable, will quickly get commoditized and could easily go to zero as LLMs improve. Des disagrees.
“There are genuinely thin ChatGPT wrappers where people build a thin shell around an API call, but most of the really good AI tools – like Granola, Lovable and Intercom – are not thin wrappers,” he emphasized.
Des says that with Fin there are 15 different sub-processes involved in answering a customer support question. “It’s not just given these docs, answer this question. It’s everything from asking who is Kyle, what screen is Kyle on, what’s the current state of Kyle’s account, what question is Kyle asking, what language does he speak, how urgent is it?”
Many incoming support questions don’t have a help doc about them. There’s a detailed reasoning and Q&A process that needs to happen in order to get it right. “Today, Fin is a long, complicated architecture of a bunch of stuff and each is extremely fine-tuned to exactly what it needs to do.”
Initially, Fin was at a roughly 25% resolution rate. It’s now at 56%. According to Des, “Each of those points was extremely hard fought with constant A/B testing.”
It’s worth noting that Intercom’s outcome-based pricing model – charging $0.99 per successful autonomous resolution – makes Intercom’s R&D team extremely motivated to fight for those basis points. Improvements to the resolution rate immediately flow through to revenue, making R&D a revenue-generation function. (Intercom also just released a bunch of new insights features designed to help customers drive continuous performance improvement.)
How to position AI products when everyone’s marketing sounds the same
I kept thinking to myself: this all sounds great, but doesn’t everyone say something similar? I could probably pull up at least a dozen websites that would all boast about having the best AI for customer service. If I were a buyer, how could I tell them apart?
Des acknowledges that this is a hard problem to solve – and it’s fundamentally different for AI products compared to B2B SaaS. It becomes not only a feature war, but also a quality war and a price war.
“In B2B SaaS there has been a competitive battleground around who has the most features. There was an assumption that if you had a feature, it worked. Now there are all these extra variables. Users of software have to ask: how well do you do this feature?”
Intercom’s bet on how to solve this: be the one to educate the buyer on how to buy while making this education feel fair and not biased. (They want to be the vendor that writes the RFP.)
They recommend that buyers pick 100 questions that are real and suitably hard to answer (not basic ones like ‘how do I reset my password?’). Then train the agents on the documents they need to answer the questions – this works to Intercom’s advantage because Fin can be deployed by pasting the URLs of the docs and then hitting go. From there, measure the performance of the vendors specifically looking for two things:
How many questions can each answer?
How good are the answers? (Grade the responses across vendors)
For other companies, Des suggests figuring out the relevant competitive metrics you’re tackling for the job you’re displacing. These metrics will look different for a support agent versus a writing tutor or a legal AI product. Ask yourself: how can you show buyers you’re completing the work at the same standard and at a far lower cost?
Des recognized that none of this answers the question of what your homepage should say. “Everyone is going to say better, cheaper, faster. There will be a bunch of products that look the same, but the ultimate vector will be quality and then taking that into your branding to show why you’re great.”
Intercom’s branding is second to none; they might have my favorite homepage out there. Their work goes far beyond the homepage, though. Intercom launched a dedicated research blog to demonstrate their depth of technical expertise – even at the potential risk of sharing sensitive information with competitors. And they provide resources to help customers understand the trade-offs between build vs. buy vs. a manual approach (below is an example).
Being at the forefront of outcome-based pricing
Of particular interest to me was Intercom’s first-to-market outcome-based pricing model. Intercom charges customers $0.99 per successful resolution by Fin.
“The thesis of Fin is that we’re doing the work for you,” Des highlighted. “We made a deliberate decision to trade off certain aspects. The biggest challenge was that we had to make sure a customer knew what a resolution was.”
Looking back, Des says that the attribution challenge “ended up being easier than we thought.” Intercom defines a resolution as when either a customer marks an issue as being resolved or exits a conversation and doesn’t come back.
Customers have responded well to Intercom’s outcome-based pricing, although their biggest pushback has been predictability. Intercom’s average resolution rate stands at 56%; however, some customers still sit in the 20% range and others see as high as 80%+ resolution. Intercom now lets customers control their budget by setting caps on their Fin spend.
“Another challenge is that people see this as sitting on top of their current spend,” Des mentioned. Intercom solves for that by helping customers model out their ability to handle a growing support volume while headcount growth can either stay flat or grow at a much slower rate. “This isn’t just a bet on right now, it’s a bet on the future, too.” It helps Intercom that customer service is a high attrition role. There’s high turnover in the function and particularly high turnover in ‘Tier 1’ roles (people either graduate to Tier 2 or move to another function).
The other message Intercom tries to reinforce is the customer’s ability to deliver better support with AI. “There’s a world of difference between getting your query resolved in nine seconds versus nine minutes. People contact support more and their user behavior changes because they no longer need to wait for a call or talk to an agent.”
Coming up with the pricing model started with modeling out the alternative cost for the customer, looking at the fully loaded cost-per-resolution of an in-house customer support rep or an outsourced vendor (BPO). Intercom modeled this out for resolving both complex and easier questions. From there, Des and team analyzed Intercom’s internal costs, most notably the cost associated with API calls to Anthropic.
There was some art to go with the science. “You don’t want to pick a price that’s 67.25 cents. It needs to be a marketable price.”
Building a UX around AI agents
Another area of interest for me: the user experience (UX) around AI agents. If Fin works mostly in the background resolving support tickets – without people needing to be in the loop – what gets exposed to the customer?
“When we improved from 25% to 56% resolution, most of those improvements weren’t user-facing. They lived in the background with constant A/B tests. It’s difficult to explain to users why things are continuously getting better. The point you need to land with them is that we’re constantly working to improve it, but those improvements will not be UI changes.”
- Des Traynor, co-founder at Intercom
Des expects that a new AI role will inevitably be created in most industries. “There was previously a support leader and now there’s going to be an AI support leader,” he hypothesized. Intercom is designing with this persona in mind, and wants to equip that persona with the resources they need to improve their AI agent. For example, there’s a Fin inbox where customers can look at all the answers Fin gave, allowing customers to see it themselves (and make improvements accordingly).
Des predicts that “every category will have their AI transformation moment”, which will force a major rethink of how work gets done and how much AI could take. Part of this transformation moment is presenting a new vision for what the job looks like in a post-AI world. “Now we have new roles like specialist roles around product ops or an owner for the docs because the docs are more important. Businesses need a new mission statement, org structure and job descriptions – and a lot of customers are still waiting for guidance.”
Advice for the next generation of builders
I asked Des what advice he gives his portfolio as they’re navigating an AI-first world.
“You need to be hungry for the most effective way to build software,” he told me. “The way we design should change. The way we develop should change. A big opportunity will open up to people who build using AI and they can now build 5-10x as fast.”
This means most feature-based moats are either non-existent or far lower than before. “If previously you said you had a moat because it took you seven years and 100 people to build your product, that argument is dead.”
Capitalizing on this opportunity requires building with AI internally – not just for your customers. Des personally recommends Claude Code, Granola, Windsurf, Cursor and Kittl.
Above all, recognize that customers are looking to you for advice about AI transformation, not just what your AI product can do for them. “It’s a mistake to put the sparkly AI button in the corner of your app. It needs to be about what happens to their business.”
Imagine building a successful software company and then deciding to innovate on pricing models. Excited to see the software industry getting one step closer to true value-based pricing.