The Meteor Is Coming: Intercom's Paul Adams on Going All In on AI

Adams opens with the story of freezing on the keynote stage at the Cannes advertising festival while at Facebook. For the first time in his life he had rehearsed the talk word for word under media training, and three or four minutes in he could not remember what came next. He walked off still wearing his microphone, swore, heard the audience laugh, then walked back on and finished. He says it hasn't happened since, but he still feels a tiny doubt every time he walks onto a stage.

His time at Google taught him a different lesson. He worked on social projects such as Buzz and Google Plus that, in his view, came "from a place of fear" of Facebook rather than from a wish to build something great for people. His own research kept finding that people need much better ways to talk with small groups of family and friends, an insight he now sees embodied in WhatsApp and iMessage. When he left Google mid-project for Facebook, some colleagues saw him as a traitor, and Google forensically examined his laptop on the way out.

Adams says those experiences are why he was drawn to Intercom, where co-founder Eoghan pitched him that at Facebook you can design the product but at Intercom you can design the company. He is a believer in high-risk, high-reward bets and accepts that big bets are often wrong. Intercom's principle of shipping fast, early and often follows from that, though he admits it sits in constant tension with a design-led company's desire never to be embarrassed by what it ships.

Adams says people fall into two camps on AI: those who are all in and those who see another round of hype after crypto, Web3 and the metaverse. He is firmly in the first camp and believes AI will transform society as much as or more than the internet. He has woken up thinking about AI every day since ChatGPT launched and still feels he is behind.

For busy people who say they have no time, his answer is that there is no shortcut: you have to make the time. That doesn't mean crazy hours, since he works around 50 hours a week and thinks decisions get worse beyond that. It means setting aside dedicated time to read and to play with the tools, such as the vision features in ChatGPT, which he tried by photographing his dinner. He compares the moment to the mobile wave a decade ago, when companies that explored it early were rewarded.

On strategy, he warns against being sidetracked by cool technology. Instead he would start with the core premise of your product and ask whether AI can do that job fully, partially or not yet. He lists what AI can already do, including writing, summarizing, answering queries, scanning text and images and, increasingly, taking actions like changing a flight. Any B2B SaaS product built around workflows, he says, is in the meteor's path and should decide where AI will replace and where it will augment.

Adams recalls that on the day ChatGPT launched, Intercom's head of machine learning, Fergal, declared it a before-and-after moment. The company threw out almost its whole strategy, asked again why people use Intercom, and went "heads down" on AI, with something of a bet-the-farm mindset. The conclusion was reinforced when Sam Altman said customer service would be among the first industries disrupted.

The result was Fin, an AI-first chatbot that is the first line of defense for a customer support team rather than a person. Adams says the technology is far ahead of customers' ability to absorb it, and the real challenge is organizational change: new roles such as conversation designers are appearing. Fin has also moved into the Intercom inbox, where it suggests answers to human agents and helps them rephrase replies.

On impact, Adams is candid that it is early. Interest is enormous, and some customers can answer 50 to 70 percent of inbound questions with Fin, but it has not yet transformed Intercom's finances. He likens it to the flat start of an exponential curve and says the company has enough evidence to believe this is the future. Support teams, who tend to be high-empathy people, still ask whether a bot can be as kind and human as a person.

Adams describes a series of moments that blew his mind. Before ChatGPT, Intercom's bots had to be taught manually by a support manager; afterward, the model could answer many basic questions by itself, nearly as well as a person. Then it appeared to work things out, and it could write code, which made him wonder how the 1-to-5 ratio of product managers to engineers would hold up and whether engineers would spend more time reviewing than writing.

Vision impressed him even more. A photo of a bike can be diagnosed, and an inscription on a 12th-century cathedral can be translated. He points to radiology, which takes seven years of training in Ireland, as a field where AI seems already better and can ingest every X-ray ever taken, so the job shifts toward checking the AI and delivering bedside manner. In a country with long hospital waiting lists, he says that could be a good thing.

He also touches on the stranger edge of the technology: cloned voices and faces that are nearly indistinguishable from the real person, real-time voice translation that would let a Spanish speaker follow the podcast live, and the prospect of talking to a dead parent. Adams admits he doesn't know whether the last one is good for people. Lenny adds that his own team is working on AI dubbing of podcast episodes.

Adams says everything starts with great machine learning engineers. OpenAI and Anthropic provide the foundation, but teams must build on top of it to get something brilliant, so Intercom has grown its ML team far beyond its former size. Some projects lean heavily on that team, while others are mostly front-end design problems, such as how an AI assistant should interject in a human agent's inbox conversation.

His second lesson is not to treat AI as bolted on. At Google he worked on the small mobile apps team in London while colleagues at headquarters shrugged that nobody would write documents on a phone. Intercom tries to avoid having a single "AI team" and instead wants everyone to learn, in part because Adams is a believer in generalists, with specialists such as ML engineers where truly needed.

For staying current, he reads widely and unsystematically: Twitter's recommended feed, some newsletters such as Matt Rickard's, and company blogs such as OpenAI's. He suggests booking half an hour, going deep and bookmarking the rest. He also tries tools like Google's Bard and Rewind, a local app that acts as augmented memory, and deliberately reads skeptics so as to balance his own optimism.

Adams says the leadership team, including CEO Eoghan McCabe and co-founder Des Traynor, was all in on AI, which made buy-in easier. Still, people asked why, and his honest answer was an educated guess shaped by experience, much like product taste. Fergal even presented a case that, two years out, much of the current roadmap could be irrelevant, which shows how large the ambiguity is. Adams wants to share his conviction without drowning out valid alternative views.

Intercom's mantra is "adapt or die." Customer support was the obvious target, but he thinks other categories are just as ripe for a new paradigm. A reporting product like Tableau, for example, could someday be a single box where you ask who your top salesperson was last January, and project management tools could be next. For companies with large design teams built around dashboards and filters, he says it's hard to accept that the UI itself might become the box.

He closes the AI discussion with a message of reassurance. He doesn't think the answer is to announce that teams will need fewer engineers, and he points to studies showing jobs tend to move around rather than vanish. Customer support has high attrition, so any reductions would likely come from roles not being refilled. His advice is not to be afraid and to lean in.

The first framework Adams discusses is "before, after": moments such as a rebrand or a new pricing launch after which nothing is the same and you must go back out and talk to customers to learn whether you were right. He is a big believer in talking to customers and not taking answers at face value.

That led to Intercom's pricing struggles. The company had a principle of aligning price to value, but value is subjective: one customer might see 10 units of value as worth $5 while another would pay $5,000. Intercom was risk averse, accumulated too many pricing models on top of old competitive mistakes, and eventually had to make the brave decision to start again. Adams ranks pricing alongside onboarding as one of the hardest design problems he knows.

His main advice is to keep it simple. Add-ons seem harmless, but each new feature that is charged for creates customers who have it and customers who don't, then tiers and tiers within add-ons, until nobody can understand their bill. He urges fighting hard against the temptation to add extra ways of pricing.

Adams's simplified version of the Kano model says people adopt a product for two reasons: differentiation, meaning what is better in ways customers actually care about, and table stakes, the boring basics required to play. Google Wave, he says, was innovative in ways nobody cared about. Intercom spent years loving differentiation until prospects who wanted to buy couldn't because it lacked a basic report or permission feature. Today the company splits resources roughly 50/50, although it has swung 70/30 each way.

For startups entering established categories like customer support, he notes that Intercom initially sold alongside giants such as Zendesk or Salesforce Service Cloud because it lacked their table stakes. Only after years of building them could customers swap those tools out for Intercom. A startup, he concludes, must invest much more in differentiation, and the balance can shift as it matures.

His related framework, "swinging the pendulum," describes spotting an undesirable state, fixing it, and then over-correcting. Intercom did it with product, and with hiring: it brought in experienced people who imported their old companies' cultures, then over-hired specialists who struggled with ambiguity. He suspects you must swing the pendulum to learn where the boundary is. At Facebook, he saw the idea that you only know the boundary by crossing it, whereas privacy debates at Google meant a year of shipping nothing.

Adams defines the market as the people, the problems they have and how much they care, and the product as the solution. Google's social projects found lots of people with the same problem, but they didn't care enough. He adds a third element, story. Sometimes a great product in a great market fails because it's explained in a convoluted way, and he cites Rdio, which many considered an excellent product but lost to Spotify, which got the story right.

He wants every product person, from designers to researchers, to think about how to explain the product as much as how to build it. Lenny connects this to positioning and April Dunford, a recent guest, and Adams agrees that the central question is why you are better and whether you can say so clearly.

On Jobs to Be Done, Adams says Intercom followed Bob Moesta's approach and found it very useful in a simple form: it keeps the team centered on a real customer problem and, crucially, on how much energy people have around it. He dismisses debates over the nuances of competing versions, because customers don't care. Intercom still uses the four forces, which cover attraction, habits and anxieties when people switch, and he jokes that Intercom's founders used them on him over a few beers.

In the lightning round, Adams recommends two very different books: Paul Arden's It's Not How Good You Are, It's How Good You Want to Be, which can be read in 20 minutes, and Ray Dalio's Principles. His recent favorite show is The Bear, which he says celebrates the grind. He is excited about ChatGPT Vision and Rewind, which he describes as augmented memory.

His favorite interview tip is to ask references: "What feedback will I be giving this person in their first performance review?" He likes that it can't be dodged and is revealing. His life mottos are on a post-it stuck over the bottom of his monitor: only work on what matters most, and stop worrying about things you can't control. They help him avoid sending the angry emails he used to fire off.

The most valuable lesson from his parents, he says, is to be nice, because you never know what is going on in someone's life. He adds that firing people, which he did not do before Intercom, has taught him that the harder thing is often the nicer thing, and that people are often relieved. He ends by inviting listeners, especially skeptics of his AI views, to contact him.