What this covers

Des Traynor, co-founder and chief strategy officer of Intercom, talks with Patrick O'Shaughnessy in a conversation split roughly between AI inside a working software company and his views as an investor. The first half is concrete. He walks through Fin, Intercom's OpenAI-powered support bot, which finds the core of a question, runs vector search across docs and past conversations, and repackages the answer in the brand's tone. He argues that answer quality depends mostly on how much business knowledge the bot can reach, not on the base model, and that rollout exposed stale help docs that human agents had quietly been compensating for. Support economics are framed through the Formula of Support: inbound volume times time to solve, divided by reps.

The middle of the conversation widens out. Traynor expects chat interfaces to reshape complex enterprise software, since they help when users know their goal but not the menu path. He also raises cautions: a human in the loop pays less attention, and less expert users cannot tell when output is wrong. The Market Size Versus Simplicity Tension gets its own treatment, as does a test for entrants, the Speed Gap Test, with its three-month versus thirty-month arithmetic. An AI meeting-notes product that failed despite good output, because Zoom and G Suite already own the budget, serves as a cautionary case.

The investor half is blunter. Traynor calls many companies dead by their last rounds' valuations and urges founders with long runway to set a six-month or one-year deadline for growth. He describes a four-part test a startup's one thing must pass, and notes that Visa wins through position rather than velocity. He closes on judging time by how you will value it at fifty.

Sharpest takeaway

Des Traynor argues that AI's biggest business impact comes from deleting entire workflows rather than bolting on features, that incumbents can win if they move fast enough, and that durable software advantage comes from brand and momentum rather than code or patents.

  • Incumbents win on AI if they move fast and rebuild core assumptions; the 'how fast can they move' test decides whether an entrant has an edge.
  • Real value comes from removing whole pieces of work (end-to-end automation), not from Copilot-style add-ons.
  • Software has few durable moats, so advantage must be converted into brand position and sustained product momentum.

The claims · ranked72 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.69

Contrary to the usual pattern where new technology platforms let startups beat incumbents, in the AI context fast-moving incumbents that recognize the shift are better positioned than new entrants.

causalhigh valuecontestednovelty 3/4durability 2/4· Des Traynor

“My experience so far has been it feels different in this context and that actually fast-moving incumbents as long as they recognize it are better positioned”

0.68

Software has very few durable moats: a product's advantage is almost by definition copied, because someone can infer the database and copy the UI, and Intercom does not rely on patents or exclusive partnerships but publishes its software broadly.

factualhigh valuecontestednovelty 2/4durability 3/4· Des Traynor

“there are very few durable moats in software. So whatever it is you have, it's almost by definition going to be copied.”

0.68

Fin's answer pipeline works by identifying the core of what the user is asking, running vector searches across all gathered documents including public docs and conversation history, deciding the right answer, and then packaging it in the brand's tone for the current conversation.

definitionhigh valueestablishednovelty 2/4durability 3/4· Des Traynor

“You look for the guts of what they're actually asking, you perform some vector searches across all the docs that you've gathered”

0.63

Many companies are effectively dead by definition based on their last one or two rounds' valuations, and venture investors would rather take pennies on the dollar and let founders restart, because the 2021 market fostered inflated valuations.

forecasthigh valuecontestednovelty 2/4durability 2/4· Des Traynor

“there are a lot of companies that are dead by definition, based on their last round or their last two rounds' valuations.”

0.60

To serve a large market, software must accumulate settings, permissions, and workflow variants for each segment, so 'for a product to be big and successful, it has to get worse' for individual users, creating a tension between market size and product simplicity that requires constant engineering effort.

causalhigh valuespeaker onlynovelty 3/4durability 4/4· Des Traynor

“So in essence, for a product to be big and successful, it has to get worse. It could get worse for any individual, but better from a market capture point of view.”

0.60

Keeping a human in the loop for AI decisions erodes attention, similar to how drivers using assisted driving pay less attention because they know the car drives itself, so high-stakes use needs a higher threshold, and it is hard to define when generative AI should be excluded.

causalhigh valuespeaker onlynovelty 3/4durability 4/4· Des Traynor

“the person has their hands in the wheel, but they're not paying anywhere near amount the same of attention as they used to”

0.57

Companies facing AI disruption should ask which pieces of customer work can be removed entirely rather than adding AI features to existing products, because removing work changes the core formula of the business, whereas AI bolt-ons (lightning bolt in the UI) are a done deal that does not differentiate.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“Of all the tasks that have to happen in our product, which can we remove as much of as possible? Take away so much of what would now be seen as undifferentiated heavy lifting.”

0.57

Answering a user's question instantly rather than seven minutes later materially changes behavior: an immediate answer lets a new Asana user create a project and expand usage, whereas a delayed answer loses them to another task, so instant support has value beyond just resolving the question.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“there's a behavioral difference between a user getting an answer in 0 seconds versus in seven minutes.”

0.57

Rolling out AI support revealed that many of a company's help docs were out of date, and human support staff had been quietly compensating for them, so documentation must be fixed before AI is allowed to answer.

factualhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“all our help docs are out of date and we didn't realize it, but our customer support team were saving our ass.”

0.57

The key question for assessing an incumbent threat is how fast the entrant can move relative to the incumbent: if the entrant builds the AI features in three months while the incumbent needs 30 months, the 27-month gap must be enough for the entrant to rebuild every other feature to the same standard, otherwise the entrant has no play.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“The question is, is 27 months enough for me and you to go and build every other feature in Mailchimp to the same standard. And if it's not, then we're goosed”

0.57

The single biggest variable in how good a customized AI support bot is comes from how much business knowledge is fed to it and how much access it has to product docs, past conversations, and user data, rather than from the base model itself.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“The knowledge sources are the single biggest variant. When we see people with like amazing Fin instances and like weaker Fin instances, it's literally how much have you fed it”

0.57

Easily copied AI features are not a moat: using an LLM over a large information source is easy, so the defensible positions are a proprietary data store or the specific workflows built on top of the model's interpretation of it, while the hard part is making the output useful enough to displace work.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“what is proprietary and easy to differentiate on is either your data store or the actual workflows that you build on top of the LLM's interpretation of the data.”

0.57

Expanding access to AI-driven tools creates a risk that users will not identify mistakes, because the less expert a user is, the less able they are to know what the correct output should look like, so the error goes undetected.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“because you've unlocked a wider addressable market, the person will not identify the mistake.”

0.57

Chat-driven interfaces will have a massive impact on complex enterprise software, because they excel when the user knows what they want to accomplish but does not know how to achieve it inside a tool with many menus, tabs, and drop-downs.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“when the user knows what they want to do, but they don't know how to achieve it, that's when chat UI is really going to take off.”

0.57

An AI meeting-notes product that produced excellent highlights could still fail because buyers will not pay a per-head fee on top of Zoom and G Suite, which already own the route to market and the budget, so distribution and the propensity to pay matter more than product quality.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“The problem is no wants to pay for that on top of Hangouts and on top of Zoom. Full stop.”

0.57

Sustainable advantage requires converting product credibility into brand position, so customers ask 'why wouldn't you use Intercom,' supported by community, advocacy, integrations, and partnerships, because feature leads eventually get copied.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“the goal as you're doing this is to transfer the credibility that exists provably in your software into the brand”

0.57

Running faster than competitors is necessary but may not be sufficient, since companies like Visa win with bureaucracy and no velocity, implying that the ultimate moat is a position so hard to copy that even fast, well-funded copiers cannot match it.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“do you think it's necessary but not sufficient? And ultimately, you need to get yourself in a Visa-like position”

0.57

Founders whose companies have 74 months of runway after a cut should set a time limit of six months or a year to show growth and otherwise wind down, because staying in a non-growing company wastes the best productive years of their career.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“I think you should set yourself a time limit of like six months or a year to get this thing growing. And if it doesn't, wrap it up.”

0.57

To make a dominant product position durable, a company must transfer the credibility of its software into a brand, through community, advocacy, rewarded power users, and technical integrations, with the sequence being best product first, then brand, then broad hearts and minds.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“job one is to honestly have the best product and the best way to do that is to move very really fast. Job two is to transfer it out into being a brand position rather than a technical position.”

0.57

A new entrant has an advantage over an incumbent only if the way the product is architected is substantially different from the incumbent's, meaning the incumbent's fundamental code assumptions are now invalid; if they are not, the entrant mostly rebuilds the same 80% of the existing stack and gains nothing.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Des Traynor

“Are there some fundamental assumptions made in the code base of Mailchimp that are now entirely invalidated? Do they just make no sense anymore?”

0.56

A startup must pass a four-part test before Traynor takes a call: its one thing must be unique to the company, valuable to the user, true, and simple to explain, because most startups fail at least one, and failing means the company is unlikely to become a large outcome.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Des Traynor

“what is the single thing you can say that is unique to you, valuable to your user, true and simple. And all four things matter and most people fail on one of the four.”

0.56

A good idea will become commonplace, so an entrant that is outpaced in speed or quality will lose, and first-mover advantage means little unless the first mover has built a brand or can keep shipping features faster than others.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Des Traynor

“If your idea is any good, you should assume it will become commonplace and everyone will have that idea”

0.56

The right way to judge how to spend time is by the future expected value it yields, such as imagining how you will value today's activities at age 50 or double your current age, because people overweight the present and regret wasted 20s.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Des Traynor

“You have to measure the ROI of the time you spend based on the future expected value of it.”

0.53

Support costs follow the formula of inbound volume times time to solve a conversation divided by number of reps, so cutting a large share of any one of those inputs, especially through AI removing work, changes the economics of the business entirely.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Des Traynor

“the formula of support, which is number of inbound, times the amount of time it takes to solve a conversation, divided by amount of support reps”

0.53

An LLM-powered ad management tool that continuously generates, A/B tests, and runs ad variants without humans logging in would make the incumbent's dashboards, reports, and configuration funnels unnecessary, so the entrant holds all the advantage in that category.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Des Traynor

“The AI is going to do it all for you. So in that world, if we're going to go after a space where, hey, we would actually build this thing substantially different were we building it today, I think all the advantage is within the entrant.”

0.53

The long-term goal is for support teams to answer any common query once, because a generated-snippets feature will learn from every live conversation and encode the knowledge so the same question never needs to be answered by a human again, freeing staff for proactive, high-value conversations.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Des Traynor

“Our vision for that will be that your customer support team deals with any common query they see, they should see it for the first time and the last time.”

0.52

Consuming information and answering questions from it is only a feature; a product also needs reporting, accuracy, collaboration, and other surrounding capabilities that make a business willing to adopt it, and that product layer is the real startup challenge.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“The ability to consume information and spit back some stuff is the feature, the product around it is reporting, accuracy, collaboration”

0.52

The largest barrier to AI adoption in customer support is people: a single interface change can force retraining of hundreds of staff, so customers adopt slowly, starting Fin only for weekends or narrow question types, which makes the rollout slower than the technology would allow.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“The largest barrier is the people aspect.”

0.52

Intercom's biggest advantage in responding to generative AI was the speed of its decisions: CEO Eoghan pushed to proceed immediately, abandoning multiple projects and roadmaps within about 14 hours of ChatGPT's release instead of running a multi-week investigation and waiting for data.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“I think that is probably the single biggest unlock that we were able to fire the starter pistol that quickly.”

0.52

Copilot-style AI features like GitHub Copilot are useful only when the task occurs often and takes significant time, so the value is frequency multiplied by time saved, and many startups default to Copilots because they are the easiest thing to build rather than the most valuable.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“if you just multiply the two things at frequency times time spent. You'll actually get a sense of where the value is.”

0.52

Intercom's customers choose it because it consistently has the latest and best features, and this persistent positioning, combined with rapid shipping, makes the company the default first provider of new capabilities like AI, which is why customers were not surprised when it launched AI.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“what our customers know is that if you want the latest and greatest, Intercom is the company that has it.”

0.52

Software is highly perishable: what was best-in-class project management three years ago is nowhere near good enough today, so a software company must keep reinvesting to maintain its position, which few non-software categories require.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“software has a really -- it's very perishable. What it took to be best-in-class in project management even just three short years ago is nowhere near good enough for today's standards.”

0.51

If someone builds a Mailchimp competitor that just adds AI for copy and design, about 80% of the existing technology still stands, since deliverability, link attribution, email rendering across clients, and brand credibility all still must be built, and the AI adds only about 20% new work.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Des Traynor

“So in that world, I think you might be in some 80-20, 80% of the existing tech still stands and 20% is going to be new stuff coming from OpenAI”

0.51

Many AI bots built on GPT-3.5 Turbo look identical when tested with obvious questions, but differences in quality become visible only in nuanced cases such as refund requests or attempts to push the bot off topic.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Des Traynor

“if you just test them all by asking a really obvious question and getting a really obvious answer, you're going to conclude they're all the same bot.”

0.50

After launching Fin, an OpenAI-powered support bot, Intercom customers have seen 50% of support volume drop in some cases, and on average customers who switch it on see 15%, 20%, or 25% of their support volume disappear immediately.

factualhigh valuespeaker onlynovelty 3/4durability 1/4· Des Traynor

“We've seen customers see 50% of their support volume drop. We've seen, on average, most customers who turn Fin on would know what would work, literally clicking an on button are getting 15%, 20%, 25% of their support volume just going away, straightaway.”

0.50

Intercom's strategy is to keep releasing features so fast that copiers are always shipping yesterday's technology, a 'gingerbread man' approach in which competitors can copy but never catch the company.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“Run. Run fast as you can. You can't catch me.”

0.50

An AI-based product still has to be built as brilliant software with strong integrations, platform, and UI, because if it only pings text to OpenAI with a clever prompt, that prompt is not a durable advantage since others can find it too.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“It can't be that you're the only person who's worked out the right prompt.”

0.50

After the Head of ML concluded ChatGPT was the biggest AI transformation of his career, Intercom recognized that its ability to summarize, extract, and propose answers mapped directly onto customer support, so the support category was going to change.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“It can learn, it can summarize, it can extract the main pieces. And given a query and given a set of information, it can actually propose an answer. And that is basically what customer support is.”

0.49

Human-led support is usually more expensive than $0.99 per resolution because most support reps are paid around $8-10 an hour at the floor, and time to onboard and train agents adds cost, though some low-value conversations may not justify the price.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“in general, we see most support reps paid somewhere between like -- the floor here would be, I don't know, $8, $9, $10 an hour”

0.49

The future of customer service will involve both humans and bots, and the key design goal is a flywheel in which humans train and improve the bots while bots relieve and augment the humans, which requires high-quality help desk software for anyone competing in this market.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“we have this idea of a flywheel where the humans help the bots and the bots help the humans.”

0.49

Intercom will switch AI providers based on who gives it access to the best technology, and reasons to change include another provider's superior tech, regional availability such as EU data residency, or price, rather than loyalty to OpenAI.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“So I think we're going to partner with whoever we think is going to give us most access to the best tech.”

0.49

Most AI investment opportunities Traynor sees are simple API calls that summarize text or change its tone, which he considers basically dead in the water, while more powerful products involve ingesting massive data and offering natural-language querying, such as Rewind.ai.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“the majority of what I see is not amazing. It's very much just we also know how to summarize a piece of text or change the tone of a piece of text”

0.49

For high-stakes topics like refunds or account lockouts, Intercom uses a separate deterministic product, Custom Answers, that matches queries with fuzzy logic to fixed scripts, because it cannot afford generative answers and must control every word.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“we actually can't afford to have generative conclusions. We actually need to control every word that it said.”

0.49

With the right amount of prompting, a GPT-4-based support agent can be constrained to stay on topic, refuse political opinions or competitor recommendations, and curtail most hallucinations, so it does not do anything a human support rep would not be allowed to do.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“we can constrain this agent to not go off-topic, to not take opinions about whatever politics or to not recommend your competitors”

0.47

Early in Traynor's investing career he was swayed by great-looking products and landing pages, but beautiful UI became commoditized as design tools improved, so a polished product no longer signals a viable business.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“Beautiful UI and stunning landing pages became more commoditized or like maybe more easy to do for people.”

0.47

All SaaS is essentially UI on top of databases, so almost everything is copyable quickly, and the goal is therefore to build a position that looks daunting to copy, such as a project tool whose copy would take years of roadmap to match.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Des Traynor

“all SaaS is basically UI on top of databases. So everything is copyable pretty quickly.”

0.45

Fin charges $0.99 per resolution because that is roughly what OpenAI charges for GPT-4 usage, and high model costs make some features, such as summarizing every conversation in real time across roughly 500 to 600 million monthly conversations, prohibitively expensive to build until cheaper models arrive.

causalhigh valuespeaker onlynovelty 2/4durability 1/4· Des Traynor

“We charge $0.99 a resolution. It's way cheaper than you pay a human, but way more than you'd guess”

0.44

Intercom customizes Fin mainly by relying on OpenAI for judgment and conversational ability, then performing vector searches across gathered documents and the conversation thread to find the right answer and package it in the brand's tone, rather than building the conversational layer itself.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Des Traynor

“We lean on OpenAI mostly to power Fin's decision-making ultimately like its judgment and also its conversational capability.”

0.38

The first challenge in deploying a customer-facing AI bot is restricting it to the business's own domain so it refuses unrelated questions, such as politics, since answering off-topic queries creates a slippery slope toward taking opinions the company does not want to take.

normativeestablishednovelty 1/4durability 2/4· Des Traynor

“if you go to any Fin instance and ask like, who is the President of America, it actually won't attempt to answer it because that's a slippery slope.”

0.34

The norm for companies facing a new technology is to form a small cross-functional team, wait three or four weeks for a readout, and finish the quarter's roadmap, which is too slow to respond to a shift like generative AI.

normativespeaker onlynovelty 2/4durability 3/4· Des Traynor

“Let's form a small cross-functional team and we'll do an investigation, we'll get a readout from them in three or four weeks”

0.34

Products that stayed simple, such as Craigslist, dodged many pitfalls by avoiding VC-driven growth pressure, while Apple's Notes app and Bear each serve nearly everyone with one product used in roughly the same way.

factualspeaker onlynovelty 2/4durability 3/4· Des Traynor

“They dodged a lot of bullets by keeping it simple.”

0.33

Traynor's mother bought him an Amiga 500 computer for his ninth birthday in 1980s Dublin despite financial hardship, and that gift set off a chain of learning, computer science, the blog, meeting Intercom's CEO Eoghan, and meeting his wife, so his whole career traces back to that single machine.

factualspeaker onlynovelty 1/4durability 4/4· Des Traynor

“The entire history of my career goes back to that one machine.”

0.31

Intercom is somewhere in the middle of the spectrum between a wholly new AI product and a rebuild of the underlying platform: whole workflows are removed, but the help desk, messenger, APIs, and knowledge base still need to be built and maintained.

factualspeaker onlynovelty 2/4durability 2/4· Des Traynor

“I think we're somewhere in the middle where whole workflows are removed in Intercom's case, but not the entire platform.”

0.31

When half of support responses go through AI, Intercom had to throw away and rebuild parts of its product, such as first response time metrics, which are no longer valid, because customers now care about AI-specific reporting.

factualspeaker onlynovelty 2/4durability 2/4· Des Traynor

“Measures like first response time right now are no longer really valid because you have this tracked AI and all that stuff.”

0.31

Deep integrations and interoperability with other tools, such as Intercom working well with GitHub or Jira, make it harder for competitors to win co-promotion and partner status, forming a technical moat distinct from features.

causalspeaker onlynovelty 2/4durability 2/4· Des Traynor

“if you try and use new app with the same, it will be harder for them to get the same co-promotion, the same partner status”

0.29

A bottle of Coca-Cola is the same product for everyone from presidents to strangers, a uniformity software rarely achieves because each customer tier needs different workflows and code.

factualspeaker onlynovelty 2/4durability 3/4· Des Traynor

“This bottle of coke is the same one the Barack Obama would drink and the same one that Bill Gates would drink”

0.28

Instead of a dashboard requiring users to hunt for the most common complaints, AI can produce a summary paragraph identifying the biggest issues, which is an example of workflow displacement rather than augmentation.

normativespeaker onlynovelty 2/4durability 2/4· Des Traynor

“maybe the new version of that report has just a summary paragraph that tells you here is the biggest issues going on today”

0.27

The natural response to a gift that shaped your life is to find an equivalent action to unlock a similar path for someone else, which is a question Traynor says he thinks about often.

normativespeaker onlynovelty 1/4durability 3/4· Des Traynor

“what equivalent thing could you do to unlock a path like that for somebody else?”

0.26

The high-order category of complete automation of work is still early, and whether incumbents or startups will capture it is an open question across many industries.

forecastspeaker onlynovelty 1/4durability 2/4· Des Traynor

“The question for me is, will it come from the incumbents or not. That's still, I think, an open question in a lot of these industries.”

0.26

Products like the iOS Notes app and Bear, which serve one use case the same way for everyone, achieve dominance because the builder only needs to build one thing for one person, which gets closer to liquid profit.

causalspeaker onlynovelty 1/4durability 2/4· Des Traynor

“So there are some products that cut through and just say, 'We are one thing, and we're that one thing for everyone and you all like it.'”

0.26

Intercom's customer service manifesto rests on four beliefs: bots and humans work together, support should be both proactive and reactive, support must be conversational and omnichannel, and all the pieces must work together rather than being stitched across multiple tools.

normativespeaker onlynovelty 1/4durability 2/4· Des Traynor

“Today, our manifesto really has four core ideas in it. The first one is bots and humans will work together.”

0.26

The spreadsheet tool Equals lets users write Excel and SQL-style queries in plain language, which dramatically expands the spreadsheet's accessible market to people who would never learn Excel, a route to growing the total addressable market through chat UI.

factualspeaker onlynovelty 2/4durability 2/4· Des Traynor

“So what that does, from my point of view, that has made the spreadsheet an infinitely more accessible tool to millions of people who never were going to learn Excel.”

0.26

A high-value AI product would take a company's common SEO terms, identify unranked opportunities, write fully specified articles, generate images, and present them for criticism, replacing an entire workflow rather than suggesting an opening paragraph.

normativespeaker onlynovelty 2/4durability 2/4· Des Traynor

“Now I'm starting to think, 'Huh. This could be replacing a workflow.'”

0.26

Intercom's head start over other Fin-style bots lasted only three or four months before many YC startups launched similar products, so competitive advantage in AI support is short-lived and the battle shifts to who has the right overall platform.

forecastspeaker onlynovelty 2/4durability 2/4· Des Traynor

“It was the case for three or four months, but now there's a lot of YC start-ups doing it.”

0.24

Generative models probably will not be used directly for doctors' diagnoses, because they might speed up doctors but still require a human in the loop, and the risk of a wrong answer that sounds right is too high for critical decisions.

forecastspeaker onlynovelty 1/4durability 2/4· Des Traynor

“I do suspect this is why you won't necessarily see GPT used in doctors' diagnosis apps.”

0.24

Most customers adopt Fin cautiously, turning it on only for weekends or for narrow queries like password resets rather than the full support volume, because they want to test performance on a smaller use case first.

factualspeaker onlynovelty 1/4durability 2/4· Des Traynor

“everyone wants to dip their toe, they don't want to say, Let's point Fin at the entire support volume.”

0.21

UI quality absolutely matters, but a nice interface is not a sustainable position on its own, and maintaining a top UI requires constant investment to stay at the top.

normativespeaker onlynovelty 1/4durability 2/4· Des Traynor

“It absolutely is not the case. It's just, Des does not think neat UI is a sustainable position.”

0.21

Intercom's first AI release in December was designed to assist customer support agents rather than face end users, because the team was not yet sure the technology was good enough for customers, though it was good enough to augment agents.

normativespeaker onlynovelty 1/4durability 2/4· Des Traynor

“that was all about assisting the customer support rep because we weren't yet sure that the tech would actually be good enough to face end users”

0.21

Intercom has dabbled in AI since 2016 and had AI products live in market before ChatGPT, but ChatGPT changed both what AI could do and how willing users were to engage with a bot.

factualspeaker onlynovelty 1/4durability 2/4· Des Traynor

“We have dabbled in AI since 2016. We've had AI products live in market. I think the world changed with ChatGPT”

0.20

Intercom charges $0.99 per AI resolution, which is far cheaper than a human but more than customers expect because the underlying model is just an API call, and this price reflects what OpenAI charges.

factualspeaker onlynovelty 1/4durability 1/4· Des Traynor

“We charge $0.99 a resolution. It's way cheaper than you pay a human, but way more than you'd guess”

0.18

Nobody yet knows where AI resolution rates will asymptote, because Intercom has only been live with Fin for about eight weeks, though resolution rates for customers are expected to keep rising as knowledge accumulates.

forecastspeaker onlynovelty 1/4durability 1/4· Des Traynor

“where does it start to asymptote. We've yet to see it, but we are only at about eight weeks.”

0.18

Intercom handles about 20,000 of its own support conversations a month, roughly 0.25 million a year, and about 0.5 million over two years, which feeds Fin's knowledge base.

factualspeaker onlynovelty 1/4durability 1/4· Des Traynor

“If you take Intercom, we do 20,000 support conversations a month, so about 0.25 million a year.”