Will the "killer app" arrive in time?

A few days ago, while updating a course I teach on digital transformation, I came across a figure from Ben Evans that got me thinking: 900 million people use ChatGPT every week. That's a lot of millions. And yet daily use in the United States doesn't reach 15% of the adult population. Eighty percent of those who use it send fewer than a thousand requests a year. Not even three a day. Far less than any of us checks our email.

It's been almost four years since ChatGPT burst into our lives. Four years, hundreds of billions of dollars of investment, and an arms race among hyperscalers that would have made the Pentagon blush. Between us all, we have turned the words "artificial intelligence" into the wildcard of every corporate presentation that hopes to be taken seriously. And into the verbal crutch of every clueless pundit who fancies himself a visionary, even if he never gets past being a crackpot.

But four years later, when an ordinary person asks me what this generative AI thing is actually good for, I don't know what to answer. Probably because there is still no "killer app" that makes it appealing to the average mortal.

What history teaches us

Every great technology platform had its killer app. That application which turned a technology from "something interesting" into "something I can't live without."

For the Internet the answer seems obvious from today's vantage point, but it wasn't in 1996. The Internet's killer app wasn't the web. It was email. Email got millions of people to pay twenty dollars a month to AOL, CompuServe or, in our case, Terra, to connect to that strange thing called the Internet. The web was flashy: a dazzling, disorderly world of unexpected links. But email was useful. And usefulness always wins. That's why, shortly afterwards, all of us providers were "giving away" free, low-capacity email accounts that would hook users on their daily fix and leave them wanting more. Does the strategy ring a bell? It's not so different from what they do to us today with those blessed token counters, is it? Some pretentious types call it "freemium models."

What later took Internet use from massive to compulsive was the arrival of the search engine. Google turned the Internet from "something you have" into "something you use twenty times a day." That transition took between five and eight years from the launch of Netscape.

For the smartphone, which turned the network from "somewhere you went" into "somewhere you lived," the killer app wasn't the one we expected either. It wasn't the phone call (which already existed) or the mobile browser (which was dreadful). It was instant messaging and social networks: WhatsApp, Instagram, TikTok. Applications that turned the phone into an extension of our nervous system. They took three or four years from the launch of the iPhone to crystallize into a mass phenomenon.

AI's uncomfortable present

And here we are, almost four years after the ChatGPT boom, without generative AI having its equivalent, however much the pundit consensus points to three candidates.

The first is coding assistants. By far the use case with the greatest real intensity. Developers using Copilot or Claude Code report productivity gains of 30% or more. Armies of token-hungry junkies hooked on APIs that move some four billion in annual revenue in enterprise environments. But it's a niche. A huge and growing one, but a niche. However much they try to tell us otherwise, the vast majority of the world's population does not code and never will.

It's no coincidence that code is precisely where AI shines brightest. A programming language is a formal system designed to eliminate ambiguity. Each instruction has a single meaning. The syntax is finite and verifiable. The result can be evaluated in binary terms: it compiles or it doesn't, it passes the test or it doesn't. When an LLM predicts the most likely next sequence of tokens within code, the distance between "probable" and "correct" is minimal, because the language itself was designed to make it so.

The problem appears as soon as you step outside that space. Human language has ambiguities, nuances, irony, things left unsaid, and an enormous dependence on context. "Not bad" can be praise or a devastating critique depending on who says it, to whom, in what tone and with what face. I don't know if anyone still remembers, but emoticons started precisely to give some human warmth to cold email in 10-point Arial. All of that is alien to LLMs that merely predict the most likely sequence. Once we leave software development, "probable" and "correct" stop being synonyms. And the subtler the domain, such as law, medicine, negotiation or police investigation, the greater the distance. Put another way: generative AI works better the less real understanding the task requires. Which ought to give us pause about the limits of a technology we insist on calling "intelligence."

The second candidate for killer app is generative search. ChatGPT, Perplexity and Google's current "AI Overviews" are beginning to change the way we look for information. But "beginning" is the key word. Google still processes (and monetizes) more than eight billion searches a day. Generative search is nipping at its ankles; it hasn't cut its legs off.

The third, for many the most promising, is agents: AI systems capable of carrying out complete tasks autonomously, interacting with other applications or agents and making intermediate decisions. The equivalent of having an infinitely patient and reasonably competent intern who does things for you, not just tells you things. But agents, as of today, are more promise than reality. The MCP protocol that is supposed to connect them to the real world has only just been adopted as a standard, and the use cases in production can be counted on one hand. On the fingers of one ear, if we're talking about mission-critical agents in demanding enterprise environments.

A (conveniently) rigged comparison

If we go back to the 900 million weekly users and compare them with previous disruptive technologies, it's tempting to conclude that adoption is going like a rocket. But that comparison is rigged, and it's worth understanding why.

The Internet had to build its own distribution infrastructure. You needed a computer, a phone line, a modem, connection software to configure, and then you had to wait for the thing to beep and connect. I don't know if any of you cursed alongside me while swapping jumpers to produce those IRQ and COM port combinations that Windows handled so unpredictably. But every user was a conquest. Reaching four hundred million users worldwide in 2000, five years after the Netscape explosion, was a feat.

Generative AI hasn't had to build anything of the sort. It climbed onto the shoulders of the Internet, of smartphones and their app stores, and of a nearly universal digital payments infrastructure. Downloading ChatGPT takes thirty seconds. It's like comparing the expansion speed of a restaurant chain that has to build every location with that of a brand delivered through Glovo. The second one grows faster, but not because its food is better.

What makes you truly relevant isn't how many users you have, but what they use you for and how intensively. And there, generative AI is, to be honest, where the Internet was back in '97 or '98. Lots of people have tried it, but not that many use it daily. And most of those who do use it for things that aren't transformative: summarizing a text, rewriting an email, asking a question they would previously have asked Google. Almost always as personal users, without their companies having a plan much more concrete than an AI-induced nervous breakdown.

Exactly like my Terra in 2000. Millions of page views. Advertisers who didn't know what to do with them. And the killer app waiting in some Mountain View garage for Larry and Sergey to finish fine-tuning PageRank.

The question of the clock

But there is a crucial difference from that era that the title of this article is trying to point out.

When the telcos of the late nineties looked to the future, they did so from three horizons at once: the core fixed-line business, which paid the bills; mobile, which was growing at double digits; and the unknown quantity of the Internet, which devoured cash without returning a dime. The portals lost money, but fixed-line and mobile underwrote the bet. They could afford to wait for the killer app because they hadn't bet everything on it.

Today patience is an asymmetric luxury, and for the same reason. Whoever builds data centers from three horizons (advertising that already prints money, enterprise cloud that keeps growing, and the bet on AI) can afford to sit and wait. Whoever has only the unknown, whoever has bet everything on the killer app arriving before the next funding round, cannot. And among the hundreds of billions the industry is burying in infrastructure, there is much more of the latter than of the former.

I've lived through this before. Terra didn't die because the Internet was a bad idea, but because the people running it stopped believing in it. They cut investment, pulled back projects and took refuge in the core business, just when we had on the table what Yraola had intuited three years before anyone else. And by voluntarily giving up on innovation, the telcos condemned themselves to being "dumb pipes": dumb plumbing that bore the cost of extremely expensive networks while others kept the promised land flowing on top. Stopping the clock wasn't a misfortune that befell them. It was a decision.

What matters isn't the content, but the context

One day, in the thick of the battle for page-view supremacy, I called the team to a two-day meeting to understand what was happening with all that traffic that wasn't generating value. Jaime de Yraolagoitia (six foot three, a degree in Pure Philosophy, and the most intuitive mind I have ever seen) spent two entire days pacing around the table like a caged beast. He didn't stop. He was driving us crazy. On that second morning I was about to get up and strangle him. But right at that moment he sat down, opened his notebook, jotted something down quickly and stood up again. I stopped him and said: "Yraola, my dear, could you tell me what on earth you've just written down?"

His answer left me cold: "Look, Julián, what matters isn't the content, but the context."

Twenty-five years have gone by. But that sentence describes the present of AI better than any Gartner report.

Today every lab is competing for the best model. They sign deals to secure exclusive access to the best content to evolve them. And the models are becoming commoditized at a frightening speed. But nobody has yet built the context that turns those models into something indispensable for the average user. AI's killer app won't be a faster chatbot or a bigger model. It will be whatever someone builds around the model. Just as Google wasn't a better portal, but a radically different way of organizing what already existed.

Ben Evans reaches a similar conclusion in his presentation: "Chat is a terrible UX. General use needs apps." Chat is a terrible interface. Widespread use needs applications. Real applications, with concrete use cases, that solve problems people have every day. Not a blank screen where you have to know what to ask... and have excellent judgment to be able to assess whether the answer makes sense.

For an executive, the lesson is uncomfortable but clear: the race won't be won by whoever has the most GPUs, or by whoever rolls out AI licenses across the entire org chart. It will be won by whoever builds the context —the data, the process, the user experience— that takes AI from "something I try out" to "something I can't work without."

Yraola defined Web 2.0 three years before the concept existed. I suspect he also defined, without knowing it, what the AI industry needs to understand twenty-five years later.

Could it be that the missing killer app isn't a technology, but a context? And if so, the question is no longer when it will arrive. It becomes who will still be standing to build it.

This article is part of a series that began with "Less Wood, It's War!", and continues with "The end of the token open bar, or something more serious?", "What if the Genesis Mission were a huge mistake?", "Where's the pea?", "From packet to token, and back to square one" and "Investing in typewriters".