Why AI Hasn't Found Its Killer Use Case

Three years ago, I was running research for a major tech corp building what they called a "revolutionary AI-native phone." When I interviewed consumers, they wanted "more AI" in the product without being able to say why. Pressed on what they'd actually do with it, the answers were consistent: easier to search, easier to edit photos, and an expectation of some magic.

Nearly a trillion dollars in cumulative AI investment later, Pew Research Center's newest data shows only 4% of US adults use AI daily enough to call it a habit. OpenAI shut down Sora six months after launching it. Meta and Apple, sitting on three billion combined daily users, still haven't shipped a consumer AI moment worth what they spent on it. And the two use cases actually pulling people in — search and companionship — are the same two reasons people came online in the first place.

For anyone building consumer AI products, that's the uncomfortable finding underneath the hype: the industry hasn't invented a new relationship with technology. It's rediscovered the oldest one, and it's now racing to decide who owns the front door to it.

---

AI adoption is growing massively, but value isn't

Two 2026 surveys back the story the industry wants you to believe: AI usage has gone mainstream. Look past the headline number in each one and that story falls apart.

Pew Research Center's 2026 study found 49% of US adults have used an AI chatbot at least once, up from 33% in 2024 and 23% in 2023. Shift Browser's 2026 AI Consumer Insights Survey of 1,448 respondents found 32% of consumers now use AI daily. Those are the numbers that get repeated as proof AI already won.

They're the wrong numbers to look at. Only 4% of Pew's respondents call their AI use "almost constant" — the one figure that actually measures dependency instead of curiosity. Shift found the same split from the other side: 20% of people say they've never used AI at all, concentrated among adults 65 and older, and only 16% trust an AI answer engine "a great deal." Fifty-eight percent say an AI answer has shaped their opinion at least occasionally. Only 60% say they trust AI even "somewhat." That's curiosity with an asterisk, not the adoption the headline number implies.

I've seen this pattern before it showed up in a survey. In my 2023 device research, the people who actually changed their relationship with AI had already built a structured, tool-based life — budgeting apps, calendar stacks, project trackers — before AI arrived. AI didn't change them. It slotted into infrastructure they'd already built. Half of US adults are more worried than excited about AI in daily life, and only 10% feel the opposite. People are adopting a tool they don't trust, and on some level, they know it.

---

Meta and Apple's AI spending hasn't produced a consumer breakthrough

If any two companies should have cracked the consumer AI problem, it's the ones with the deepest distribution on earth.

Meta's 2026 AI-related capital expenditure: $115–135 billion, nearly double the prior year. Llama 4, released in April 2025, was widely panned by developers and benchmarks alike. The response was a $14.3 billion stake in Scale AI to acquihire Alexandr Wang as Chief AI Officer, followed by a nine-month rebuild culminating in Muse Spark. Employees burned 73.7 trillion tokens in roughly 30 days using AI tools internally, and Meta shut down its own usage leaderboard rather than keep publishing what that activity actually added up to. CTO Andrew Bosworth's memo to staff: "all motion is not progress." 600 jobs were cut from Meta Superintelligence Labs. Chief AI scientist Yann LeCun left.

Apple's version is quieter and arguably worse. The Siri AI overhaul, announced with Apple Intelligence in 2024, has been delayed repeatedly — targeted for 2025, pushed to iOS 26.4 in spring 2026, then slipped again. Internal testing reportedly turned up basic errors and slow responses. The new Siri ultimately runs on Google Gemini, not Apple's own models. The AI health coach quietly shrank from a standalone product to scattered Health app features.

Between them, Meta and Apple reach roughly three billion daily active users and have committed tens of billions of dollars each. Meta is still rebuilding its stack after a public misfire. Apple shipped its flagship assistant running on a rival's model. If two companies with that much money and that much distribution can't buy their way to a consumer breakthrough, don't believe anyone who tells you the answer is just a bigger model.

---

OpenAI's Sora shut down six months after launch

Even the company that defined the category tried to manufacture a second consumer hit beyond ChatGPT. Sora launched as an iOS app on September 30, 2025: one million downloads in the first week, four million by the end of October. Disney signed a roughly $1 billion content licensing deal, handing over character IP for Sora-generated video.

The economics never worked. Forbes estimated Sora's annualized burn at $5 billion-plus. OpenAI's own head of Sora called the cost structure unsustainable. Downloads spiked and faded — described by analysts as a "faddish response" that collapsed once the novelty wore off. Hollywood largely rejected it despite the Disney deal, citing IP and job-loss fears. In March 2026, OpenAI shut Sora down, citing failure to turn a profit and being "eclipsed by competitors."

What OpenAI did next tells you more than the shutdown did. Facing projected losses of ~$14 billion in 2026 (after ~$8 billion in 2025, cumulative ~$44 billion by 2029), OpenAI moved into advertising — the same business model as the incumbent it set out to disrupt. Ad revenue projections: $2.5 billion in 2026, scaling to $100 billion by 2030 — a target one analyst called "extremely aggressive." ChatGPT Plus subscribers are projected to fall 80% — from 44 million to 9 million — as OpenAI shifts to an ad-supported free tier called ChatGPT Go. The company hired Meta's VP of global ad solutions, struck partnerships with Criteo, Smartly, and StackAdapt, and reportedly hit $100 million in annualized ad revenue within six weeks of the pilot launch.

The company that told us it was building something categorically new is now selling ads inside a chat window — the same business model as the incumbent it swore to replace.

---

AI is reliable on structured tasks, unreliable on everyday ones

When you hear people say they don't trust AI, listen to them — the benchmarks back them up. Enterprises are the only ones getting real value out of this technology, and not because they trust it more. It's because their tasks are repetitive and structured.

Writing a coding function, triaging a support ticket against a known set of categories, summarizing a document against a fixed template — those are the repetitive, well-defined jobs enterprise AI is actually good at, and why the "enterprise AI is real" numbers later in this piece hold up. The moment a task gets open-ended, the performance collapses. On WebArena, the standard benchmark for exactly that kind of open-ended web task, the best agent hits 61.7% against a 78% human baseline — a gap that hasn't closed in two years. A March 2026 reliability study of 6,259 production AI agents across 4.5 million tasks, covering the full range of what those agents were actually asked to do rather than just their best-case use, found an aggregate 56.6% success rate — barely better than a coin flip. Even inside company walls, where tasks are far more bounded than anything a consumer improvises at home, enterprise agentic systems still show a 37% gap between lab benchmark scores and real-world deployment.

A work day is already broken into discrete, repeatable tasks. A personal evening isn't. Deciding what to cook, whether to trust a stranger's review, how to work around a kid's schedule — nobody hands you a template for that, and those are exactly the ad hoc calls AI still fails on. Reliability tracks structure, not effort, and daily life doesn't come with much structure built in.

---

Searching & companionship — the top use cases

What if all the hype about people loving AI and ChatGPT is just noise that's covering the same old pattern that's been around since the internet started: people use the web to search and find companionship.

Start with search, since it's the easiest to prove. 37% of consumers now start searches with an AI tool, and 60% read AI-generated summaries instead of clicking through. That's the same behavior people have had since 1995, just typed into a different box.

The second use case pulling people in is even older: companionship. Harvard Business Review's 2025 ranking of the top 100 gen-AI use cases puts therapy and companionship at #1, up from second place the year before. The broader theme it belongs to — personal and emotional support — grew from 17% of reported use in 2024 to 31% in 2025. People aren't asking AI to reinvent their lives. They're asking it to look things up, and to listen.

Sit with the irony of that for a second. Investors poured hundreds of billions of dollars into a technology sold as a categorical leap in how humans relate to computers, and the two things people actually do with it are the two things that built the internet in the first place — a very expensive way back to 1995.

---

Google recognized the threat to their search monopoly

Google understood the stakes here before almost anyone else. When ChatGPT launched in November 2022, Google declared an internal "code red" within weeks, reassigning teams across the company because a chatbot threatened to replace the search box that funds its entire business. Three years later, the fear ran the other direction: as Gemini closed the gap, Sam Altman declared his own "code red" at OpenAI, pulling resources off ads, shopping, and a personalized assistant to defend ChatGPT's position. Both code reds were about the same thing: whoever owns the front door to the internet owns the business. It's the same fight that played out in the late 1990s, when portals like Yahoo and AOL controlled the entry point to the web until dedicated search engines proved they could do the one thing users actually wanted, faster.

None of that means AI has nothing new to offer. Where it shows a real, measurable edge is coaching people through decisions and personalizing how they learn. A 2025 randomized controlled trial published in Nature's Scientific Reports found a purpose-built AI tutor outperformed in-class active learning by an effect size of 0.73 to 1.3 standard deviations — large enough that it's one of the more credible "this technology does something genuinely new" results in the field. Scenario planning, personalized coaching, adaptive learning paths — that's a use case a search box or a portal never could have delivered.

That strength and the companionship data above are the same finding wearing two different outfits, and the second outfit should worry you more than the first one excites you. A coach that adapts to you is valuable. A companion optimized the way social platforms optimized for engagement is not — it's the same extraction model social media ran, upgraded with a system that can hold a conversation. AI's most promising consumer use case and its most dangerous one are the same use case. Which one wins depends entirely on who builds it and what they're optimizing for.

---

AI Overviews are cutting the traffic Google used to send publishers

There's a second half to the "better search" story, and Google would rather you not connect it to AI: the company is using the technology to keep users on its own results page instead of sending them anywhere else — the exact traffic flow that built the open web.

An Ahrefs study of 300,000 keywords comparing December 2023 to December 2025 found top-position click-through rates on AI-Overview-triggering queries fell from 7.3% to 1.6% — a 58% drop tied to AI Overviews after controlling for underlying trends. Zero-click rates hit 80–83% specifically on queries where an AI Overview appears.

The casualties are named and countable. Chegg reported a 49% decline in non-subscriber traffic from January 2024 to January 2025, cited in its antitrust filing against Google. Business Insider's organic traffic fell 55% from 2022 to 2025, and the company cut 21% of its staff. HuffPost lost roughly half its search referrals. The New York Times saw search's share of its site traffic fall from 44% to 37%. Chartbeat and Reuters Institute data show global publisher referral traffic from Google down a third in 2025, with US publishers down 38%.

Google built its advertising empire by sending traffic to sites like these for two decades. Penske Media's antitrust filing accuses the company of now using AI to keep the answer and discard the source, calling it "cannibalizing" publisher traffic. Google calls it a lawful product improvement. Call it what the numbers actually show: a transfer of value from the people who built the web to the company standing between them and their readers.

---

Enterprise AI adoption is real, but narrower than the headlines suggest

Enterprise AI has the stronger numbers, and the industry leans on them hard. McKinsey's State of AI survey reports 88% of organizations regularly use AI in at least one business function, with 72% using generative AI specifically, up from 33% in 2024.

Look underneath and the picture gets a lot less impressive. Nearly two-thirds of those organizations haven't started scaling AI enterprise-wide. Only about 6% qualify as McKinsey's "AI high performers" — attributing more than 5% of EBIT to AI impact. The PwC Global CEO Survey found 56% of CEOs reported zero measurable ROI from AI in the past twelve months, and MIT's GenAI Divide research found 95% of generative AI pilots never make it past the experimental phase.

Coding assistance, support triage, document synthesis — those are genuinely transformed. Everything else stays narrow. And if the best-funded, best-staffed organizations on earth can only make AI stick in a handful of workflows, don't expect an individual with none of that infrastructure to do better on their own.

---

Why the value hasn't followed the adoption

Three more factors explain the gap between how many people have tried AI and how few have changed anything about how they live because of it.

Nobody trained ordinary users the way businesses trained employees. Enterprises have a direct incentive to teach staff how to use AI well — ROI to justify, competitors to beat. Even so, Microsoft's 2025 Work Trend Index found 77% of employees already use AI at work, but fewer than a third feel confident doing so, with structured onboarding and management actively pushing adoption. 82% of enterprise leaders say they provide AI training; 59% still report a persistent skills gap. Organizations with mature AI upskilling programs are nearly twice as likely to report meaningful ROI. Training is the biggest lever on outcomes, and companies spending tens of millions on it are still getting this wrong. The average consumer gets a chat box and a tooltip.

The comparison everyone reaches for is off by seven years. When ChatGPT launched, the popular line was that we were living in 1999 — the year before the internet remade shopping, communication, and business overnight. That's wrong. This is 1992: computers that were unreliable, required specialized knowledge, and were too expensive and complicated for most people to bother with. Read the benchmark gap and the 56.6% production success rate from the section above through that lens and they stop looking like a permanent ceiling. They look like evidence the tools are early. That only means something if the industry is honest that "early" is actually where it is, instead of selling 1999 to raise the next round.

Most people's lives simply aren't structured in a way AI can act on. This is the one training doesn't fix. The people getting real value from AI already had a calendar that reflects reality, a budget tracked in a queryable system, a workflow already broken into steps across tools and APIs. AI doesn't build that structure. It orchestrates on top of what's already there. Someone whose life runs on memory, paper, and habit has nothing for an AI assistant to read, connect, or act on, no matter how well they learn to prompt it. Enterprise AI's own vendor logic makes the same point from the other direction: an assistant's value is defined by what internal systems it can plug into, and one with nothing to connect to is "a consumer tool with an enterprise price tag." If companies with dedicated IT teams and structured databases still struggle with data readiness, an individual with none of that starts even further behind. Training can teach someone to write a better prompt. It can't hand them ten years of tracked spending or a maintained calendar to point the AI at.

The trillion dollars bought capability. It didn't buy the structure most people's lives are missing, and no model release fixes that on its own.

---

How to win average consumers

That structural gap won't stay this wide forever — infrastructure catches up, models keep improving. The real question is what actually moves someone from "tried it once" to "changed my life," once the technology is ready to earn that. Between the AI-native phone research that opened this piece, the AI-native applications I've since worked on, and the AI education work I run alongside it, the same three conditions keep showing up wherever adoption actually sticks.

Eliminate a problem from our lives. The products that win aren't the ones bolting a chat window onto an existing task — they're the ones that make a recurring frustration disappear. Weather apps show what happens when you get this wrong: a decade spent wrapping meteorological uncertainty in false precision — one percentage, one icon — instead of building an interface honest about the range of what might actually happen. AI that wins consumers won't add more conversation on top of that abstraction. It will resolve the uncertainty the product has been hiding.

Demonstrate magic, don't just talk about it. Consumer attitudes toward CGI were skeptical for years — clunky, uncanny, a gimmick — until Pixar and then Avatar showed, rather than argued, that the technology had crossed a threshold. AI is still mostly making its case in press releases and benchmark charts, which convinces exactly the people who were already sold. The products that win consumers won't explain the model. They'll produce a moment good enough that the explanation becomes unnecessary.

Make us more connected. The AI that earns a permanent place in someone's life will be wired into their health, their friends, their community, and the other applications they already depend on, so the value compounds every time they use it. That's the opposite of today's dominant use case — a single-player search box with no memory of yesterday's conversation. Connection is what turns a tool into infrastructure, and infrastructure is what the small daily-dependent group already had before AI ever showed up.

That's the work I spend most of my time on now. At PH1 Research, I run behavioral evals and build the golden datasets that tell an enterprise product team whether an AI feature is actually creating value for a user, rather than just performing well in a demo — the same distinction this entire piece has been about. Teams at Microsoft, Spotify, Mozilla, The Weather Network, and Hims & Hers have used that work to find out which of their AI bets were real before their customers did. If you're trying to answer that question for your own product, reach me at [email protected].