If You’re Still Building an “AI Product,” You’ve Already Lost the Next Three Years

Founders and product leaders ask me some version of it every week, across the banks, retailers, and enterprises I advise on AI: “How do we build an AI product?” It’s a category error, and the longer a company keeps asking it, the further behind it falls without ever noticing the ground shifting underneath. Apple just spent ninety minutes showing why — and the most instructive part wasn’t what it announced. It was what it pointedly refused to announce.

1. AI is infrastructure, not a destination

Apple did not walk on stage and unveil an AI app. What it unveiled instead was intelligence dissolved into everything that was already there — woven, quietly and almost invisibly, into Siri, Photos, Messages, Mail, Safari, Calendar, the Phone app, even the screenshot tool on the Mac. There was no new icon to tap, no separate product to download, no destination to navigate toward. In one demo, you could screenshot a festival schedule, let the system read it, and have the performances you cared about land in your calendar without your ever opening a calendar; in another, the Phone app pulled context from your Mail and Messages mid-call, surfacing what you needed precisely when you needed it.

Read the trade coverage closely and you’ll notice something even sharper: Apple deliberately under-sold the whole thing. The keynote led with fixes before features — 70% faster photo loading, 80% faster AirDrop transfers, a steadier operating system — and framed a dramatically better Siri as one entry on a long list of improvements rather than the headline act. That restraint wasn’t timidity. It was the entire thesis, expressed structurally. A company that genuinely believed AI was a destination would have built a cathedral around it. Apple buried it in the plumbing.

This is not a new pattern; it is the oldest one we have

In 2000, the companies that treated the internet as a place you go — the portals, the web-only storefronts — were quietly dismantled by the ones that treated it as infrastructure that made everything else better. Amazon never built “an internet store.” It rebuilt the entire logic of retail with the internet running underneath, load-bearing and unseen. A decade later, mobile wasn’t won by whoever shipped the most apps; it was won by the companies that redesigned their core product around what a camera, a location signal, and an always-on connection suddenly made possible.

AI is that shift again, only faster and with more at stake. The winners will not be the companies with the most impressive AI feature. They will be the companies that have embedded intelligence so deeply into the product that pulling it out would cause the product to stop working — the way pulling the internet out of Amazon would not leave you with a slightly worse Amazon, but with no Amazon at all.

So the question was never “how do we build an AI product?” The question is: what would every part of this product become if intelligence were native to it — assumed, ambient, foundational — rather than bolted on as a feature someone has to find?

Most teams still underestimate how far this ripples into pricing, onboarding, even the permission model — long before anyone touches a line of code.

2. The interface is changing, and most product teams haven’t noticed

For thirty years, software has obeyed a single grammar: you open an app, work through it, and arrive at a result. You navigate to the right screen, locate the thing you need, and take the action yourself. Every menu, every onboarding flow, every painstakingly debated piece of information architecture exists to serve that one sequence.

The new Siri — and, more broadly, the entire generation of agentic systems now coming into view — quietly rewrites that grammar into something else: you ask an agent, the agent orchestrates the apps, and the result comes back.

The agent becomes the primary surface. The apps recede into services it orchestrates in the background, invisibly, on your behalf. This is not, fundamentally, a Siri story. It’s an architectural one, and it’s bigger than any single company. OpenAI’s ChatGPT Agent, Google’s Gemini Agent, Anthropic’s Claude operating a computer on its own — all of them point at the same horizon, where the next generation of software is orchestrated by agents rather than navigated by users.

And here is the tell, the detail that should settle any remaining doubt: the standalone versions of these agents have already begun to die. OpenAI’s Operator and Google’s Project Mariner — both launched as separate, standalone agent products — were quietly shut down within the past year and absorbed back into the core assistants, ChatGPT and Gemini. The destination didn’t survive; the infrastructure did. The agent was never meant to be a place you go. It was always meant to be the layer through which everything else happens.

When the agent is the interface, almost everything product teams have spent careers optimizing begins to lose weight. Navigation matters less. Information architecture matters less. The lovingly crafted onboarding funnel matters less. What rises in its place are two questions that most roadmaps don’t yet have a column for: Does the agent have enough context to act on the user’s behalf? And does the user trust it enough to let it?

I see the cost of ignoring this most vividly inside large, traditional enterprises — the banks, the retailers, the businesses that built themselves long before “digital” was a department and now spend much of my advising time trying to catch up. They will pour a quarter’s worth of design budget into restructuring a navigation menu — refining the taxonomy of screens a customer will, within a few years, simply stop visiting — because that customer will be asking an agent to move the money, reorder the supplies, or settle the account, and the agent will not care how elegantly the menu was organized. The honest, uncomfortable question every product team should be putting to itself right now is this: are you designing for the world in which users navigate, or the one in which agents act? Because you are almost certainly designing for the first, and the second is already arriving.

3. The new moat isn’t the model — it’s context

Here is the single most important strategic idea I’ve carried out of the past six months, and I’ll state it as plainly as I can:

AI alone is a commodity. AI plus context is a product. AI plus context plus action is a moat.

Let me make that concrete, because the distinction is where fortunes will be made and lost.

AI alone — a generic large language model answering questions — is available to literally everyone, for pennies per query. There is no differentiation to be found there; at the task level, the frontier models are roughly interchangeable, and they are getting cheaper and more interchangeable by the month. Building a business on the raw model is building a business on tap water.

The constraint was never really the model. Every team wiring an agent into real tools runs into the same wall almost immediately — the model reasons just fine, but the product hasn't decided what the agent is allowed to touch, in what order, or with how much oversight.

AI plus context is where products begin to diverge from one another. It’s precisely what Apple is constructing with Apple Intelligence: a system whose value lies not in the model but in what the model knows about you — your relationships, your history, your habits, the contents of your week. Glean does the same thing inside the enterprise, where the differentiator was never the LLM but the index of a company’s own scattered, hard-won internal knowledge. Harvey does it for law firms, wrapping a generic model around legal knowledge no competitor can conjure overnight.

We saw the same pattern building a health app from the ground up: the model was almost incidental once the product had to decide what it actually knew about a given patient, and when it was allowed to act on that.

AI plus context plus action is where the genuine moats get dug. This is the moment the system stops merely answering and starts doing — finding the email, booking the meeting, filing the document, reconciling the account, sending the reply with enough earned confidence that you no longer feel the need to inspect every step. And here is the part competitors cannot shortcut: context compounds. Every interaction makes the system a little more useful, a little more accurate, a little more trusted — and that accumulating trust, built one correct action at a time, is the asset a rival starting from zero simply cannot copy, no matter how good their model is.

So the question worth taking into your next strategy session isn’t which model you should license. It’s this: what do you know about your users, and what can you reliably do on their behalf, that a competitor starting today could not easily replicate? That is the only place your AI investment will ever truly compound.

What this actually means on Monday morning

Take these three ideas seriously — AI as infrastructure, the agent as interface, context as the moat — and they collapse into a short, demanding set of priorities.

Stop building AI features. Start rebuilding your core value proposition around intelligence. The right question is not “where does AI fit into what we already have?” but “what would this product look like if we’d designed it, from the first line, assuming intelligence was always going to be native to it?”

Design for intent, not navigation. As agents become the interface, the products that win will be the ones that understand what a person is trying to accomplish — not merely what they happened to click.

Know your user — and mean it. Not surveillance, not indiscriminate data hoarding, but meaningful, structured context: what the system should remember, what it should be trusted to act on, how memory should compound into value over months and years. These are design decisions long before they are engineering ones, and most companies are treating them as neither.

The companies asking these questions today will be very, very hard to catch in three years. The companies still convening committees to decide whether to “add an AI chatbot” will spend those same three years wondering, with genuine bewilderment, what exactly happened to them.

The next generation of software will be context-aware, agentic, and threaded so deeply into ordinary life that we’ll stop noticing it — which is precisely the point. The winners will be the ones who designed those experiences around humans, not around models. Apple just spent ninety minutes on stage trying, in its understated way, to tell everyone that. The tragedy is how few people in the room were actually listening.