General-purpose AI agents have stopped running ads; only two paths remain for this industry going forward.

For most of the past six months or so, I've had a pretty clear feeling: ads for general-purpose AI Agents have gone almost completely quiet, all at once.

Not long ago it was a completely different picture. Names like Coze, Manus, and Dify seemed to shove themselves in your face with every feed refresh—build an "agent" just by talking, as if everyone could hire a whole team for themselves. Now those ad buys are basically gone.

So does that mean AI Agents are finished?

I don't see it that way. My judgment is: it hasn't disappeared, it has forked. Going forward it will head in two directions—one is industry solutions, the other is private, custom AI assistants. And the second path is one the big companies can't take. I'll get to that later, because it matters far more than "which platform is easier to use."

Ads disappearing doesn't mean demand disappearing

The phase of burning cash to buy users has ended, and there are usually just two reasons: retention is bad, or the math doesn't work. Both are true of general-purpose Agents.

I don't have any industry data to back that statement up, so I can only talk about the logic as I see it: the awkward thing about a general-purpose Agent is that it can do a bit of everything, but there isn't a single thing that "must be handed to it."

Tool products survive because not doing it hurts—skip your taxes and you get fined; skip reconciliation and the money doesn't add up. What makes general-purpose agents uncomfortable is precisely that if you don't use them, nothing hurts. Users try it once for the novelty, then forget it's even running.

The uses that actually stick are all ones that save people time on one specific task. That point determines that going forward it has to move toward the narrow.

Direction one: industry solutions sell outcomes, not capabilities

Industry solutions don't sound sexy, but they're currently the only thing whose math adds up: the client pays not because you have an Agent, but because this job used to take three people and now one person plus a set of processes can finish it.

What makes this direction impossible to rush is that you have to genuinely get inside the industry—know what the consequences are if that one thing goes wrong by a step, who signs off, and who's on the hook when something breaks. Not one of these can be solved by piling on more model capability; they can only be solved by a person sitting there on-site.

And precisely because of that, it can't build hype with a launch event. However loudly you blow your own horn, the client's first question will always be "did it get last batch's data right?"

Direction two: private, custom AI assistants—the bottleneck is trust, not technology

The second direction interests me more. It also has three hard conditions, and it doesn't hold up without any one of them:

First, privacy must be absolute. This isn't something you can solve by adding a privacy clause or promising "not used for training"—the data simply cannot leave the premises; the model has to sit locally and run offline. Once it goes into the cloud, it means it's subject to audit, can be pulled on request, and internal employees can see it—which is inherently in conflict with "absolute privacy."

Second, it has to be as informed as a private butler. The owner's schedule, family finances, health, who he doesn't get along with, where his kids go to school—it has to know all of these to truly save him time. An assistant that only knows the weather and your calendar is called an alarm clock.

Third, the person delivering it has to stick around. This kind of assistant is grown, not installed; six months to a year is normal, and swapping people halfway through is the same as starting over.

This direction actually already has early forms: plenty of programmers are using AI Agents to build this kind of thing for themselves, and some are even already using it.

But that's not yet a business. Because the barrier is still technical—you have to understand models, know how to set up machines, write scripts, and debug problems yourself. Programmers can handle it; ordinary people can't. So right now it's just a little toy within the community.

The people who will really turn this into a business aren't programmers

I'm sure that in the future some big bosses will discover this thing.

Bosses make judgments differently from programmers. They look at only two things: can this thing save me time, and can it help me make fewer mistakes. An assistant that "knows everything about you" isn't a toy to him—it's a necessity.

And he's even less likely to hand it to a big company's cloud account—precisely because he has too many things he can't let outsiders see.

So he will definitely hire someone on the sly to build a bespoke setup just for himself.

The phrase "on the sly" is key: he won't post about it on Moments, won't give interviews, and certainly won't go up on a funding-launch stage to talk about it. He'll find someone he trusts, sign an agreement privately, take it slow, spend six months to a year on it, and the result will be for him alone to use.

This business will never have ads. It's a word-of-mouth business within circles—serve one boss well, and the few bosses around him will soon all want one.

This kind of order has another characteristic: price-insensitive, results-sensitive. Charge him a lot and he doesn't care; leak even a little of his business and it's completely over.

So why don't the big companies do it

Some will ask: the business is this good, why wouldn't a big company do it?

It's not that they can't technically do it—it's that they can't and don't dare to.

The business model doesn't fit. Big companies make standardized products; they only recoup the investment by selling one thing to a million people. Private customization is "one set per person"—by that measure the math never works, and it won't get past the internal project-approval stage.

They can't bear the data liability. Once the things mentioned above enter a cloud system, they're subject to audit, can be pulled on request, and internal employees can see them. Big companies don't even dare to promise the sentence "never leaves the premises," because there are simply too many people inside their systems.

Responsibility is asymmetric. Big companies answer to shareholders, regulators, and public opinion; a single leak is a monumental incident. So they can only do general capabilities, cutting risk off at the outermost layer, and will absolutely never go deep into one person's life.

And one more: people are not products. A boss wants an insider who knows him inside out and won't be replaced for years; on the big-company side it's a customer-service team on rotation, a new batch every three months. There's simply no way to deliver this.

So the conclusion is simple: big companies will build the foundation—models, compute, platforms; the layer that truly hugs one person's life is destined to be the work of small teams, or even a single person.

What would prove my judgment wrong

I need to be clear about this, because everything above is my reasoning, not established fact.

The most likely thing to overturn it is phone manufacturers: if a system-level assistant one day comes default in your phone, with data kept local, backed by the operating system, and no need to hire anyone separately, then the "hire someone to customize" path would lose a big chunk of its ground. That path still looks far off, but if it materializes, it's a crushing blow from a higher dimension.

Second: the deployment barrier for local models drops to the level of "install an app." At that point, "knows how to set up machines" is no longer scarce, and the person delivering it shifts from a technical worker to "the person who understands the owner's life"—the business still exists, but its form changes completely.

Third: a truly trustworthy private hosting arrangement emerges in compliance. Then the "must be local" wall may no longer be so solid; privacy may become a matter of a contract rather than a machine.

If any one of these three comes true, the direction in the second part has to be rewritten. My current judgment is based only on today's technology and today's cost of trust.

Back to the question at the beginning

The disappearance of ads isn't the funeral of this industry; it's more like a placement exam.

The generic, undifferentiated, anything-goes-for-everyone crowd will quietly fade; what remains are two paths — one demands that you understand the industry, and the other demands that you be trusted. Neither can be bought with funding; both can only be earned with time.

So I don't think AI Agents are dead. They've simply shifted from "telling everyone I can do everything" to "being accountable to one person for everything they do."