Every buyer is now asking one question, even if they never say it out loud: why am I paying for this when I can just ask the AI to do it?
That question is repricing whole companies. Since the start of 2026, public software has been hit hard on exactly this fear: Salesforce, Adobe, Intuit and ServiceNow all down 25 to 30 percent in a matter of weeks, a sell-off a16z named the SaaSpocalypse. The market has worked out something a lot of boardrooms are still catching up to. A great deal of what software sold for the last twenty years was convenience wrapped around a capability that has just become close to free.
Here is the part worth holding onto from the first paragraph: the same force doing the repricing is the best gift in a decade to anyone building something deep. I argued in why the traditional moats are dissolving that size and capital no longer protect you. This is the companion question, the one that piece does not answer. Is your own product one of the deep ones left standing, or one of the shallow ones about to be exposed? Here is how to tell.
The model does not kill products. It exposes them
For a long time, competition was bounded and sensible. You looked across the street at the companies selling roughly what you sold, and you tried to be a bit better, a bit cheaper, a bit faster to ship. That world had rules. You could see your rivals, study them, out-execute them.
It is tempting to cast the foundation model as a new kind of predator that does not play by those rules. That framing is wrong, and it lets too many products off the hook. The model is not killing anyone. It is exposing how thin a layer of logic some products wrapped around a common task. Summarisation tools, basic writing assistants, simple data extraction, first-draft design, boilerplate code generation: each model upgrade reveals that the thing was never deep to begin with. The capability was always sitting just under the surface, waiting for a general tool to reach it.
So the uncomfortable question is not "when will the model come for me." It is "how much of what I sell was ever more than a prompt away from free." That is a question you can actually answer, and answering it honestly is the whole game.
The new bar: be AI-resistant
There is a higher bar to clear now. Your product has to be AI-resistant. By that I mean it has to do something a model plus a good prompt will not do well on its own.
How long does that buy you? Not forever. I do not believe anything is permanently safe from a technology improving this fast, and I would not trust anyone who hands you a precise forecast, given that a capability can become a default feature of a general assistant on a Tuesday with no warning. My honest bet is that real depth buys you years, not months, which is long enough to build a real business and compound a real advantage while the shallow stuff is competed down to commodity margins. Treat it as a direction, not a deadline.
Here is the blunt version. If a chatbot subscription and a good prompt can replace you, you are not a product. You are a feature with a price tag. That is not an insult, it is a diagnosis, and it is recoverable. But you cannot recover from a problem you will not name.
What actually makes a product AI-resistant
The instinct, when people hear this, is to reach for the model. Add more AI. Bolt a copilot onto the side. That is precisely the wrong move, because the model is the commodity, available to you, to every competitor, and to the buyer directly. Wrapping a thin layer of your own around a capability everyone can summon is the definition of the shallow software now losing its pricing power. I have written before about why the model is rarely where the value is, and this is the sharpest version of that point.
What makes a product AI-resistant is depth the model cannot reach on its own. In my experience it comes down to four things, and the test is not whether your product touches them but whether it would survive without them.
Proprietary data the model has never seen
A general model knows the public internet. It does not know your customers' private, messy, hard-to-gather data, the kind that takes years and trust to accumulate and cannot be scraped. The model is brilliant at reasoning over information and helpless without information it has never been given. If that information is yours alone, the model needs you. A blunt test: if a competitor could reproduce your data with a weekend of scraping, it was never a moat.
Hard-won quality in a narrow domain
General-purpose tools are general. They are impressively competent at everything and genuinely excellent at almost nothing. The gap between "good enough to demo" and "good enough that a professional will stake their reputation on it" is enormous, and it is filled by people who have spent years getting the edge cases right in one specific field. That quality is not a prompt. It is a thousand small decisions earned the hard way, and the model cannot shortcut its way to them.
Real integration into how the work happens
A chatbot is a blank box waiting for instructions. Real work is not a blank box. It is a sequence of steps, approvals, handoffs, and systems that have to talk to each other in the right order at the right moment. A product woven into the actual workflow, that knows what comes before and after and lives where the work already lives, is doing something a general assistant in a separate window cannot. The deeper the integration, the higher the switching cost, and the harder it is to rip out and replace with a prompt.
Accountability a customer can actually shoulder
This is the one people get wrong, usually by overstating it. The naive version says "a model cannot be sued, so a human is always needed, so you are safe." That does not follow. Accountability attaches to the operator, not to whoever sold the software. A customer can run a commodity model in-house and have their own employee sign off, capturing the accountability without paying you a penny.
So the real moat is narrower and more interesting. In domains with real consequences, finance, healthcare, law, infrastructure, the thing a professional needs in order to put their name to a decision is not just an answer. It is the audit trail, the validated workflow, the liability-grade quality controls, and the domain guardrails that make shouldering the risk survivable. A bare model plus an in-house operator cannot cheaply assemble that. A deep vertical product can bundle it. The defensible position is not "someone must be accountable." It is "we make being accountable survivable, and a prompt cannot."
Notice that all four are about everything around the model, not the model itself. That is the same lesson the repriced incumbents are learning in public, and it gives the old build-or-buy question a third option hiding inside it: the thing you were about to buy may simply be a prompt away from free, and the thing worth building is the part that is not.
What this looks like when it works
I do not write this from the sidelines. The clearest case I have lived is an autonomous claims system I built for a European insurance brokerage. It now resolves 67 percent of cases autonomously and lifted sales KPIs by 23 percent. None of that came from the model being clever. It came from the four things above: proprietary claims data the model had never seen, quality earned by getting insurance edge cases right one painful exception at a time, integration straight into the broker's existing workflow rather than a chatbot off to one side, and a structure that let a human own the decisions that carry real consequences. Strip the model out and replace it with a newer one, and the system still stands, because the value was never the model.
The work I do now at OpenAsset is the same shape, deliberately narrow and deep, built for the architecture, engineering and construction industry on proprietary data that took years to gather and structure. I will be honest about where even that is exposed: of the four tests, integration is the one I think about most, because integration depth is the thing a fast competitor can chip away at workflow by workflow. Data and accountability are the hardest to copy. If you only have time to deepen one axis, deepen the one a rival cannot reproduce in a quarter.
It does not have to be my work to make the point. Look at the deep vertical tools your own profession refuses to give up, the ones that have your data, your edge cases, and your sign-off baked in. They are rarely the most impressive in a thirty-second demo. They are the ones still standing while the generic wrappers are repriced.
The optimistic flip side
If I stopped there, this would read as a warning, and that is not how I see it at all. The same force that exposes shallow products flips the whole board, and it flips it in favour of small teams willing to go deep.
The advantages that used to be unbeatable were capital and headcount. You needed armies of engineers and piles of money to build anything serious, and those two things were the wall that kept newcomers out. That wall is draining away. One person with the right tools can now build what used to need a hundred. The same model that exposes the shallow product is the one that lets a tiny team build a deep one at a speed that would have been unthinkable five years ago.
So the conclusion is not "be afraid." It is "build differently." I want to be careful not to overstate it: shallow software is not suddenly worthless. Switching costs, procurement inertia and plain trust will keep plenty of thin products earning for years. But their pricing power is collapsing, and betting a new business on selling that is the losing game. Meanwhile the cost of building something genuinely deep has fallen at exactly the moment depth became the thing worth selling. I cannot remember a better time to build for someone with conviction and a narrow, hard problem to solve.
The one honest question
If you take nothing else from this, take the question I keep coming back to, and answer it honestly rather than hopefully.
If your customers could get eighty percent of what you do from a chatbot tomorrow, would they still pay you? And if the answer is yes, can you name the twenty percent only you can supply, the part built on data, depth, integration, or accountability a model cannot?
If you can name that twenty percent and point to where it lives in your product, you have something AI-resistant, and you should pour everything into making it deeper. If you cannot, you have not lost. You have just found the most important piece of work in front of you.
If you are looking at your own product and quietly unsure which side of that question you are on, that is one of the more useful things an outside pair of eyes can help you work out. I have no product to push and nothing to upsell, so if it helps to think it through with someone independent, let's talk.
Related: Traditional moats are dissolving. Size and capital no longer protect you · Why your AI spend isn't showing up in the numbers · You don't need to build a brewery to drink a pint of beer · What AI-resistant depth looks like in production
