The 1997 Test for Your AI Roadmap

The AI tooling market has hit the part of the cycle where every founder I talk to has bought, trialed, or canceled at least three tools in the last 90 days. The pitch decks look identical. The demos look identical. The retention curves, as far as anyone will admit on the record, look identical too. Most of it will not be here in two years.

Benedict Evans has been making the case on Lenny Rachitsky’s podcast and on his own analyst site that AI right now looks a lot more like the internet in 1997 than the internet in 2007. The infrastructure works. The capital is flowing. The applications that will actually capture the value have mostly not been built yet, and a lot of what gets called a use case today is going to look very obvious or very silly in retrospect. That framing is the most useful filter I’ve found for deciding what to buy, what to build, and what to ignore for the rest of the year.

The test, in one sentence

Would this product still matter if AI plateaued tomorrow?

That’s it. Run any AI tool, internal build, or vendor pitch through that question and most of them collapse on contact. The ones that pass tend to be doing boring, durable work (good data, good distribution, good positioning) with AI as one input among several. The ones that fail are usually a thin wrapper around a model that someone else trained, sold at a price that assumes capability keeps improving on the same curve, with no defensibility once it stops.

This is not a contrarian opinion. It is the same point Ben Thompson has been making for two years at Stratechery, going back to his “value chain” essays on where AI economics actually accrue. His read, broadly, is that the application layer wins when it owns a workflow, a dataset, or a distribution channel that the model can plug into. Everything else gets compressed by the next OpenAI or Anthropic release, often within a single quarter.

What the dot-com analogy actually says

The 1997 framing gets misread in two directions. The first misread is “this is all hype, none of it will matter.” That was wrong about the internet and it is wrong about AI. The second misread is “everything is changing all at once, place every bet now.” That was also wrong about the internet, and the people who placed every bet in 1999 mostly went out of business.

The real read is closer to this: the platform shift is real, the timeline is longer than the discourse suggests, and the companies that will be most valuable in 2034 are mostly not the ones with the loudest 2026 launch posts. A few will be. Most won’t. The ones that survive will look, in hindsight, like they were doing something obvious that no one else got around to building.

Tomasz Tunguz has been writing about this same pattern on the SaaS side, tracking how AI-native startups are pricing, retaining, and burning. His data keeps showing the same thing: the AI products with durable retention are the ones solving a specific operational problem for a specific buyer, not the ones marketing “AI-powered” as the headline feature. The headline feature is the model. The model is rented. The moat has to come from somewhere else.

Why most AI marketing tools fail the test

I spend most of my week looking at the AI tooling stack inside small B2B teams. The pattern is consistent. A founder-marketer buys a tool because it demos well. They use it for a few weeks. They cancel because the output is generic, the costs scale linearly with use, and the next model release at the foundation layer makes half the product irrelevant.

Run the 1997 test on the typical AI marketing tool and the answer is no. It does not still matter if AI plateaus, because the entire premise is that the model keeps getting better and cheaper. Strip that assumption and you have a wrapper around a feature that the foundation labs will ship natively within a year.

The tools that pass the test look different. They sit on top of proprietary data the team already had. They integrate with workflows the team is already running. They produce outcomes the team was already trying to get, faster and with less manual lift. The model is an input. The compounding capability is the data, the distribution, and the operator on top of it. The same shift is what makes the founder-marketer role newly valuable: one person with taste, owning the data and the distribution, plugging models in where they actually help.

The job-loss panic gets the same thing wrong

The other place the 1997 analogy lands hard is on the labor side. The current discourse on AI and jobs has the same texture as the late-1990s discourse on the internet and jobs. Big confident predictions. A lot of forecasts about which roles disappear. Most of them, in retrospect, wrong in both directions. The internet did destroy jobs (travel agents, classifieds, video stores) and it created categories no one had named yet (SEO, growth, community management, every job that has the word “platform” in it). Net employment went up. The mix changed.

Ethan Mollick has been making a related point on One Useful Thing: the org-level productivity story for AI is still mostly anecdotal because the systems and habits to actually capture the gains have not been built inside most companies. The capability is in the model. The compounding return is in the workflow on top of it. Companies that have not rebuilt the workflow are running model-of-the-week experiments and reporting flat productivity. The thesis predicts that, and so far the data agrees.

What this means if you’re a founder

Three practical moves if you’re trying to decide where AI dollars actually go for the rest of the year.

Stop chasing model-of-the-week. A new state-of-the-art model gets posted roughly every two weeks. None of those launches is a buying signal on their own. The question is whether your team has a workflow that gets meaningfully better when the underlying model improves, or whether you’re just running a tool that happens to have a model behind it. The first compounds. The second resets every time the vendor’s pricing page changes.

Buy capability that compounds, not novelty that depreciates. Capability that compounds: a content engine your team owns the data and voice on, a customer research workflow that captures every conversation, a pricing surface that updates from live usage data. Novelty that depreciates: a tool that does one demo-friendly thing, priced as if model costs only fall. The same logic that broke the free-to-paid funnel for pure-play AI products is the same logic that breaks novelty tooling on the buyer side. The economics assume a trajectory. If the trajectory bends, the value goes with it.

Manage to the workflow, not the model. The teams I see getting durable wins out of AI right now are the ones treating models the way a marketing org treats ad networks. Inputs to a system the team owns. Swap them out when something better ships. Keep the system. Most of the panic and most of the disappointment comes from teams that built the system around a specific model and now have to start over every quarter.

The 1997 analogy is useful because it sets the right pace of expectation. Real change. Long timeline. Most of the value goes to people who are still building, in 2032, what they were quietly working on in 2026. Mark Suster has been making a version of this argument at Upfront about venture timing for years, in a different shape. The dollars chase the demo. The returns sit with the operators who kept building after the demo cycle moved on.

If you’re trying to figure out where AI actually fits in your business and what to stop buying this quarter, the AI Fit Quiz is a five-minute reality check. Five questions, ninety seconds, a real answer on whether your roadmap passes the 1997 test or quietly fails it.

Jared Castronova is the founder of JAC Growth Marketing, where he builds AI-powered GTM systems for B2B companies.