This is the website moment, again.

Somewhere in the late 1990s, in thousands of conference rooms, a sensible executive made a sensible argument: the web was immature, the early projects were money pits, and the prudent move was to wait for the technology to settle. Every clause of that argument was true. The companies that acted on it spent the following decade being disrupted by competitors who acted on the opposite conclusion from the same facts.

The same argument is being made about agentic AI right now, and it has better evidence behind it than the 1990s version ever did. The skeptic's case is published, peer-reported, and numerically impressive. It is also, for the second time in 30 years, the wrong conclusion drawn from accurate data.

The twin warnings

Warning one: most attempts fail. MIT's Project NANDA found 95% of corporate AI pilots deliver no measurable return (MIT, 2025), and the wreckage is still accumulating, Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027 (Gartner, 2025). A board reading those two numbers together can be forgiven for reaching for the brake.

Warning two cuts the opposite way: the capability is compounding anyway. Fewer than 10% of organizations have scaled an AI deployment beyond the pilot stage (McKinsey, November 2025), which means the gap between the scaled few and everyone else is widening every quarter the technology improves. The failure data describes the median attempt. It says nothing about the trajectory of the prize.

The 1990s produced the same statistical wreckage, and the same wrong lesson was available then. The dot-com collapse vaporized most early web ventures, and a board in 2001 could cite the casualty list as proof the web was a fad, right up until the survivors of that casualty list restructured retail, media, and travel around themselves. High failure rates among early movers measured the difficulty of the attempt, not the value of the destination. They never have.

Held together, the twin warnings produce a conclusion neither produces alone: this is hard enough that most attempts fail, and consequential enough that not attempting is its own failure mode. That combination, high attempt mortality, compounding payoff for survivors, is precisely the structure the web had in 1997. It rewards disciplined entry. It punishes both recklessness and absence.

Why waiting reads as safe and isn't

Waiting feels neutral because the cost is invisible on this year's P&L. No write-off, no cancelled project in the 40%, no embarrassing pilot post-mortem, just a clean balance sheet and a competitor compounding 5 minutes at a time. The 1990s holdouts had clean balance sheets too, right up until the revenue line explained what the expense line had hidden.

The structural difference this time is what the technology takes. The web took the storefront: a company could bolt a website onto an unchanged business and survive, because the disruption stopped at the channel. Agentic systems take the operating model itself, how decisions get made, how information moves, what a unit of work costs, and an operating model cannot be bolted on in a quarter when the pressure arrives. The 95% who fail (MIT, 2025) mostly fail by trying exactly that.

That is why the waiting strategy is worse now than it was in 1997, not better. A storefront could be built late in 18 months of catch-up spending. An operating model is years of sequenced organizational change, and the clock on those years starts when the company starts, not when the board finally believes.

The decade question

Played forward 10 years, the twin warnings stop being a project-selection problem and become a survival question: whether the mid-market enterprise survives in its current form. The mid-market has no enterprise lab to absorb failed attempts and no app-store simplicity to make attempts cheap, so the segment most exposed to the compounding curve is the one with the least native capacity to climb it. The website moment punished the absent. This one will too, at the level of the operating model rather than the marketing budget.

The honest version of the decade question deserves stating without euphemism. A mid-market company that reaches 2031 with a 2021 operating model will not necessarily disappear; it will compete for capital, talent, and customers against operators whose unit of work costs a fraction of its own, and most positions in most industries do not withstand that arithmetic for long. The skeptical executive of 1998 was not wrong about the web's immaturity. He was wrong about which risk was larger, and the asymmetry between those two errors is the entire lesson of the period.

The answer to a high-mortality, high-stakes entry is neither a leap nor a pause. It is a ladder: an audit-only assessment in 4–6 weeks before any build, one function brought live in 14 weeks before any expansion, then 3–5 functions over 6–18 months (internal planning estimate, 2026), 18 months of our own operating record stand behind that sequencing (internal operating record, 2026). The 1990s offered no instrumented way in, so companies chose between gambling and sitting out. This decade offers a third option. The companies that decline all three will be remembered the way the websiteless are, and this time the obituary will cite the operating model.