We have been shipping production software since the mid-1990s. In that time we have lived through “the internet changes everything,” “mobile changes everything,” “the cloud changes everything,” “blockchain changes everything,” and now “AI changes everything.” Three of those were genuinely true. Two were not. AI looks more like the first three than the last two, but the hype pattern is identical, and reading it correctly matters, because the pattern is where fortunes get made and budgets get burned.
The pattern that repeats
Phase 1. Pure novelty. Everyone demos the thing. The demos look magical. Coverage is breathless. Real production deployments are rare and quiet. In 1996 that was a company putting up a brochure website and issuing a press release about it. In 2023 it was every SaaS product bolting a chat box onto the sidebar. The demo is cheap in this phase; the distance between demo and dependable is invisible to almost everyone.
Phase 2. Aspirational application. Companies announce huge initiatives. Most of them are marketing announcements with no working product behind them. A handful of teams ship real things and learn fast. This is the phase where the gap opens: the announcers are optimizing for the press cycle, the shippers are accumulating the unglamorous knowledge (what breaks, what users actually adopt, what it costs to run) that cannot be bought later at any price.
Phase 3. Disillusionment. The announced initiatives fail to materialize. Trade press starts writing “is it overhyped?” pieces. Stocks correct. People who never built anything declare the technology dead. The dot-com crash of 2000 is the canonical example, and it is worth remembering what actually happened next. The crash killed the companies that were stories, not the technology. Online commerce kept growing straight through the wreckage.
Phase 4. Boring success. The technology turns out to be exactly as transformative as the optimists claimed, on a timeline twice as long, deployed in places far less glamorous than predicted. Nobody in 2009 predicted that mobile’s biggest business impact would be logistics dispatch and field-service apps, and nobody writes headlines about the cloud anymore because it is simply how software is run. The companies that learned in phase 2 quietly dominate. It stops being news precisely when it becomes infrastructure.
Blockchain is the control case: it ran phases 1 through 3 on schedule and then never earned phase 4, because outside a few niches it was a technology in search of a problem. That is the diagnostic: the fakes die in phase 3; the real ones get boring.
Where AI sits today
As of this writing, AI is somewhere in phase 2 verging on the front edge of phase 3. The signs are familiar: huge corporate initiatives announced without a working product, a press cycle that swings from “transformative” to “bubble” inside the same week, and a small but growing population of unglamorous teams quietly putting real systems into production. We see both halves in our own client conversations: companies burned by a flashy pilot that never shipped, sitting next to companies whose AI-assisted workflow has been quietly saving them hours a day for a year.
Four questions that separate real from 1999
When a vendor, a consultant, or your own team brings you an AI promise, the pattern suggests four questions. Is there a working system or an announcement? Ask to see it run on your kind of data, not a curated demo. Does it attack a problem you already pay to solve? The durable wins automate existing cost (document processing, support triage, data entry) rather than inventing a behavior nobody asked for. What happens when it is wrong? Anyone selling AI without a clear answer about error handling, review paths, and accountability is selling phase 1 novelty. Can it be measured? If nobody can state what success looks like in numbers, you are funding a press release. These are the same filters we apply before taking on AI development work, because a project that fails them will stall no matter who builds it.
The right move at this point in the cycle
It is not to make a corporate-PR announcement, and it is not to wait for the dust to settle. Phase 4 winners are decided in phase 2 and 3, not after. The right move is to start one focused, defensible AI use case, ship it to production with proper evals and guardrails, and let the people who learned from doing the work set your direction. Small enough to survive being wrong, real enough to matter if it is right. We wrote up how a disciplined engagement actually runs (discovery, proof of concept, production, with a stop-or-continue decision at each gate) and what separates pilots that ship from pilots that stall. A first production release in 8 to 16 weeks is a realistic goal for a well-scoped use case; a company-wide transformation program announced in a press release is not.
The disillusionment phase, when it arrives in force, will be your friend if you are building. Vendors get honest, talent gets available, and the noise that makes evaluation hard right now dies down while your system keeps compounding its lead.
The internet, mobile, and the cloud all rewarded companies that shipped one real thing while the press was declaring the technology overhyped. AI will reward the same discipline. If you missed those three, you do not have to miss this one.