AI vs. the Web Revolution in Retrospect

A working analysis, May 2026. Written to answer one question: if the AI revolution rhymes with the web revolution, where are we on the curve — and what does that mean for someone about to start a company?

Preface: why this comparison earns its keep

Every technology wave gets compared to the last one, usually lazily. This comparison is worth making carefully because the web revolution is now complete — we can see the whole arc, from the 1994 consumer moment through the 2000 crash to the durable winners that emerged on the other side. That hindsight is a gift. It lets us ask not "what's exciting" but "what actually survived, and why" — and then hold the AI revolution up against that template to locate ourselves in time.

The thesis in one sentence: we are at roughly the web's 1998 — past novelty, real money moving, but before the knee — and the front of the adoption curve is moving 2–3x faster than the web did while the back of the curve is moving at the same speed or slower. Everything else in this book is an elaboration of that claim and its consequences.

Chapter 1 — The world before, so we remember what a rupture feels like

It's easy to forget how total the discontinuity was, because the result is now invisible infrastructure. Before the web, every computer system was an island. Companies ran on mainframes and minicomputers — DEC VAX, the AS/400 — accessed through dumb terminals with no graphical interface. Inter-company data moved by fax, phone, and mail; the only electronic exchange was EDI (Electronic Data Interchange), expensive enough that only large players with banking-grade EFT roots could afford it.

"Online," to the extent it existed, meant Bulletin Board Systems — geographically fragmented dial-up boards you connected to one at a time, often at 300 baud (about 30 characters per second). In 1994 there were roughly 60,000 BBSes serving about 17 million US users — bigger than CompuServe at the time. And the walled gardens — AOL, CompuServe, Prodigy — did not interconnect. You were in one garden or another.

The web's rupture was that HTTP + HTML + the URL made everything publicly addressable and linkable at once. Mosaic plus cheap dial-up arrived in 1994–95, and the BBS market crashed almost immediately; many BBS operators simply became website operators. The lesson worth carrying forward: a real revolution doesn't improve the islands, it dissolves the boundaries between them. When you're evaluating whether AI is that kind of rupture, ask what boundary it dissolves — not what task it speeds up.

Chapter 2 — The implementer: a role that exists to abolish itself

Every revolution mints a generalist "master" role to ferry organizations across the chasm. In the web era it was the webmaster — a term coined by Tim Berners-Lee. It grew out of research institutions where people built sites on top of their day jobs; the canonical example is a head librarian at SLAC who built one in her spare time. By 1996 it was a real job: a Web Week survey found 35% of respondents listed "Webmaster" as their title, against effectively none the year before.

The webmaster was a generalist — configuring servers, ensuring uptime, deploying software, updating content, managing databases, access rights, and DNS, all at once. The role was broad precisely because nobody had specialized yet. And then, within about five years, it fragmented — into web designer, web developer, SEO specialist, and sysadmin. The generalist master role dissolved the moment the work became legible enough to divide.

The cloud revolution repeated the pattern with the "cloud architect / migration partner," organized around frameworks like Gartner's migration "R"s (rehost, replatform, repurchase, refactor, retire, retain). It too partially fragmented — into DevOps, SRE, and platform engineering.

The durable lesson for the AI moment: the generalist "AI enablement" role — the one I occupy, the one Squid Club trains — is real and lucrative and has a half-life. It exists to carry organizations across a gap. Once the platforms absorb the complexity, the role specializes or dissolves. This isn't a reason to avoid it; it's a reason to treat it as a vantage point rather than a destination. (Chapter 6 turns that vantage point into a strategy.)

Chapter 3 — The adoption curve, measured honestly

The single most important discipline in this whole analysis is refusing to compare a junk number to a real one. The headline AI-adoption statistics ("88% of organizations use AI") are the junk tier — they measure "used a model in at least one function once," which is the web equivalent of "has an employee who has seen a website." Useless for locating ourselves in time.

The web's real curve (businesses with a website)

Using a consistent national series (the ABS "web presence" measure, methodologically stable across years):

Year	Businesses with a website
1998	6%
2000	16%
2001	22%
2002	24%

Then it kept climbing through the rest of the decade toward ~60%. Note the shape: a slow crawl from the 1994 consumer moment, single-to-low-double digits for years, then a bend upward around 1999–2001. Note also the gap that ran ahead of it: internet access was at 29% in 1998 when website presence was 6% — the access-far-outruns-value-creation gap that every revolution shows.

AI's real curve (agents in production — the strict definition)

The honest equivalent of "has a website" is not "uses AI" — it's agents actually deployed in production, doing real work. By that measure, roughly three years past the late-2022 consumer moment: about 11% deployed in production, ~14% scaled org-wide, with ~78–79% stuck in pilots. The pilot-to-production failure rate is brutal — for every 33 pilots, about 4 make it out; ~88% never graduate.

The two curves, indexed to year zero

Web year-zero ≈ 1994 (Netscape). AI year-zero ≈ 2022 (ChatGPT). Lay agents-in-production over the website curve and today maps to roughly the web's 1998–99 — sitting on the flat shelf before the knee.

But the two ends of the funnel are moving at different speeds, and this is the crux of the whole book:

The front of the funnel (awareness, individual usage, pilots) is moving 2–3x faster than the web did. Individual human usage is already ~37% and climbing toward what would have been the web's eventual plateau.
The back of the funnel (pilot → production) is moving at the same speed or slower, because the hard part — reliability, memory, governance, integration — is genuinely hard and does not compress just because the models got good.

So the felt experience of "AI is everywhere, incredibly fast" is true — about the top of the funnel. The thing that actually defines the revolution's maturity is still at web-1998 levels and climbing the slow part.

Chapter 4 — Which startups survived, and the exact shape of survival

This is where hindsight pays the most. Look at the small-and-midsize startups of the web era — the picks-and-shovels companies, not the doomed dot-com retailers — and a precise survival template emerges.

Vignette (CMS, founded Nov 1995). Got its break by partnering with CNET, which had built an internal tool called PRISM for database-driven publishing. CNET transferred the technology plus $500K in cash for a 33% stake. Vignette productized an internal tool, went enterprise, and was ultimately acquired by Open Text in 2009. The move: someone's internal tool became the product.

NetObjects (Fusion, founded 1995). Held the first website-builder patent; IBM invested $100M for a majority in 1997 at a $150M valuation. Revenue ~$34M by 2000. Then it got squeezed (see below) and faded, re-established small in 2009. The cautionary half of the template.

Akamai (CDN). Born from a challenge Tim Berners-Lee posed at MIT in early 1995 about network congestion. Pure deep infrastructure. IPO'd 1999; still a backbone of the web. The deep-infra winner.

Speedera (CDN, founded 1999). The first CDN to turn a profit — and was acquired by Akamai in 2005. The reliable exit: get bought by the category king.

Macromedia / Dreamweaver (1997). Built Dreamweaver around the professional, not the novice ("Roundtrip HTML" respected hand-coders). Sold to Adobe in 2005 as part of a $3.4B deal; Dreamweaver still ships in 2025. The professional-grade tool that kept deepening.

FrontPage (1995), the anti-pattern. Microsoft acquired Vermeer in Jan 1996 to get it, bundled it in Office, and discontinued it by 2006. It was the "make it easy for novices" WYSIWYG tool that won the first wave and died in the second.

The killer dynamic that decided who lived

There's a specific mechanism that crushed the middle, and it's the most transferable thing in this chapter. Once the upstarts proved what was possible, Adobe (PageMill) and Microsoft (FrontPage) entered with consumer-grade tools they could run as loss leaders to a larger product suite — driving the market price toward zero. Vignette and NetObjects were forced to either specialize upmarket into the enterprise or get commoditized to death competing with free. The market outside the giants fractured and fled to niches.

The survival recipe, distilled

The durable winners sat at two extremes, never the middle:

Own the deep infrastructure the giants won't or can't build (Akamai).
Build the professional-grade tool that keeps deepening and respects the expert (Dreamweaver) — never the novice WYSIWYG abstraction (FrontPage).
Go enterprise before the loss-leaders commoditize your category.
The reliable exit is acquisition by the platform/category king (Vermeer→Microsoft, Speedera→Akamai, Vignette→Open Text).

The middle — WYSIWYG abstractions, point tools, "easy for everyone" — either died or got absorbed.

Chapter 5 — The web-services consultancies: paper fortunes, structural fragility

Now the closest mirror to "doing the revolution for established companies" — the digital integrators. This is the most important chapter for anyone whose current income comes from helping other companies adopt the new thing.

In the boom they looked spectacular. At its fall-2000 peak, Scient had $100M in quarterly revenue, 1,180 employees, and a $133 stock price. Razorfish reached a $5.5B valuation. Sapient grew revenue from $950K to $503M between 1992 and 2000.

Then the floor gave way, and the speed of the collapse is the entire lesson. By August 2001 — less than a year after its peak — Scient's quarterly revenue had fallen from $100M to $11M, and it bought its dying rival iXL on the way down before going defunct in 2002. Razorfish fell to an $8.2M fire sale. Nearly the whole category — MarchFIRST, USWeb, Viant, Whittman-Hart, Scient, iXL — is now literally on Wikipedia's "list of former consulting firms."

The survivor explains why. Sapient lived by cutting 2002 revenue to about one-third of its 2000 level and rebuilding on blue-chip clients. And here is the number that matters most, stripped of bubble psychology: as it recovered, Sapient targeted a 10% operating margin — in line with Accenture's 11%. That is the honest, structural profitability of a well-run services firm. The billions in market cap were fiction; the durable reality is a solid double-digit-margin professional-services shop. (The market eventually agreed: even recovering, Sapient traded at 3.5x sales while Accenture — the honest comp — traded at 1.6x. Sapient was ultimately acquired by Publicis for $3.7B in 2015.)

Why the services model is structurally capped

Three constraints, all visible in the data:

It doesn't scale non-linearly. Revenue is bolted to headcount; you can't 10x revenue without ~10x-ing skilled humans, and good people take time to hire and train. The opposite of software economics.
The revenue is project-based and evaporates. $100M→$11M a quarter is what happens when your income is discretionary transformation budgets that freeze in a downturn. No recurring base, no moat.
The category's own thesis has a half-life. As HBR's "IT Doesn't Matter" argued, the new capability becomes ubiquitous infrastructure — like electricity — and stops being a competitive differentiator. The transformation you're paid a premium to deliver is, by definition, the thing that becomes commoditized once everyone has it.

The direct warning for "AI enablement" as a business: the consulting wedge is real and lucrative early, most valuable in exactly this pre-knee window — and it decays as agents-in-production becomes table stakes.

Chapter 6 — Entrepreneurial implications: building in the pre-knee window

Pulling the threads together into what this means for starting a company now.

Where we are dictates strategy

We're pre-knee, at web-1998. The web's value didn't come from the 1994–98 shelf; it came when adoption bent upward in 1999–2003 and the infrastructure to cross the production gap matured — CMSes, CDNs, payments, hosting. The equivalent AI infrastructure — eval harnesses, memory/context systems, agent reliability and observability, governance, integration scaffolding — is being built right now and mostly does not exist yet. Capability is a commodity; eval infrastructure and integration depth are the moats. That is the exact sentence a 1998 web-tooling founder would have recognized.

The corollary is reassurance about timing: you're early, not late. At year three on this curve the web hadn't yet produced its durable tooling winners at scale — Akamai IPO'd in 1999, the CMS wars were just starting. The Vignette/Akamai-equivalent AI companies are being founded in this 18-month window.

Don't build the pure-play "we'll do your AI for you" shop as the endgame

That's a Razorfish/Scient. It will feel incredible for 2–3 years — revenue, logos, team, conference — and then either the knee arrives and the big SIs crush you on scale, or the next wave commoditizes your craft and the project pipeline freezes. Best realistic outcome is acquisition by the category king (genuinely how Sapient's people won), but it caps at ~10% margins under someone else's logo.

Instead: use services as sensing apparatus to find the productizable layer underneath

This is the single most important move in the whole web playbook, and almost nobody did it deliberately. Vignette's CMS was CNET's internal PRISM tool, productized. Akamai came from one specific MIT problem. The consultancy that paid attention to what it kept rebuilding for every client found the product. The richest possible version of that sensing apparatus is a network of operators telling you, every week, what breaks when they try to put agents into production. That is the asset to mine.

The product is in the production gap, not the usage layer

Per the curve: usage is saturating (top of funnel, already commoditized). The whitespace is the back of the funnel — the ~88% of agents that never reach production. Whatever makes pilot→production reliable is the Akamai-shaped / Dreamweaver-shaped opportunity. It is professional-grade (not "AI for novices" — that's the FrontPage that dies in wave 2), it's infrastructure the giants haven't built because the knee hasn't forced them to yet, and it's recurring rather than project-based.

How to time the knee (watch the right number)

The signal won't be "more people use AI" — that's already saturating. It will be the pilot-to-production conversion rate moving: when "5–14% in production" becomes 25%, then 40%. That conversion already shifted once — programs never reaching payback fell from 34% (2025) to 19% (2026), because vendor agents began shipping with eval harnesses and integration templates that custom builds previously had to invent themselves. That declining failure rate is the early tremor of the knee — and a direct signal about where the durable tooling value sits.

Chapter 7 — What we might be missing (the honest unknowns)

Intellectual honesty requires marking where the analogy could break, because that's also where the asymmetric opportunity hides.

The biggest unknown: is the AI knee demand-driven or supply-constrained? The web's knee was demand-driven — consumers showed up, businesses followed, and the bottleneck was just building enough infrastructure to serve obvious demand. That made the web's knee essentially inevitable, a question of when not if. AI's bottleneck looks supply-side and technical — reliability, trust, governance — which means the knee could stall rather than merely delay. Gartner forecasts that more than 40% of agentic-AI projects will be canceled by end of 2027 over costs, unclear value, and weak risk controls.

That ambiguity is itself the opportunity. If the bottleneck is reliability, the company that solves reliability isn't riding the knee — it's causing it. In a demand-driven revolution you surf; in a supply-constrained one you can move the curve. That's a fundamentally more valuable position than the web's tooling winners ever had.

Other things the comparison suggests we may be underweighting:

The danger is our own success. Scient at $100M/quarter felt invincible 11 months before $11M. A services business that's working is the easiest place to fail to notice you should have been building the product. The boom masks the structural cap right up until it doesn't.
The implementer role's half-life is shorter this time. The front of the funnel moves 2–3x faster, which likely means the generalist "AI enablement" role specializes faster than the webmaster did. The window to convert that vantage point into a product is correspondingly tighter.
The loss-leader squeeze is already loaded. The model providers (OpenAI, Anthropic, Google) are the Adobe/Microsoft of this era — perfectly positioned to ship "good enough" versions of mid-layer tooling as loss leaders to their platform. Any product that sits too close to the model layer is FrontPage-exposed. The defensible positions are the same two extremes: deep infrastructure the providers won't own, or the professional-grade tool/workflow they won't bother to build well.

Coda: the one-paragraph version

The web's services firms made fortunes on paper and almost all died; the durable money was made by whoever turned what they learned serving clients into productized infrastructure for the production gap. We are at web-1998: the front of the curve is sprinting, the back is crawling, and the knee — the prize — is ahead, not behind. The move is to use the services/enablement layer as a sensing network, find the thing every operator keeps rebuilding to get agents into production, and build that — at one of the two durable extremes (deep infra or professional-grade tool), aimed at the enterprise, before the model providers commoditize the middle. And if the bottleneck really is reliability, the reward for solving it isn't a ride up the curve — it's the power to bend it.

Sources: US Census Bureau and Australian Bureau of Statistics business-technology adoption series; company histories of Vignette, NetObjects, Akamai, Speedera, Macromedia/Dreamweaver, FrontPage, Razorfish, Scient, iXL, Sapient/Publicis Sapient; St. Louis Fed individual-usage data; industry production-deployment and pilot-conversion surveys; Gartner agentic-AI forecasts. Compiled from primary research, May 2026.

AI vs. the Web Revolution in Retrospect

About

Preview