What a conversation about the Semantic Web reveals about the argument the AI debate keeps having with itself.
In the summer of 2008, the inventor of the World Wide Web stood up at a conference in New York and told the room how the next era of the internet would work. Get the world’s data onto the web in one common format, Tim Berners-Lee argued, agree on a simple shared set of rules for linking it, and a more intelligent web — the Semantic Web, Web 3.0 — would follow. His instinct was the natural one for a technologist of his standing: the path to a smarter system runs through a better standard.
I wrote a response to that argument at the time. Not because I thought Berners-Lee was wrong about where things were heading — he plainly wasn’t — but because I thought the mechanism he was reaching for would run into the same wall I had already watched a decade of procurement technology run into. My objection was a single sentence, and it is the same sentence I find myself making again in 2026, to a different person, about a different technology:
You cannot standardize your way past the real world. A system that asks the environment to adapt to it will lose to a system that adapts to the environment.
That is the whole argument. In 2008 it was aimed at the Semantic Web. In 2025 and 2026 it is aimed at the case being made for AI and “data sovereignty.” The technology in the sentence changes. The sentence does not.
If that sounds like a history lesson, here is why it isn’t. In 2025, MIT’s Project NANDA found that roughly 95% of enterprise generative-AI pilots produced no measurable return on the P&L — and located the cause not in the models, the talent, or the infrastructure, but in what its authors called the learning gap: integration, context, organizational fit. The technology did not fail. Organizations failed to absorb it. That is the same diagnosis I brought to Berners-Lee’s standard in 2008, which is why a conversation about the Semantic Web turns out to be a conversation about your AI program — and, if you run IT, about the failures you keep being handed to explain.
What I was actually describing
The term I used in 2008 was “Web 4.0.” I meant something specific by it, and it is worth being precise, because the precision is the point.
Berners-Lee’s Web 3.0 was about meaning within a single thread — teaching a machine to understand that two things are related because their words and context line up. What I described as the step beyond it was a mechanism that could “assemble and manage seemingly disparate streams of information into a collective outcome that has real-world applicability” — multiple independent threads, drawn together autonomously into a result a person could actually use, with the answer differing legitimately from one user to the next.
I want to be precise about what that was, because the precision cuts both ways — it keeps me from overclaiming, but it also keeps me from underclaiming what was already built. This wasn’t a thought experiment. The Metaprise™ architecture underneath RAM, developed in 1998 under Canada’s SR&ED program, ran on exactly this principle: multiple internal and external agents, each carrying a different strand of input, resolved dynamically and in real time. I did not call them large language models — the mechanism that now does this at scale did not exist, and I won’t pretend I foresaw transformers or any other specific technology. But the architecture was agent-based, multi-strand, and real-time by design, and the 2008 “Web 4.0” piece simply named the capability that architecture was built to deliver. Eighteen years later the industry has its own word for that capability. It calls it agentic AI: multiple information streams, autonomously orchestrated into a usable outcome. The name changed. The architecture was already running.
And the architecture is not even the main point. Plenty of people gestured at a smarter, more connected web; the description was never the rare thing. The rare thing was the diagnosis attached to it — because the diagnosis is what the present debate keeps getting wrong.
The same error, seventeen years apart
Set the two arguments side by side.
Berners-Lee, 2008: adopt the common standard, link the data under one simple set of rules, and the intelligent web follows. The standard is the cause; capability is the effect.
Michael Gale, in his August 2025 essay responding to MIT’s finding, argues that the enterprises that win with AI are the ones that chose to own their AI and data platform — that made data sovereignty a mission-critical decision. The choice is the cause; the flywheel of advantage is the effect. Gale’s case is well argued and his data is real — 2,050 executives across thirteen countries, a thriving cohort of around 13% reporting outsized returns, a named roster of firms he offers as proof. I take all of it as given. My disagreement is not with his numbers and certainly not with his motives. It is with the direction of the arrow.
Both arguments make the same move. Both take a thing that is visible in the winners — and treat it as the thing that caused the winning. And both, in doing so, skip the question that actually determines the outcome: what had to already be true inside the organization for that choice to produce results rather than wreckage?
Berners-Lee’s standard fails in a diverse environment for the same reason a centrally mandated procurement platform fails across a sprawling enterprise: the world it is trying to organize is too varied to be coerced into one set of rules, and the moment you force it, the people and conditions on the ground route around you. Gale’s sovereignty fails for the same reason: you cannot buy a hybrid, open, well-governed data estate as a product. You can only earn it, by having already aligned the departments, the workflows, and the operating conditions that a data estate sits on top of. The firms in Gale’s thriving 13% did not succeed because they bought sovereignty. They were able to make sovereignty work because they were already the kind of organization that could — disciplined across prior technology eras, long before AI gave that discipline a new name.(The eVA story, including the supply-base growth and the contrast with the Canadian federal effort, is documented here: [Yes Virginia! There is more to e-procurement than software!])
Sovereignty is the visible strategy. Discipline is the underlying cause. The choice did not create the capability; the capability made the choice possible.
Why this particular exchange matters now
I have made some version of this argument for a long time, across domains that look unrelated on the surface — strategic sourcing, supplier economics, sustainability, government procurement reform. Each time, from a different starting point, the same conclusion: the technology is the final piece, not the first, and what determines the outcome is the alignment of people, process, and conditions before the technology arrives.
The 2008 Berners-Lee exchange is the one I keep returning to in the context of the AI debate, and here is the specific reason. It is the one place where the argument was made at the level of information architecture itself — not about a procurement tool or a sourcing strategy, but about how the underlying system of data and intelligence should be built. That is the exact layer the AI debate now lives at. When the conversation in 2026 is about data estates, platform ownership, and whether you can purchase your way to an AI advantage, it is having a fight about information architecture. And the position I took on that fight — that real-world conditions cannot be coerced, standardized, or purchased into alignment — was on the record before the category that would test it existed.
This is what I mean when I say the 2008 reference got both the function and the failure mode right. The function: a mechanism that orchestrates many streams into a usable, context-specific outcome — what we now call agentic AI. The failure mode: the belief that the right standard, or the right purchase, is the cause of success rather than its visible symptom. The first was a description of where things were going. The second was a warning about how most organizations would get there wrong. Both have held — which is also why MIT’s finding reads as corroboration rather than catastrophe: the divide it identified is the substrate argument arrived at independently, from inside the most-quoted study in the field.
Same platform, opposite outcome
If you run technology for a living, this is the section that matters, because it’s the one the sovereignty argument can’t account for.
If owning the right platform were the cause of success, then the same platform, competently implemented, should produce the same result. It doesn’t — and the cleanest demonstrations come from cases where the technology is held identical by fact rather than by assertion.
The Commonwealth of Virginia’s eVA procurement program and a comparable Canadian federal initiative ran on the same class of platform. Virginia treated reform as a problem of understanding its operating conditions first, and built one of the most durable public e-procurement programs on record. The federal effort modeled itself on Virginia’s outcome, imported the same technology, and skipped the diagnosis. Same instrument. Opposite song. The variable that separates them is not the platform, and — this is the part worth sitting with — it is not the competence of the technologists either. Both had capable people. One had done the upstream work; the other had not.
The point sharpens in a case from my 2008 CATA Alliance paper that I’d put in front of any CIO. Hewlett-Packard was not a victim of unfamiliar technology — it was building a practice to rival IBM’s at implementing the very SAP systems it then failed to merge cleanly inside its own walls, at a cost the paper documented at roughly $400 million in lost revenue. An acknowledged expert in the product could not make the product work. As the paper put it: if a high-technology company with extensive experience in the product can’t succeed, what does that say about anyone else’s chances? The expertise was never the variable. The sequence was.
I lay these out as the comparisons my archive documents, not as laboratory trials — but the logic is exactly what survivorship-bias-prone winner lists lack, because the technology is held constant and only the conditions differ. And the conclusion they point to is not the one CIOs usually brace for. It is not that the platform doesn’t matter, or that IT’s skill doesn’t matter. Both are necessary, and both are yours to own. They are simply not sufficient — and the missing piece sits upstream of IT, in the operating conditions the platform is dropped into. Which is precisely why IT keeps inheriting the blame for failures that were decided before the platform was ever chosen. The substrate failure isn’t IT’s to cause. It is, too often, IT’s to explain.
Where the evidence is strong, and where it isn’t
I hold this to the confidence the evidence supports — which, for an IT argument, is the difference between a claim that survives the room and one that gets picked apart in it.
That these firms — and organizations like them — exhibited the discipline to absorb successive technology eras before AI is well supported, and it is a direct, evidenced answer to the implication that AI-era success is an AI-era phenomenon. That this discipline is the cause of their success rather than one correlate among several is a more provisional claim, and it stays provisional for reasons no amount of conviction removes: the firms held up as proof are, by construction, the ones that won, and a pattern among winners is not the same as a proven cause. The stronger evidence for causation does not come from any roster of winners. It comes from controlled comparisons of the kind above — the same technology deployed into organizations with different underlying conditions, producing opposite outcomes. Those carry the causal weight that a list of successes cannot.
As for the 2008 connection itself: I haven’t seen anyone else make this specific link — taking the information-architecture argument from the dawn of the Semantic Web and running it forward to the AI-sovereignty debate as a single, continuous diagnosis. If someone has, point me to it. I’ll state it as what it is — a connection I haven’t seen made elsewhere — because that’s accurate, and accurate is harder to knock down than “first.”
The point for the 87%
None of this is an argument against AI, and it is not an argument against Gale’s winners. It is an argument about sequence, and sequence is good news for the majority of organizations that MIT found in the failing column.
If success came from a purchase — the right platform, the right standard, the right sovereignty posture — then the firms without it would be permanently behind, waiting to buy their way in. But if the visible strategy is downstream of an underlying discipline, then the work that matters is work any organization can actually do: diagnose your own real-world conditions, align the people and processes that the technology will sit on, and then introduce the technology to a substrate that can absorb it. That is not a product you acquire. It is a sequence you follow. The winners were not luckier or better-funded in some way the rest can never reach. In most cases they simply did the alignment first — and AI, like every technology before it, arrived to prove the argument, not to make it.
Technology changes capability. Substrate determines survivability. That was true when the question was how to build a smarter web in 2008, and it is true now that the question is how to make AI deliver in 2026. The standard was never the answer. Sovereignty isn’t the answer either. The answer is, and has always been, what you align before the technology shows up.
This piece sits alongside a more detailed working paper that lays out the verification method, the controlled comparison cases, and the contemporaneous record behind the 2008 argument, including the published Procurement Insights archive from that period. I’m glad to share it with anyone who wants to examine the evidence rather than take the argument on faith — including Michael Gale, whose essay prompted this one and whose challenge to it I’d welcome.
ABOUT HANSEN: Hansen Procurement Technologies is a next-generation analyst and advisory firm for the AI era. We help organizations determine whether they are ready to absorb, operationalize, and govern new technologies before significant investments are made. Through diagnosis, validation, and compliance assessment, we focus on the conditions that determine whether technology delivers measurable outcomes.
-30-
Related
The Standard Wasn’t the Answer in 2008. Sovereignty Isn’t the Answer in 2026.
Posted on June 8, 2026
0
What a conversation about the Semantic Web reveals about the argument the AI debate keeps having with itself.
In the summer of 2008, the inventor of the World Wide Web stood up at a conference in New York and told the room how the next era of the internet would work. Get the world’s data onto the web in one common format, Tim Berners-Lee argued, agree on a simple shared set of rules for linking it, and a more intelligent web — the Semantic Web, Web 3.0 — would follow. His instinct was the natural one for a technologist of his standing: the path to a smarter system runs through a better standard.
I wrote a response to that argument at the time. Not because I thought Berners-Lee was wrong about where things were heading — he plainly wasn’t — but because I thought the mechanism he was reaching for would run into the same wall I had already watched a decade of procurement technology run into. My objection was a single sentence, and it is the same sentence I find myself making again in 2026, to a different person, about a different technology:
That is the whole argument. In 2008 it was aimed at the Semantic Web. In 2025 and 2026 it is aimed at the case being made for AI and “data sovereignty.” The technology in the sentence changes. The sentence does not.
If that sounds like a history lesson, here is why it isn’t. In 2025, MIT’s Project NANDA found that roughly 95% of enterprise generative-AI pilots produced no measurable return on the P&L — and located the cause not in the models, the talent, or the infrastructure, but in what its authors called the learning gap: integration, context, organizational fit. The technology did not fail. Organizations failed to absorb it. That is the same diagnosis I brought to Berners-Lee’s standard in 2008, which is why a conversation about the Semantic Web turns out to be a conversation about your AI program — and, if you run IT, about the failures you keep being handed to explain.
What I was actually describing
The term I used in 2008 was “Web 4.0.” I meant something specific by it, and it is worth being precise, because the precision is the point.
Berners-Lee’s Web 3.0 was about meaning within a single thread — teaching a machine to understand that two things are related because their words and context line up. What I described as the step beyond it was a mechanism that could “assemble and manage seemingly disparate streams of information into a collective outcome that has real-world applicability” — multiple independent threads, drawn together autonomously into a result a person could actually use, with the answer differing legitimately from one user to the next.
I want to be precise about what that was, because the precision cuts both ways — it keeps me from overclaiming, but it also keeps me from underclaiming what was already built. This wasn’t a thought experiment. The Metaprise™ architecture underneath RAM, developed in 1998 under Canada’s SR&ED program, ran on exactly this principle: multiple internal and external agents, each carrying a different strand of input, resolved dynamically and in real time. I did not call them large language models — the mechanism that now does this at scale did not exist, and I won’t pretend I foresaw transformers or any other specific technology. But the architecture was agent-based, multi-strand, and real-time by design, and the 2008 “Web 4.0” piece simply named the capability that architecture was built to deliver. Eighteen years later the industry has its own word for that capability. It calls it agentic AI: multiple information streams, autonomously orchestrated into a usable outcome. The name changed. The architecture was already running.
And the architecture is not even the main point. Plenty of people gestured at a smarter, more connected web; the description was never the rare thing. The rare thing was the diagnosis attached to it — because the diagnosis is what the present debate keeps getting wrong.
The same error, seventeen years apart
Set the two arguments side by side.
Berners-Lee, 2008: adopt the common standard, link the data under one simple set of rules, and the intelligent web follows. The standard is the cause; capability is the effect.
Michael Gale, in his August 2025 essay responding to MIT’s finding, argues that the enterprises that win with AI are the ones that chose to own their AI and data platform — that made data sovereignty a mission-critical decision. The choice is the cause; the flywheel of advantage is the effect. Gale’s case is well argued and his data is real — 2,050 executives across thirteen countries, a thriving cohort of around 13% reporting outsized returns, a named roster of firms he offers as proof. I take all of it as given. My disagreement is not with his numbers and certainly not with his motives. It is with the direction of the arrow.
Both arguments make the same move. Both take a thing that is visible in the winners — and treat it as the thing that caused the winning. And both, in doing so, skip the question that actually determines the outcome: what had to already be true inside the organization for that choice to produce results rather than wreckage?
Berners-Lee’s standard fails in a diverse environment for the same reason a centrally mandated procurement platform fails across a sprawling enterprise: the world it is trying to organize is too varied to be coerced into one set of rules, and the moment you force it, the people and conditions on the ground route around you. Gale’s sovereignty fails for the same reason: you cannot buy a hybrid, open, well-governed data estate as a product. You can only earn it, by having already aligned the departments, the workflows, and the operating conditions that a data estate sits on top of. The firms in Gale’s thriving 13% did not succeed because they bought sovereignty. They were able to make sovereignty work because they were already the kind of organization that could — disciplined across prior technology eras, long before AI gave that discipline a new name.(The eVA story, including the supply-base growth and the contrast with the Canadian federal effort, is documented here: [Yes Virginia! There is more to e-procurement than software!])
Sovereignty is the visible strategy. Discipline is the underlying cause. The choice did not create the capability; the capability made the choice possible.
Why this particular exchange matters now
I have made some version of this argument for a long time, across domains that look unrelated on the surface — strategic sourcing, supplier economics, sustainability, government procurement reform. Each time, from a different starting point, the same conclusion: the technology is the final piece, not the first, and what determines the outcome is the alignment of people, process, and conditions before the technology arrives.
The 2008 Berners-Lee exchange is the one I keep returning to in the context of the AI debate, and here is the specific reason. It is the one place where the argument was made at the level of information architecture itself — not about a procurement tool or a sourcing strategy, but about how the underlying system of data and intelligence should be built. That is the exact layer the AI debate now lives at. When the conversation in 2026 is about data estates, platform ownership, and whether you can purchase your way to an AI advantage, it is having a fight about information architecture. And the position I took on that fight — that real-world conditions cannot be coerced, standardized, or purchased into alignment — was on the record before the category that would test it existed.
This is what I mean when I say the 2008 reference got both the function and the failure mode right. The function: a mechanism that orchestrates many streams into a usable, context-specific outcome — what we now call agentic AI. The failure mode: the belief that the right standard, or the right purchase, is the cause of success rather than its visible symptom. The first was a description of where things were going. The second was a warning about how most organizations would get there wrong. Both have held — which is also why MIT’s finding reads as corroboration rather than catastrophe: the divide it identified is the substrate argument arrived at independently, from inside the most-quoted study in the field.
Same platform, opposite outcome
If you run technology for a living, this is the section that matters, because it’s the one the sovereignty argument can’t account for.
If owning the right platform were the cause of success, then the same platform, competently implemented, should produce the same result. It doesn’t — and the cleanest demonstrations come from cases where the technology is held identical by fact rather than by assertion.
The Commonwealth of Virginia’s eVA procurement program and a comparable Canadian federal initiative ran on the same class of platform. Virginia treated reform as a problem of understanding its operating conditions first, and built one of the most durable public e-procurement programs on record. The federal effort modeled itself on Virginia’s outcome, imported the same technology, and skipped the diagnosis. Same instrument. Opposite song. The variable that separates them is not the platform, and — this is the part worth sitting with — it is not the competence of the technologists either. Both had capable people. One had done the upstream work; the other had not.
The point sharpens in a case from my 2008 CATA Alliance paper that I’d put in front of any CIO. Hewlett-Packard was not a victim of unfamiliar technology — it was building a practice to rival IBM’s at implementing the very SAP systems it then failed to merge cleanly inside its own walls, at a cost the paper documented at roughly $400 million in lost revenue. An acknowledged expert in the product could not make the product work. As the paper put it: if a high-technology company with extensive experience in the product can’t succeed, what does that say about anyone else’s chances? The expertise was never the variable. The sequence was.
I lay these out as the comparisons my archive documents, not as laboratory trials — but the logic is exactly what survivorship-bias-prone winner lists lack, because the technology is held constant and only the conditions differ. And the conclusion they point to is not the one CIOs usually brace for. It is not that the platform doesn’t matter, or that IT’s skill doesn’t matter. Both are necessary, and both are yours to own. They are simply not sufficient — and the missing piece sits upstream of IT, in the operating conditions the platform is dropped into. Which is precisely why IT keeps inheriting the blame for failures that were decided before the platform was ever chosen. The substrate failure isn’t IT’s to cause. It is, too often, IT’s to explain.
Where the evidence is strong, and where it isn’t
I hold this to the confidence the evidence supports — which, for an IT argument, is the difference between a claim that survives the room and one that gets picked apart in it.
That these firms — and organizations like them — exhibited the discipline to absorb successive technology eras before AI is well supported, and it is a direct, evidenced answer to the implication that AI-era success is an AI-era phenomenon. That this discipline is the cause of their success rather than one correlate among several is a more provisional claim, and it stays provisional for reasons no amount of conviction removes: the firms held up as proof are, by construction, the ones that won, and a pattern among winners is not the same as a proven cause. The stronger evidence for causation does not come from any roster of winners. It comes from controlled comparisons of the kind above — the same technology deployed into organizations with different underlying conditions, producing opposite outcomes. Those carry the causal weight that a list of successes cannot.
As for the 2008 connection itself: I haven’t seen anyone else make this specific link — taking the information-architecture argument from the dawn of the Semantic Web and running it forward to the AI-sovereignty debate as a single, continuous diagnosis. If someone has, point me to it. I’ll state it as what it is — a connection I haven’t seen made elsewhere — because that’s accurate, and accurate is harder to knock down than “first.”
The point for the 87%
None of this is an argument against AI, and it is not an argument against Gale’s winners. It is an argument about sequence, and sequence is good news for the majority of organizations that MIT found in the failing column.
If success came from a purchase — the right platform, the right standard, the right sovereignty posture — then the firms without it would be permanently behind, waiting to buy their way in. But if the visible strategy is downstream of an underlying discipline, then the work that matters is work any organization can actually do: diagnose your own real-world conditions, align the people and processes that the technology will sit on, and then introduce the technology to a substrate that can absorb it. That is not a product you acquire. It is a sequence you follow. The winners were not luckier or better-funded in some way the rest can never reach. In most cases they simply did the alignment first — and AI, like every technology before it, arrived to prove the argument, not to make it.
Technology changes capability. Substrate determines survivability. That was true when the question was how to build a smarter web in 2008, and it is true now that the question is how to make AI deliver in 2026. The standard was never the answer. Sovereignty isn’t the answer either. The answer is, and has always been, what you align before the technology shows up.
This piece sits alongside a more detailed working paper that lays out the verification method, the controlled comparison cases, and the contemporaneous record behind the 2008 argument, including the published Procurement Insights archive from that period. I’m glad to share it with anyone who wants to examine the evidence rather than take the argument on faith — including Michael Gale, whose essay prompted this one and whose challenge to it I’d welcome.
ABOUT HANSEN: Hansen Procurement Technologies is a next-generation analyst and advisory firm for the AI era. We help organizations determine whether they are ready to absorb, operationalize, and govern new technologies before significant investments are made. Through diagnosis, validation, and compliance assessment, we focus on the conditions that determine whether technology delivers measurable outcomes.
-30-
Share this:
Like this:
Related