Enterprise AI has impressive models and infrastructure, yet inside companies the pattern remains pilots and bespoke deployments. The missing piece, in this telling, is an application layer like the web provided for the internet—persistent context, business sem
In the demo rooms, enterprise AI can look like the future. Models write. They summarize. They reason and code. They search and retrieve. They translate, classify, plan, and increasingly act. The machinery is there.
But in company offices, the story often lands somewhere else: pilots sprout everywhere, and transformation stays stuck near the finish line.
The argument unfolding here is that the whole industry is running on raw capability without the missing layer that would make it usable as an everyday business environment. The writer’s answer is blunt: enterprise AI is in 1991—exactly when the internet worked, but before it had the web.
In 1991, the internet already moved packets using TCP/IP. Email connected people across institutions. FTP moved files. Telnet enabled remote access. Universities, research labs, and technically sophisticated organizations could use the network. It was powerful. but it wasn’t yet consumable for a normal company as a business space in the way the web later became.
Then the World Wide Web arrived with a thin, decisive overlay: URLs, HTTP, HTML, servers, and browsers. CERN’s history of the web is used to anchor the timeline—by Christmas 1990. Tim Berners-Lee had defined the basic concepts of HTML. HTTP. and URLs. and written the first browser/editor and server software. In 1991. CERN released the WWW software more broadly and announced it on internet newsgroups. helping the idea spread beyond its original context.
The point of the analogy is not that networking began with the web. It’s that the web made networking legible, usable, and buildable for everyone else. Enterprise AI, the argument goes, still lacks the equivalent layer that turns powerful infrastructure into something organizations can consistently run.
Large language models, in this framing, are extraordinary infrastructure—one of the most important technological substrates of the time. But infrastructure is not the same thing as an application layer.
A company trying to deploy LLMs today is compared to a bookstore trying to sell online before the web existed. The network is there: packets move, servers exist. But each transaction would require custom machinery—custom protocols, custom interfaces, custom logic, custom deployment, custom integration, and custom everything. That isn’t commerce; it’s engineering.
That’s offered as a reason the enterprise AI market still leans heavily on pilots. bespoke deployments. forward-deployed engineers. and consulting-heavy implementations. The issue isn’t that the intelligence is fake. It’s that the layer that makes it consumable by ordinary organizations is still immature.
A model can generate an answer. A company. though. needs a system that places that answer into the real world of work: where it fits. what data it can use. what constraints apply. who has permission to act. what process is affected. which outcome matters. and how the system learns from what happens next.
That’s presented as the missing layer—not “more AI,” and not just better prompts. The gap is described as specific and identifiable: enterprise AI needs an equivalent of the web layer. a structured application layer that turns raw intelligence into something organizations can use repeatedly. safely. and at scale.
The writer lays out seven properties the missing layer would have to deliver.
First is persistent context: the system cannot behave as if every interaction starts from zero.
Second is business semantics: it must understand customers, products, policies, workflows, roles, and constraints in company-specific terms.
Third is process state: it has to know where work is, what has happened, what is pending, and what depends on what.
Fourth is permission and governance models: it must operate inside organizational boundaries, not around them.
Fifth is feedback loops: it needs to learn from outcomes, not merely generate outputs.
Sixth is interoperability: it must connect to systems of record, tools, data, and workflows without rebuilding bespoke integration every time.
Seventh is repeatability: it should be deployable as architecture rather than artisanal consulting.
The emphasis on context engineering by Anthropic is offered as an indicator of where the industry may be headed. Its engineering team explicitly describes context as a critical but finite resource for agents. arguing that the challenge is to curate and manage the information surrounding the model—not simply write better prompts.
The direction of travel is stated as a shift in what counts as the product: the model is no longer the whole offering. The environment around the model becomes the product.
A second analogy sharpens the claim that enterprise AI is in a pre-industrial phase of enterprise software. Before ERP systems became standardized platforms, corporate software was often a patchwork—custom implementations, integrations, internal systems, and consulting projects. SAP’s history is cited for the long arc from specialized business software toward enterprise application platforms. with SAP eventually becoming the market leader in enterprise application software.
That evolution mattered because it didn’t just digitize individual functions. It industrialized a way of representing the company—finance. inventory. procurement. manufacturing. HR. logistics. and reporting became standardized enough to create repeatable implementations and a partner ecosystem. The same pattern is linked to CRM and SaaS. Salesforce’s history is used to point to AppExchange. a marketplace for independent software vendors and applications that helped turn Salesforce from a product into a platform ecosystem.
The difference, as described, is between a category that scales and one that depends on custom projects.
Today, enterprise AI is described as still stuck in that custom-project phase. Each company, the writer says, must map processes, clean data, understand permissions, reconstruct workflows, encode constraints, and define outcomes. That work can be necessary—but when it has to be done manually in every deployment. it shows the platform layer hasn’t arrived.
The strategic sting comes next. In the web transition, the decisive question wasn’t who owned the cables. In enterprise software, it wasn’t who owned the database or server hardware. The critical question was who defined the system layer that made networking or business representation usable—and who built the ecosystem around it.
In this view of AI, the next winners may not be the companies with the largest models or biggest clusters. Those companies are still important. But category-defining power could belong to whoever builds the missing application layer—one that makes enterprise intelligence persistent. governed. contextual. process-aware. and repeatable.
That’s also tied to the critique of the industry’s current obsession with model performance, context windows, and benchmark scores. Better models are described as necessary, but not sufficient.
McKinsey’s 2025 research on AI adoption is cited to support a different measure of value: companies seeing the most value aren’t just deploying tools; they’re redesigning workflows and embedding AI into processes.
Deloitte’s work on agentic AI is also cited, saying many organizations hit a wall because they try to automate processes designed for humans instead of reimagining how the work should actually be done.
Taken together, the bottleneck is portrayed as shifting upward in the stack.
Transitions of this kind, the writer notes, are hard to see while they’re happening and obvious afterward. Before the web, the internet looked like a domain for specialists. After the web, it became a business environment. Before ERP and SaaS platforms matured, enterprise software looked like custom automation; afterward, it became repeatable architecture. Before cloud platforms matured, infrastructure looked like procurement and systems administration; afterward, it became programmable capacity.
Enterprise AI is now described as approaching a similar threshold—currently looking artisanal, with pilots, prototypes, integrations, forward-deployed engineers, consulting-heavy engagements, and custom workflow mapping.
That period is treated as normal. Experts often have to carry powerful technology across the gap manually at first. But the destination is the layer that makes expert intervention less central.
The next five years are framed as decisive. The web didn’t commercialize overnight. ERP didn’t standardize enterprise overnight. Salesforce didn’t create a platform ecosystem in a single release. These transitions take years. Still. the decisive moment usually arrives when someone defines the missing layer well enough that everyone else can build on it.
The writer says that moment is what enterprise AI still lacks. The industry has models. infrastructure. early agents. the consulting wave. pilots. frustration. proof that isolated tools aren’t enough. and growing recognition that context. workflows. constraints. memory. and outcomes matter more than prompts.
What remains absent is the equivalent of the browser, the URL, the ERP layer, or the AppExchange—the standard application layer that would make enterprise AI consumable by ordinary companies.
Until it appears, the industry remains in a paradox: extraordinary intelligence delivered through extraordinary effort.
So the central question becomes less about which model wins and more about who will define the layer that turns intelligence into enterprise infrastructure. When that layer arrives. the writer argues. the current debate will look different—forward-deployed engineers won’t vanish but become less central. custom deployments won’t disappear but stop dominating the pattern. and pilots won’t go away but the path to production will shorten.
In that scenario, AI stops being something companies experiment with and starts being something companies are built on. The industrial era is described as not arrived yet. But once the missing layer appears, it’s predicted that it will feel obvious all along.
MISRYOUM
enterprise AI large language models pilots application layer web analogy TCP/IP World Wide Web HTTP HTML URLs context engineering agents interoperability governance McKinsey 2025 Deloitte agentic AI ERP SAP Salesforce AppExchange