The consumer AI market still knows how to put on a show.
One week the conversation is Anthropic versus OpenAI. The next week it is GPT, Opus, then Fable, then whatever name gets attached to the newest model that feels a little faster, a little more natural, a little more willing to reason its way through the strange little tests we keep inventing for it. The product launches have become a kind of technology weather system. Everyone watches the front move in. Everyone argues about which one feels warmer.
For consumers, that makes sense. The model is the product, or at least close enough to the product that the distinction barely matters. If one assistant writes better, thinks longer, speaks more naturally, or handles a messy prompt with less handholding, you feel the difference immediately. You can switch tabs and compare the thing directly. The competition is visible.
Enterprise AI does not work that cleanly.
Inside a large organization, the smartest raw model is rarely the whole answer. Sometimes it is not even the main answer. The work is buried under permissions, data boundaries, systems of record, workflow exceptions, security policies, compliance requirements, and all the operational scar tissue that never appears in a launch demo.
That is why the consumer framing starts to break down when it gets carried into the enterprise. Three years in, the question is no longer just which model is best. It is which fabric can turn a capable model into something the business can actually use.
The first conversation was about the model. Which one was smartest. Which one reasoned better. Which one had the larger context window, the better benchmark score, the more convincing demo, the cleaner answer to the impossible prompt someone found on the internet.
That conversation mattered. It still matters at the frontier.
It is also not the main enterprise question anymore.
Most large organizations are not trying to solve pure intelligence. They are trying to solve work. That is a different problem. Work does not happen inside a model. Work happens across systems, permissions, records, workflows, exceptions, approvals, audit trails, and all the weird connective tissue a company has built over twenty years of buying software.
The model is part of that system. It is not the system.
The model race compressed faster than the operating model
The early AI cycle was naturally model-centric. That was where the visible progress was.
One month a model could summarize a document. A few months later it could write code. Then it could reason through a multi-step task, read images, use tools, hold more context, and stop embarrassing itself quite as often in front of executives. Every release felt like a new ceiling getting lifted.
For frontier research, deep reasoning, novel discovery, and the very hardest technical tasks, that race still matters. If you are using AI to search a chemical space, prove something difficult, design a new architecture, or reason through a genuinely novel problem, the frontier matters. There are use cases where a few percentage points of model quality are the whole game.
Most enterprise work is not that.
Most enterprise AI work is summarizing messy information, drafting from known patterns, answering questions grounded in company data, routing work, checking policy, preparing decisions, moving information from one system to another, and helping people operate the business with less friction.
For that class of work, many models are now good enough.
Not perfect. Not magic. Good enough.
That changes the shape of competition. When intelligence is scarce, the smartest model wins. When intelligence becomes broadly available, the advantage moves to the thing that makes intelligence useful.
Intelligence is abundant. Context is scarce.
That is the shift.
Fabric is not a metaphor
The word "fabric" can sound suspiciously like something invented in a conference room after the better nouns were taken.
I mean it concretely.
The enterprise AI fabric is the layer around the model that lets it do useful work inside an organization. It is the data pipeline that decides what information the model can see. It is the retrieval layer that grounds an answer in current documents rather than internet soup. It is the identity system that determines whether the agent is allowed to read the contract, update the ticket, or touch the financial record.
It is the integration layer into CRM, ERP, HR, procurement, finance, ticketing, email, chat, document stores, and all the line-of-business systems people politely call legacy because calling them load-bearing and terrifying feels rude.
It is security and compliance. Classification. Audit. Retention. Policy. Data loss prevention. The part of the stack that sounds boring until the first time an AI assistant confidently summarizes a document it should never have been able to open.
It is agent orchestration and tool use. Which agent does what. Which tools it can call. What happens when two agents need the same system. Where a human approval gets inserted. How the action is logged. What the escalation path looks like when the agent runs into a judgment call instead of a task.
It is workflow automation. Vector stores. Retrieval systems. Compute. Deployment. Monitoring. Cost controls. The operational plumbing that turns a prompt into a governed business process.
None of those pieces are optional at enterprise scale.
A generic model can write a decent email. A model connected to the right fabric can understand the account context, apply the policy, draft the response, route it through the right approval, update the source system, and leave an audit trail behind it.
That is the difference between a clever tool and work getting done.
Relevance beats raw intelligence inside the enterprise
The harder problem is not whether the model can think.
The harder problem is whether it knows what matters here.
Every company has its own language. Not the values-on-the-wall language. The real language. The shorthand in tickets. The names of internal processes. The version of the customer record that is trusted and the version everyone knows to ignore. The policy exception that lives in a PDF from six months ago. The approval path nobody wrote down because the people who know it have been doing it for years.
That context is where enterprise value lives.
A model without it can sound smart and still be useless. It can produce a polished answer to the wrong question. It can summarize a policy without knowing which policy is current. It can propose a workflow that makes sense in the abstract and breaks the minute it touches the actual business.
This is why the frontier narrative often misleads enterprise buyers. The public story is about whether the model can solve a harder benchmark. The internal reality is about whether it can find the right file, respect the right permission, use the right system, and know when not to act.
In the enterprise, relevance is not a feature. It is the product.
We have seen this movie before
Technology markets usually start with components and end with platforms.
The early personal computer market looked like a battle of applications and hardware specs until the operating system became the control plane. The cloud market looked like compute, storage, and networking until the platform wrapped those components in identity, governance, deployment, observability, and a buying motion the enterprise could actually absorb.
Databases are another version of the same story. Query engines matter. Performance matters. But the systems that captured durable value were the ones that became the place where data lived, moved, got governed, got backed up, got secured, and got trusted.
Enterprise software ecosystems work the same way. Point solutions can win categories. Platforms win workflows. The component gets evaluated by a technical team. The platform gets embedded into how the organization operates.
AI is moving through that same curve, only faster.
The model is the component everyone can see. The fabric is the platform forming around it.
The vendor landscape starts to sort itself
This is where the market gets interesting.
Model-first companies have the cleanest story and the hardest enterprise problem. They can build extraordinary capability. They can move quickly. They can define the frontier. But unless they own the context layer, the workflow surface, or the control plane, they risk becoming an intelligence supplier inside someone else's system.
That can still be a very good business. It may not be the best enterprise moat.
Platform-first companies have the opposite problem. They may not always have the model everyone is talking about this month. What they do have is distribution, identity, admin surfaces, compliance posture, data gravity, procurement muscle, and existing presence inside the workflow. They are already sitting where the work happens.
That matters more than people want to admit.
If your AI can inherit the same identity model, security posture, document permissions, meeting context, business records, and admin controls the company already uses, you are not asking the enterprise to invent a new operating model. You are extending one it already trusts.
The hybrid players are trying to do both. Build or tune models, while also building the fabric around them. That is a hard road, but it is the right ambition. The winners will not be the ones with the most dramatic demo. They will be the ones whose AI shows up in the flow of work, grounded in the right context, governed by the right controls, and boring enough for a regulated enterprise to say yes.
That last phrase is doing a lot of work.
Boring enough to say yes is underrated.
Owning the workflow beats owning the model
The temptation is to think the model layer will keep all the value because it feels like the magic.
It probably will not.
The workflow is where value gets measured. Did the invoice get processed. Did the claim get triaged. Did the sales team find the right next step. Did the employee get onboarded. Did the analyst make the decision faster with better context. Did the agent complete the task without creating a mess someone else had to clean up.
Those outcomes depend on the model, but they are not created by the model alone.
They depend on whether the system can reach the data, understand the permission boundary, call the tool, handle the exception, and leave behind a record the business can defend. They depend on whether the AI is embedded into the work or bolted onto the side as a very articulate chat box.
The companies quietly winning this shift are the ones embedding into the systems of record, the productivity surfaces, the governance layers, and the automation paths. They are not always the loudest. They do not always have the flashiest benchmark slide. They are doing the less glamorous work of becoming the place where enterprise context is assembled and acted on.
That is where the moat is moving.
The next state probably looks less like one model answering one prompt and more like a workflow that can call several models, tools, and agents in sequence. One model reasons through the intent. Another classifies the record. Another retrieves the right context. Another drafts the action. An agent checks policy, another routes for approval, and the platform keeps the whole chain inside the right permission boundary.
In that world, the individual model matters, but it matters the way a processor matters inside a larger system. You care that it is capable, reliable, and cost efficient. You care less about the logo on the chip than whether the machine does the job. The workflow becomes the product. The platform becomes the control plane. The models become swappable capacity underneath it.
The next phase is control of context
Models will keep improving. The best ones will get faster, cheaper, more capable, and less strange around the edges. That is good. It will expand the set of tasks AI can handle.
It will also compress differentiation.
When every serious vendor can access capable models, the fight moves up and around the model. Who controls the context. Who owns the workflow. Who orchestrates the agents. Who governs the action. Who gives the enterprise a single place to manage what the AI can see, what it can do, what it did, and what it cost.
That is the next control plane.
The enterprise buyer will still ask which model is underneath. They should. But the better question is larger now.
What fabric does this model inherit?
Because that is where the real work happens. Not in the answer alone, but in the chain of systems, permissions, context, and actions that turns the answer into something the business can use.
The model stopped being the moat.
The moat is the fabric around it.
Views expressed are explicitly that of my own.