Businesses are not disappointed in AI. They are disappointed in a shopping strategy that was never going to work. You cannot purchase your way into an intelligent company one tool at a time, because intelligence does not live in tools. It lives in context. And context is the one thing no vendor can sell you.
Here is the pattern I keep walking into.
A company decides it is time to get serious about AI. So it buys. An AI feature inside the CRM. An AI notetaker for meetings. An AI assistant bolted onto the help desk. A chat subscription for the team. Each one demos beautifully. Each one earns its line on the budget.
Eighteen months later, leadership is staring at the stack and asking the same question: why does none of this feel like it changed anything?
The tools work. The workflows don't.
The notetaker knows what was said in the meeting but not who the client is. The CRM assistant knows the deal stage but not what was promised on Tuesday's call. The help desk bot answers policy questions from a knowledge base nobody has touched since the reorg. Every tool is smart inside its own walls and blind past them.
And the diagnosis leadership reaches is almost always the same: our data is siloed.
That diagnosis is half right. The wrong half is the expensive one.
The problem is not where your data lives. It is that no system sees enough of your business to be useful.
Here is my position, stated plainly. AI runs on context. Give a capable model the full picture and it does remarkable work. Give it a keyhole and it produces confident, polished, useless output. Most companies have bought ten keyholes.
This is why the disappointment is so widespread and so uniform. It is not a model problem. The models are extraordinary. A bolt-on tool fails for the same reason a brilliant new hire fails when nobody gives them access to anything: not for lack of talent, but for lack of context. The difference is that the new hire earns context over months. The bolt-on tool never does. It is a first day that repeats forever.
Everyone wants Jarvis. Few people look at how he was built.
When leaders describe what they want from AI, they rarely describe a tool. They describe Jarvis, the system from the Iron Man films. You think a half-formed thought out loud, and the right thing happens. It already knows the situation, the history, the stakes, and your preferences. It drafts, schedules, flags, and prepares, and it brings you the decisions that are yours to make.
Notice what makes Jarvis work in those movies. It is not the voice, and it is not the raw intelligence. It is that the system sees everything that matters and knows the man completely. Tony Stark never opens seven applications to assemble the picture himself. He never explains who a person is or what happened last week. The context is standing, so the intelligence lands.
Jarvis is not a gadget Stark bought. Jarvis is the accumulated context of one operation, wired together, with intelligence layered on top.
That distinction is the whole game, so let me take it apart. There are three layers between the AI stack you have and the system you want, and knowing which layer you are buying is the difference between an operating advantage and a subscription graveyard.
Layer one is connection. It is real, and it is getting cheap.
Connection is the plumbing: the pipes between your email, calendar, accounting system, project tools, and CRM. This layer is being commoditized as we speak. Open standards now let AI systems plug into business software the way appliances plug into a wall. The major platforms are racing to wire their own products together for you.
So when a vendor's pitch is "we connect to all your systems," understand what you are being offered. Plumbing. Necessary, genuinely useful, and less differentiated every quarter. Anyone can connect.
Connection is also the smallest layer. Connected data is not understanding, any more than a phone line is a conversation.
Layer two is context. No vendor can ship it, and it decides everything.
Context is what the data means in your business. This client is mid-renewal and the relationship is delicate. That invoice is an exception we approved in March. This email needs the owner's eyes no matter what it says. That number from the field is wrong because the crew logs time a day late. We never send anything to a client without a human approving it first.
None of that lives in any of your systems. It lives in your people, your habits, and your unwritten rules. Context is your company's judgment, and today it is stored almost entirely in the heads of your most senior people. Institutional knowledge that lives in one person's head is not institutional knowledge. It is institutional risk.
This is the layer that has to be built, because the vendor does not have your context and never will. Building it means sitting with the workflows, naming the rules, encoding the exceptions, and deciding where a human must stay in the loop. It is slower than swiping a card. It is also the asset.
And it is worth saying clearly: this is not a data cleanup project. Stanford's Digital Economy Lab studied 51 successful enterprise AI deployments and found only 6 percent began with AI-ready data. The winners did not centralize everything into a warehouse first. They wired access to the messy systems they had and encoded judgment on top. If you are waiting for clean data to start, you are waiting for a prerequisite that the successful deployments skipped.
Layer three is operation. The system is never finished, and someone has to run it.
An AI system wired into your business is not a project with an end date. It is a system with an operator.
Your business changes weekly. Clients arrive, rules evolve, a policy shifts, an edge case appears that no one predicted. A system carrying your context has to be corrected, tightened, and audited as it runs, the way any critical operation is. I update my own system almost every day, and after a year of running it, the maintenance is not a burden I tolerate. It is where the compounding happens. Every correction makes the system smarter about the business, permanently.
Anyone selling you an install-and-walk-away AI solution is selling layer one in a costume. And if no one in the room can name the operator, you do not have an AI strategy. You have an unattended process wearing your company's name.
Bolt-on tools fail because they sell layer one and leave you to discover the other two on your own.
This is the honest answer to the siloed data complaint. Siloed data is the symptom you can see from the executive chair. The disease is that nobody in the building owns context and operation. The vendors will not own it, because they cannot. Your IT team will not own it by default, because it is not a technology question. It is a question about how your company works, and that question belongs to leadership.
Underneath the tooling frustration, something bigger is available here. The company that builds these layers ends up with its judgment documented and executable: how decisions get made, what the exceptions are, where the approvals sit, what good looks like. That asset outlives any single tool, any vendor relationship, and any employee's tenure. The automation is the visible return. The encoded judgment is the durable one.
What to do with this, starting now.
Not a framework. Four moves.
- Stop asking which AI tool to buy. Start asking which workflow to wire. Look for the workflows that cross the most systems and consume the most senior hours. Billing that touches accounting, email, and project records. Client onboarding that touches five departments. That intersection is where context dies today, and where the return lives.
- Pick one workflow, not the whole organization. Wire its systems together, encode the rules and approval gates, and run it until it is trusted. Your existing tools do not get ripped out. They become sources feeding a system that finally sees across them. Then take the next workflow. The foundation you build for the first one serves every one after it. Org-wide is the destination. It was never the first step.
- Interrogate every AI purchase with one question: what can it see beyond itself? If the answer is its own silo, you are buying another keyhole, whatever the demo looks like.
- Name the operator before you build anything. A person accountable for the system's judgment and a team accountable for keeping it running. If you cannot fill in those names, you are not ready to build, and knowing that now is cheap.
I will say this once, because it is where my conviction on this comes from: I run my business on a system built exactly this way, one workflow at a time over the past year, and it is what my team and I now build inside client organizations. The reason I trust these layers is that I live on top of them every morning.
The companies that end up with their Jarvis will not be the ones with the biggest tool budget. They will be the ones that stopped shopping for intelligence and started building context.
Jarvis was never a product. He was built, one system at a time, by someone who knew the operation. Yours will be too.
