What Should Enterprises Own in Their AI Stack?

The next phase of AI transformation will be about what companies can learn from every decision it makes, as well as who owns that learning

Conversations surrounding AI have been focused on one main question: what can these models do? They were benchmarked, piloted, and marveled at, and it was the right question for the experimentation phase. However, it’s not the right question for where we stand today.

AI is now doing work, monitoring changes, preparing deliverables, and participating in key decisions in this next phase — the agentic enterprise — to move business forward rather than just answering questions. That shift is accelerating: IBM’s 2026 CEO Study found that CEOs expect the share of operational decisions made by AI to nearly double to 48% by 2030. Once AI is doing the work rather than merely describing it, the strategic question changes from “What can this model do?” to “What does my company learn every time it makes a decision, and who owns that learning?”

If enterprise value accrues one decision at a time, then a company’s most important asset in the age of AI is the loop that allows the organization to learn from every decision, specifically the context behind it, the evidence that led to it, the outcome that resulted from it, and the pattern that connects them. Own that loop, and every decision makes the next one sharper. Miss it, and you are automating activity without accumulating advantage.

By simply reframing it, almost everything changes about how a serious enterprise should architect its AI.

The Mistake of Betting Your Firm on One Model

The instinct has generally been to consolidate by picking a leading model provider, standardizing on it, and building everything on top of it. However, I think that’s a mistake, although not for the reason people usually give.

The usual objection is lock-in and pricing leverage. Those are real, but they are secondary. The deeper problem is about incentives and where your learning goes. Model providers are in the business of building the best general-purpose models in the world, and the way they do that is by learning from usage. When you route your firm’s proprietary decisions, data, and workflows through a single provider, you are handing the most differentiating material you have — the record of how your company thinks — to a vendor whose incentive is to generalize it. Your edge becomes an input to a commodity your competitors can also buy.

The frontier is also moving too fast to bet on any one model staying best for long. The right architecture should instead allow enterprises to route tasks to the model or specialized agent best suited for that task, while owning the agent harness and learning loop that capture how the business works and makes decisions. Models can change; that proprietary foundation remains yours and grows more valuable over time.

That proprietary foundation may also have measurable strategic value. IBM’s study also found that CEOs incorporate proprietary data and IP into custom AI models and agents expect 13% more of their 2030 revenue to come from products and services they do not offer today.

You Can’t Build Every Agent, and You Shouldn’t Try

Owning your harness, loop, and routing does not mean building every agent that runs on top of them. The lead enterprise agent — the one that understands your organization, your priorities, your proprietary data — is yours to build and yours to own. But the work of the business spans domains no single team can master: sales, HR, legal, finance, market intelligence, and more. Each of those will be served best by a specialized agent that is excellent at that one thing.

No enterprise is going to build a best-in-class agent for every function internally. It is not a resourcing race you can win. The winning pattern is a strong lead agent orchestrating a roster of specialized sub-agents, each one customizable to your firm’s requirements, each plugging into your harness through a clean interface. Yet enterprises are still early in building that kind of architecture, with a 2026 Deloitte survey finding that only 15% of organizations have scaled orchestrated, cross-functional multi-agent adoption.

So, the real work for enterprise leaders is twofold: build the north-star lead agent and the loop it feeds and then identify the winners among specialized agents and integrate them. Ownership where it compounds; partnership where it doesn’t.

A Specialized Agent Versus a Wrapper

A specialized agent has to bring something the enterprise cannot easily recreate. The durable advantage is proprietary data plus a purpose-built harness to retrieve and reason over it. An agent is only as good as what it can find and how well it understands what it found.

That is the bar enterprise leaders should use when deciding where to build and where to partner. If a capability depends on unique content, domain expertise, and years of purpose-built infrastructure, rebuilding it internally is unlikely to create an advantage. Focus those resources instead on the proprietary data, decision loop, and orchestration layer that are uniquely yours.

What the Winners Will Understand

The firms that win the agentic era will be those that architect around what makes them uniquely smarter over time. They will own the loop that allows them to learn from every decision, while surrounding it with the best specialized agents rather than trying to build everything alone.

The goal is not to own every piece of the AI stack. It is to own the pieces where your advantage compounds and partner for the rest. Connected is not the same as intelligent. In the agentic enterprise, intelligent means owning what compounds and partnering for the rest.