In 2023, Duolingo began building new AI features on top of GPT-4. The company didn't build the model. It built everything around it: the scenarios, curriculum and product experience that turn a general-purpose model into a language tutor, plus Birdbrain, its own system for judging what each learner can handle next. Morgan Stanley did the same, putting GPT-4 inside an internal assistant that searches the firm's own research for its advisors. In both cases the model was the commodity and the system around it was not.
Meanwhile, Gartner forecasts global AI spending will hit $2.67 trillion in 2026, a 49.5% jump from last year. Yet a RAND report cites estimates that more than 80% of AI projects fail, twice the rate of conventional IT projects. It raises an uncomfortable question: are enterprises investing in intelligence, or just buying IT software and calling it AI?
The Procurement Reflex
For most large organizations, AI adoption follows a familiar path. A vendor pitches a platform, procurement evaluates it against competitors, a contract is signed, and the tool is rolled out. This is how enterprises have bought technology for decades, and for most software it works well enough.
But AI does not work like ERP. You cannot configure it once and leave it running. The value of an AI capability depends heavily on how well it understands the context it operates in: the organization's data, its industry dynamics, its decision-making culture. A model arrives pre-trained on general patterns. The relationships and priorities that define how a particular business operates sit outside its reach unless the company puts them in. So the useful split isn't off-the-shelf versus tailored. The better rule is to buy the commodity layers and own the layers that create differentiated intelligence.
The Ceiling Nobody Talks About
Generic AI tools perform well in stable, data-rich environments. Geopolitical realignment, regulatory fragmentation and supply chain disruption have made that the exception. An off-the-shelf system cannot read the intent behind a policy shift, or weigh a trading relationship built on trust rather than contract terms. It cannot factor in what a senior executive with twenty years of regional experience would recognize immediately: that certain signals mean something different depending on who is sending them and why. A model alone can't close that gap.
Two companies doing the same basic job, answering questions from their own information with a language model, show the difference. Morgan Stanley tested its assistant against what the firm's experts would say before each use case went live, and OpenAI reports that over 98% of its advisor teams now use it. Air Canada's chatbot told a customer in late 2022 that a bereavement discount could be claimed after travel, while another page of its own website said otherwise. A Canadian tribunal held the airline responsible, finding it hadn't taken reasonable care to ensure the bot was accurate. One company could check answers against the people who knew the right ones. The other found out from a tribunal.
From Procurement to R&D
The organizations getting real value from AI aren't treating it as a purchasing decision. They treat it as an ongoing research and development capability, built, tested, refined and adapted continuously within the business. Bought on its own, AI is a product you deploy and hope fits. Built into your own systems, it learns from your operations, absorbs your institutional knowledge, and compounds in value the longer it runs.
In practice, four things are worth owning. The first is context: customer history, engineering knowledge, internal research and years of decisions by experienced employees, turned into something a model can use. The model can be bought; the context cannot. The second is workflow. Systems that sit outside how people actually work struggle to create lasting value, so the strongest ones are embedded in decisions people already make. The third is evaluation, the most overlooked part. Collect a couple of hundred real tasks, each with the answer your best people would give, and run every model and change against them. It also makes buying safer: when a new model ships, you know within a day whether it's better for you. The fourth is the feedback loop. Every correction ignored suggestion and escalation to a human shows where the system fails and what to change. Over time, that is harder to copy than access to the same model.
That's what treating AI as R&D means. You form a hypothesis, such as "giving the assistant three years of account notes will cut errors in renewal briefs," test it against your evaluations, release it to a small group, study what breaks and go again. Until recently, building this in-house meant large teams and significant capital, which locked most organizations into procurement by default. That constraint is lifting: foundation models can now be rented, leaving the layer around them as the part worth building.
This isn't an argument against external tools. Document processing, transcription and routine automation are fine to buy. The danger is applying procurement logic to strategic AI, the systems that inform decisions about risk, markets and competitive positioning. Those systems need to understand your world, not the average of everyone else's.
The Compounding Question
The real cost of the procurement approach never appears on a balance sheet. It is the opportunity cost. The enterprise next door that started building around its models eighteen months ago has systems that understand its operations deeply enough to surface insights no vendor could replicate. That advantage compounds. A competitor can buy the same licenses next week. They cannot buy your corrections, your tested workflows or your evaluations.
Enterprise leaders in 2026 are past the question of whether to invest in AI. What remains unanswered is whether they are building a capability that gets smarter as their business evolves, or buying a product their competitors may acquire. One of those strategies compounds, the other has a shelf life.
