What is AI? It is a tool, an environment, or a new form of intelligence? Those ideas are all interesting fodder for engaging—and sometimes enraging—dinner conversations, but none of them clarifies the work required for organizations to confidently deploy and manage these systems. That may, in some cases, be intentional. After all, if the systems are so complicated that you can’t reasonably control them, but so powerful that you can’t afford not to use them—what’s a company to do? Perhaps shuffling them into a general-purpose category will do the trick.
But is AI a general-use technology? Every time someone calls AI—and generative AI (GenAI) specifically—a general-purpose tool, I imagine a spork. Will a spork work in a pinch? Yes, it does. But does a spork ever really work well? No, it does not. And I can confidently surmise that 9 out of 10 diners will agree.
To be fair, I suspect those folks who are invoking this terminology hope to conjure the image of a Swiss Army knife. What this analogy can fail to account for is that multi-tools aren’t a singular thing. They don’t rely on a single tool but provide multiple, discrete options crafted with an explicit job in mind. Some generalize more than others. A basic paring knife is ubiquitous across multi-tools while other fixtures are strictly specialty offerings. Urban picnicker or backwoods survivalist? There is a multi-tool for you—it’s just not the same one.
In this sense, the language of AI as general-purpose tech is somewhat coherent. As long as organizations recognize that the singular, umbrella term AI references a spectrum of raw materials from which multiple products may be forged. Keeping that firmly in mind, this perspective can be helpful in shifting from the “one AI (or one AI provider) to rule them all” mentality to understanding AI as a portfolio. A portfolio in which both raw AI materials and the offerings derived from them feature as discrete products. Each of which must succeed on its own merits. This brings us to the thorny question of trust.
Is AI adoption (or lack thereof) a trust problem? Here, I’d say yes, but with a caveat. Namely, the problem of trust is that humans often stubbornly refuse to conform and adopt to new tech. It’s a condition that AI purveyors like to attribute solely to fear, illiteracy, or habitual conditioning. In reality, the tech is, at times, the problem. Not because an AI system chooses to misbehave or is too powerful to be controlled, but because it is deployed in ways that are inappropriate, unmanageable, or unnecessary.
None of this is the tech’s fault, exactly. The tech is what it is: a product with discrete characteristics. The lack of specificity or boundaries for a foundational AI offering is, in and of itself, a core feature of the product. The ways in which a human instinctively responds to confidently presented language are also a feature of the product. A feature that is deliberately reinforced through the design of AI products.
Is a lack of trust in AI ever simply a human problem? Yes. But only if the system in question is trustworthy to begin with, which is entirely a property of the product itself.
Why not talk about AI systems as products? That large language model (LLM) or, to be fancier, frontier model you are using—it’s a product. The AI componentry you are building into your business processes? Raw products. The applications being generated with, or on top of, an algorithm? They are also products. Some products may be more singular than others: a piece of content versus a customer-facing chatbot versus an agentic workflow. However, they are products.
As such, employing the language of product development and procurement helps ground fanciful AI discussions in operational reality. AI systems have some unique characteristics. But cloaking our understanding and governance of them under the veil of unquantifiable uniqueness only serves to obscure their nature. Building on the well-established lexicon by which we understand, define, and assess products brings much-needed clarity.
Will all AI products neatly conform to existing categories? No, but that’s OK and even expected. Asking the question is the first step toward knowing which outliers are distinctions with a difference or merely distractions. That is the point of the exercise. In defining where boundaries do and do not exist, when favored outcomes can and cannot be assumed or guaranteed, we learn something important about what it is we are choosing to build, AI or otherwise. Customer need, reliability, stability, accessibility, warrantability, tolerance, assurance, and risk— these are just a few characteristics we habitually assess for products of all stripes. Why not AI, too?