I was in a conversation recently with a senior executive who said something I've heard a dozen times in the last eighteen months: "We're going to develop our AI strategy around our specific use cases."
Product development frameworks at their best have always followed this method: Define the problem first, then apply the technology.
The problem is that it's exactly backwards, and the organizations that figure this out late are going to spend the next few years holding expensive, half-built implementations of problems that either got solved by a foundation model vendor for $20 a month, or turned out to require infrastructure they never invested in.
Here's what's actually going on:
The Velocity Problem
AI is not a stable technology. The pace of development of new models, new interaction paradigms, and new capabilities is measured in weeks and months, not years.
The average enterprise software implementation cycle runs six to eighteen months from scoping to deployment. These timelines exist for good reasons: procurement, integration, change management, testing, training. These are key parts of the product development lifecycle that can't (read: shouldn't) be compressed away.
But here's the arithmetic problem: If your implementation cycle is twelve months and the technology's meaningful capability shift is three to six months, you are structurally guaranteed to be building for a version of the technology that no longer represents the frontier by the time you ship. The use case you scoped in Q1 may be commoditized, superseded, or architecturally awkward by Q4.
This isn't a hypothetical—we've already watched it happen with document summarization, code completion, basic customer-facing chatbots. All of these were serious internal AI initiatives at major organizations eighteen months ago.
All of them are now table-stakes features in off-the-shelf tools, and the teams that built bespoke implementations are now stuck maintaining something they could license today for a fraction of the cost.
Use-case-first AI strategy doesn't just risk being wrong about the use case. Its dependency on known-knowns structurally ensures you're optimizing for where the technology was, not where it's going.
The Misclassification
The core of this article today doesn't focus on velocity though: There's a deeper problem underneath the velocity issue, and it has to do with how organizations are categorizing this investment.
AI is being funded and governed as capital expenditure. That means a specific asset, a specific expected return, a specific accountability horizon. The board approves the budget, the team builds the thing, and eighteen months later someone asks whether it delivered on the expectations set out for it.
AI in its current state doesn't behave like CapEx, though. It behaves like R&D.
R&D investment is justified by optionality, not by a predetermined output. The goal of R&D investment is to develop the organizational capability to discover and build things you can't fully specify in advance. The return is emergent, meaning it comes from iteration, learning, and being positioned to move when the right opportunity appears.
The problem is that "we're running an R&D program" is a very hard thing to sell to a board that wants AI ROI on a twelve-month horizon. So instead, organizations run a program that is ultimately self-defeating: They sell AI investment as CapEx to get it approved, then try to run it like R&D without saying so.
They pick a use case that sounds concrete enough to justify the budget, build around it, and hope that the underlying organizational learning is enough to show something defensible at review time.
The AI hangover we're starting to see across the market is coming to fruition because the accountability structure that governs most internal AI projects has been wrong from the start.
The use case investments haven't delivered on CapEx terms, and the R&D learnings are happening but not being made legible. In organizations where this happens, everyone is left holding an expensive lesson with no clear path forward.
The Wrong Layer
Both the velocity trap and the misclassification trap share a common root: The investment being made at the wrong layer of the stack.
Think about AI capability in terms of layers:
At the top are use cases: specific applications, specific workflows, specific outputs. These are the most visible layer, which is why they attract the most attention and drive most budget conversations. They are also the most brittle and most directly exposed to model changes, vendor commoditization, and shifting interaction paradigms.
In the middle are harnesses and tooling: The firm-specific middleware that connects models to your data, systems, and workflows. These are more durable than use cases, but still tightly coupled to how a specific generation of models works. A harness built for one architecture may be awkward or obsolete for the next.
At the bottom is the data platform: the infrastructure that makes your organization's knowledge accessible, normalized, documented, and connectable. This layer is use-case agnostic. It doesn't care which model you're using, which vendor you've chosen, or which application turns out to matter. It just makes everything above it work (or not work).
The best data platform keeps working regardless of what changes above it.
Most AI investment is concentrated at the top two layers, while the durable value is almost entirely at the bottom. Counter-intuitively, the first two layers look like capex investments but due to market dynamics are in reality far more speculative, resembling R&D in practice, while the AI platform looks like a cost center, but when built correctly operates as a functional base for all other technology implementations, present and future.
The Universal Adapter
The question most organizations are asking is: What will AI do for us? That's a use-case question. It assumes you can identify the right application in advance and build toward it.
The better question is: How quickly can we plug into whatever AI becomes next?
This is the question solved by the platform. It doesn't require you to predict which use case will matter. Rather, it requires you to build the organizational infrastructure that makes you fast and flexible enough to capitalize on opportunities as they emerge.
At its core, an AI data platform functions as a universal adapter: A data foundation that any model, any application, and any workflow can connect to cleanly.
Data becomes accessible rather than siloed, documented rather than tribal, and normalized instead of being inconsistent. These traits make your organizational information—about customers, operations, products, performance—legible to the tools that need to use it.
This is not a glamorous description. It doesn't have the forward-looking energy of "deploying an AI agent," "developing a proprietary harness," or "building a custom LLM." But it's the thing that makes all of those downstream investments actually work. And critically, it's the thing that makes them reusable—so when the next capability shift happens, you adjust your heading rather than starting over from scratch.
The Resolution
What makes the platform investment argument genuinely useful rather than just intellectually correct is that it solves both traps at once:
It resolves the velocity problem because it's use-case agnostic. Instead of betting on a specific application, you're betting on your ability to move fast when the right application appears. Velocity becomes an asset rather than a threat.
It resolves the misclassification problem because the platform is a genuinely durable asset: It is defensible as CapEx on its own terms, not as a proxy for a use case that may or may not deliver. You can honestly represent it to a board as a capital investment, because that's what it is. And underneath that honest representation, you're running something that functions exactly like R&D infrastructure: fast, cheap iteration across whatever use cases emerge, with learning that compounds rather than expires.
You're not choosing between selling it as CapEx and running it like R&D. You're investing in the one piece of that stack that is true for both, simultaneously.
The Question to Ask
The question we should be asking at the executive level is: Are we structured to move fast enough to keep up with the rate of technological change?
From where we sit, the answer is clear for 90% of companies: No.
If your answer depends on having picked the right use case in advance, you're not structured to move fast. You're structured to be right once. In a technology environment moving at this pace, being right once isn't enough—if you're even right at all.
Companies have historically underinvested in their data platforms, treating them as cost centers rather than strategic assets, and the miscategorization of this asset class is the reason why companies overlook the potential within the platform in favor of downstream use-cases.
It's rare that we as humans take the time to think about what it is we're standing on. Whether it's the shoulders of giants or the global stage, in the end it's just another platform.
Snowpack Data's AI platform engineers help organizations build the data infrastructure that makes AI investment durable. If you're evaluating your AI platform strategy, we'd like to talk.
