- What Gets Sold as AI-Ready Versus What AI Actually Needs
Sometime in the last two years, “AI-ready” became the most expensive phrase in enterprise technology. Boards approved it. Budgets followed. Initiatives launched. And then, most of them stopped producing anything measurable.
This is not a niche problem. MIT’s Project NANDA, after reviewing more than 300 publicly disclosed AI initiatives and interviewing representatives from 52 organizations, found that 95% of organizations deploying generative AI are getting zero return on the investment. Not marginal returns. Zero. S&P Global Market Intelligence reported that 42% of companies abandoned most of their AI initiatives in 2025, up sharply from 17% in 2024, with the average organization scrapping 46% of its proofs-of-concept before they ever reached production. The RAND Corporation, after interviewing 65 experienced data scientists and engineers, concluded that more than 80% of AI projects fail, which is more than double the failure rate of non-AI technology projects. Among the root causes they identified: organizations lack the necessary data to adequately train effective models, and inadequate infrastructure makes it harder to deploy completed AI work into production.
The scale of spend running alongside these numbers makes it harder to ignore. Amazon announced $200 billion in capital expenditure for 2026, almost entirely for AI infrastructure. Every major cloud provider is in a full sprint. The infrastructure investment is real, and so is the gap between what’s being built at the platform level and what most enterprises actually have underneath their AI initiatives.
That gap has a specific shape, and I’ve spent 25 years inside it. As Chief Data and Engineering Officer at Availity, the nation’s largest real-time health information network, I’ve led a full cloud migration to AWS, overseen early enterprise adoption of generative AI through Amazon Bedrock and Amazon Q, and made the kinds of hard infrastructure decisions that most organizations are still debating. What follows is what I learned.
The Complexity That Accumulates
When we completed our cloud migration, I was confident we were positioned for AI. We had modernized. We were on AWS. We had tooling. Looking back, we had moved years of accumulated complexity from one environment to another and called it transformation.
Every layer in that stack was somebody’s solution to a real problem. Each one made sense at the time it was built. And collectively, they had become a system that nobody fully understood, that required tribal knowledge to operate, and that was going to fight any AI initiative we tried to run on top of it. That is how organizations end up with eight middleware layers and four conflicting data schemas. Nobody plans it. It accumulates, decision by decision, deadline by deadline, under pressure.
Complication gets rewarded organizationally. Your pay grade scales with how many systems, processes, and dependencies you manage. The person who eliminates the problem often gets less credit than the person who builds the team to manage it. That incentive structure quietly shapes everything.
AI does not accommodate that. It performs relative to what it is working with, and it makes the consequences of accumulated complexity visible faster than anything that came before it.
What We Actually Did
When we took a hard look at the stack during our migration, I asked a question that changed how we approached it: what would we actually build today, from scratch? The honest answer was not what we had.
IBM was asking us to keep OpenShift. Cloudera was making the case for their platform. StreamSets argued we needed them. Informatica made the same pitch. Every vendor had a reason we couldn’t function without them.
We removed all of it.
Not out of hostility toward any of those tools, but because the honest answer to the from-scratch question left no room for them. They existed to support decisions made under different constraints, at a different scale, with different goals. Carrying them into an AI context meant carrying all of that history too. Simplification is not an IT project. In an AI context, it is the strategy.
We are not done simplifying. We had a hard deadline to exit the data center, so I made priority decisions and some things waited. The work continues. What I can say is that AI has turned out to be the most powerful simplification tool I’ve worked with, not because it adds capability on top of what exists, but because it allows you to subtract. We are retiring entire features and replacing them with AI-driven solutions. Fewer systems. Less surface area. Less to hand off to the next engineer who inherits this stack.
What the Foundation Actually Requires
Readiness for AI lives several layers below whatever tooling or platform sits on top of it. It requires data that is accurate and complete for the specific use case being deployed, not just theoretically clean in a general sense. It requires genuine accessibility, meaning AI systems can actually reach the data they need with appropriate permissions and acceptable latency. It requires governance: clear ownership, documented lineage, privacy controls, and compliance infrastructure that holds when the model touches regulated data. And it requires structural integrity built deliberately over time, not retrofitted because an AI initiative is now on the roadmap.
Informatica’s 2025 CDO Insights survey, based on responses from 600 data leaders globally, found that 43% cite data quality, completeness, and readiness as the leading obstacle preventing AI pilots from reaching production. Nearly two thirds of those same leaders said they have been unable to transition even half of their generative AI pilots into production. The bottleneck is not the models. It is what the models are being asked to work with.
This is also where the strategic picture clarifies. The new moats in enterprise technology are compounding data assets, deep integrations, and compliance positioning that takes years to earn. A feature can be replicated over a weekend. An AI agent can be built on a competitor’s API in days. Eight years of governed, contextualized, high-quality operational data cannot be purchased or assembled on a short timeline. That is the actual strategic asset, and it is also the actual prerequisite for AI output that is trustworthy enough to act on.
AI is an amplifier. Organizations with clean, compounding data foundations will find it accelerates their advantage. Organizations carrying years of accumulated complexity will find it accelerates that too, just in the wrong direction.
Where to Start
AWS CEO Matt Garman, speaking at re:Invent 2025, described AI inference as a fundamental new building block in computing. “The world invented a new Lego,” he said. Before inference, developers had compute, storage, and databases. None of those could independently make decisions or take actions. That has changed. But that building block performs exactly as well as the data foundation it sits on.
If your AI initiative is stalling, or if you’re about to launch one, start with the question I asked during our migration: what would you actually build today, from scratch, if the goal were to support AI at scale? Audit the gap between that answer and what you currently have. That gap is the real infrastructure problem, and it is worth mapping honestly before another budget cycle passes.
Define success in business terms before the first sprint begins. Revenue impact, cost reduction, a workflow change that shows up in the numbers. Those are the outcomes worth building toward. Treat data readiness as a prerequisite rather than a parallel workstream. The RAND research is clear that organizations which skip this step are not dealing with a model problem. They are dealing with a foundation problem, and the model will tell you so in production.
AI does not lower its standards for the data it works on. What it does is remove any remaining ambiguity about whether those standards were met.
Michael Privat is Chief Data and Engineering Officer at Availity, one of the largest health information networks in the United States. He writes about AI, engineering leadership, and enterprise technology on Substack at michaelprivat.substack.com. Follow him on LinkedIn at linkedin.com/in/michaelprivat.
