What the 5% of Companies Getting Positive AI Returns Are Doing Differently

Why data readiness and modernization determine whether AI investment compounds or evaporates

A few months ago, a CFO at a mid-market software company looked at the AI line in the quarterly budget review. The company had spent several million dollars across the prior twelve months on generative AI tooling, agent platforms, Copilot subscriptions, and a small team of engineers tasked with building internal AI workflows. They wanted to know what the company had to show for it. The honest answer, once the room got past the demo footage and the productivity surveys, was very little. A few pilots had been quietly retired. Others were stuck at the proof-of-concept stage. The one that made it into limited production was producing outputs that the legal team would not let anyone act on without human review.

The CFO closed the meeting with a sharper version of the question every executive in the room had been avoiding: what would have to change for the next twelve months to look different?

Similar conversations are happening in boardrooms across the country. The instinct is to treat this as an AI maturity problem, but it is actually a data and infrastructure problem that the AI program inherited from work that nobody got around to doing. That pattern shows up in the numbers. As McKinsey recently reported, 88% of organizations now use AI regularly in at least one business function, yet the companies investing most aggressively are largely the same ones reporting the smallest returns. Adoption is widespread, but the foundation underneath it is not.

The AI Pilot Loop Is a Data and Infrastructure Problem

MIT's NANDA Project found that despite $30-$40 billion in enterprise AI spending, 95% of organizations report no measurable business return. McKinsey calls this the gen AI paradox. The diagnoses are consistent across both, claiming that the technology is not the bottleneck. The integration is. 

Pilots stall in the same places every time. The data the agent needs lives in fifteen different systems, each with a different access model. The infrastructure is fragmented across cloud accounts spun up by different teams with different governance assumptions. The instrumentation needed to evaluate whether the agent is actually working was never built into the systems it's supposed to operate against. These are not AI problems. They are foundation problems the AI program inherited from work that nobody got around to doing, and that distinction is what the CFO's question was really pointing at. 

Why misdiagnosis is so common

The reason organizations misdiagnose this as an AI problem is structural. The AI budget is new, the AI vendors are loud, and the AI team is highly visible. The foundation problems are old, the people who could fix them are usually busy keeping the lights on, and the work itself does not produce headlines. When something breaks, attention flows to the new and visible work, even though the actual cause lies in the older and quieter work beneath it. This is exactly why so many organizations are entering their third year of AI investment with the same set of unresolved questions they started with.

The Cost Economics Make The Prerequisite Work Urgent

The misdiagnosis would be expensive enough in stable pricing conditions. The current pricing environment is anything but stable, and that is where the prerequisite work becomes genuinely urgent.

The current generation of LLM and agent products is operating at a substantial loss. Sam Altman publicly acknowledged that OpenAI is losing money on $200-per-month Pro subscriptions because users consume more compute than the pricing model accounts for. Internal projections reported by the Wall Street Journal recently revealed that OpenAI expects to lose roughly $14 billion in 2026 alone, with operating losses climbing toward $74 billion by 2028 before the company projects a return to profitability around 2030. That gap between revenue and the underlying cost to serve is currently funded by venture capital and provider strategy. It will not be funded forever.

When pricing eventually rises to reflect the reality of compute, the economic exposure for enterprises will lie along two axes. The first is direct subscription cost, where a $200 line item becomes a $500 or $1,000 line item without the underlying value proposition changing. The second is infrastructure efficiency, where companies running AI across fragmented systems will burn dramatically more tokens to accomplish the same work as companies running AI across consolidated, well-instrumented systems.

What this means for AI program economics

The infrastructure efficiency dimension is the one most executives have not modeled. An AI agent that has to reconcile data across five different schemas burns tokens reconciling. An agent that has to query four different monitoring tools to assemble a coherent picture of the system state burns tokens querying. An agent that has to navigate inconsistent authentication patterns burns tokens when authenticating. Each of these inefficiencies is invisible at today's subsidized prices and will be very visible at tomorrow's market prices. 

Companies that consolidated, standardized, and instrumented their infrastructure before deploying AI will absorb the shift in pricing. Whereas companies running AI on fragmented foundations will see their AI line items double or triple without producing more output. We see this dynamic in our customer conversations almost every week, and we view it as the single most underweighted risk in enterprise AI strategy today.

The Conservative Domains Are Where The Prerequisite Work Matters Most

Much of today’s marketing narrative makes it sound like AI is being adopted evenly across all functions within the enterprise, but the reality looks different from inside production environments. The most conservative parts of any organization, the ones running production infrastructure, financial systems, and regulated workloads, are actually the last to adopt AI, and the lag is rational. The cost of an AI mistake against a customer-facing database, a payment rail, or a healthcare workflow is catastrophic, and the data foundations in those domains are often the weakest in the entire company because the systems were built before anyone imagined AI would need to reason across them.

This pattern also shows up consistently in our customer base. Product organizations that have spent years talking publicly about their AI initiatives are typically talking about AI within their products, rather than AI running their products. Operating AI in production is a different problem, and it is one most consulting organizations will not touch. The standard model is to make recommendations and work in pre-production contexts, then hand the work off to the customer for them to bring it into production. We take a different position. EverOps engineers operate within our partners' production. That commitment shapes every aspect of our approach to AI readiness work. We build AI deployments that work in production from day one, because we are the team standing behind them when they do.

The Companies Pulling Ahead Are Doing The Foundational Work

The pattern across companies that are actually generating measurable AI returns is consistent. Data consolidation. Infrastructure modernization. Observability buildout. Governance frameworks. None of these appear on the analyst slide on AI trends, yet all of them appear in the operational history of the small percentage of companies that are extracting real value from AI investment.

Our work with Zendesk is a representative example. The engagement resulted in a 70% reduction in infrastructure costs across 2,500 EC2 instances, delivered through immutable AMI workflows, automated instance recycling, OS modernization, and migration to auto-scaling groups. The immediate return was financial. The less obvious return is what that work sets up for everything that comes next.

None of it was designed with AI readiness in mind, but it became AI-ready anyway.

That outcome is not a coincidence. The practices that produce efficient, consolidated, well-instrumented operations are the same ones that create an environment AI can actually reason across and deploy into. A modernized infrastructure does not need to be rebuilt for AI. It is already the substrate AI requires.

The same pattern shows up across our work on cloud cost optimization, observability foundations, and platform modernization. Partners often come to EverOps with a business problem, uncontrolled cloud spend, unreliable incident response, or fragmented telemetry. We solve it by having our senior, embedded engineers operate directly in their production environment. The byproduct of that work is a foundation that supports whatever AI strategy leadership eventually wants to run on top of it. 

We now recommend that customers view AI readiness as a downstream benefit of modernization work, rather than treating AI strategy as a separate program that ignores the underlying substrate. That framing has changed how leadership teams evaluate the ROI of foundational engagements, and it is the same framing that separates the companies generating real AI returns from the ones still waiting for their pilots to graduate.

AI Is A Tailwind For Teams That Built The Foundation

As software gets commoditized by AI, the value in technology services shifts toward the people who can understand a specific environment, customize the pieces, configure them correctly, and operate the resulting system. Off-the-shelf SaaS platforms cannot deliver this. Quickly generated tooling also cannot deliver this. The work requires deep operational knowledge of a specific company, accumulated experience across many customers, and the discipline to translate generic AI capability into specific organizational outcomes. The services layer becomes more valuable as the underlying software becomes more commodified, not less.

The companies that laid the foundation will capture the upside of this shift. Their data is clean enough for AI to use. Their infrastructure is consolidated enough for AI to run on. Their governance is mature enough to allow AI to operate within. Each of these conditions makes the next AI capability cheaper to deploy and more valuable when deployed. The companies that skipped the foundation will face the opposite trajectory. Each AI deployment will be more expensive than it needs to be and less valuable than it could be, because the work the foundation should have done is now being done inefficiently inside the AI itself.

The compounding effects of these two paths will be visible in financial statements within eighteen months. Leadership teams that recognize this early will redirect AI investment into the foundation that determines whether the rest of the program works.

Let EverOps Build The Foundation Your AI Strategy Needs

If your AI program is stalled in pilot purgatory, the problem is rarely the model and almost always the foundation underneath it. Fragmented data, sprawling infrastructure, inconsistent governance, and unmodernized platforms produce AI deployments that are too expensive to scale and too unreliable to trust. 

EverOps embeds engineers directly into your environment to do the foundational work that makes AI strategy executable. Our modernization, observability, and cloud cost engagements deliver immediate financial returns and produce the infrastructure substrate your AI program depends on to scale.

Talk to our team about an AI readiness assessment for your current environment and find out exactly what it would take to make your AI investment compound rather than evaporate.

Frequently Asked Questions

Why are so many enterprise AI pilots failing to produce measurable returns?

Most pilots fail because the data and infrastructure foundations underneath them were never built to support AI workloads. The technology is not the bottleneck. Integration with fragmented data, sprawling cloud accounts, inconsistent instrumentation, and immature governance is. MIT's NANDA findings point to the same root cause from a different angle. Foundation work, not model selection, separates the AI programs that generate returns from those that produce write-offs.

Is AI pricing going to rise significantly?

The current pricing reflects substantial provider subsidies. Public statements from OpenAI's CEO confirm the company is losing money on its $200 ChatGPT Pro subscriptions, and internal projections reported by the Wall Street Journal show OpenAI expects operating losses of roughly $14 billion in 2026 alone. When pricing eventually reflects actual compute costs, enterprises running AI on inefficient infrastructure will see disproportionate cost increases because their AI consumes more tokens to accomplish the same work as enterprises running AI on consolidated, well-instrumented infrastructure.

Why is the conservative-domain reality important to AI strategy?

The most operationally critical systems in any enterprise are also the last to adopt AI, because the cost of an AI mistake against production infrastructure, financial systems, or regulated workloads is catastrophic. The data foundations in these domains are also typically the weakest, because the systems predate AI by years or decades. Closing that gap is where the highest-stakes prerequisite work lives, and it is also where most AI strategy decks underweight the work required.

What does data readiness for AI actually mean?

Data readiness for AI means that the data an AI system needs to reason across is complete, consistently structured, current, and governed by a clear policy about what the AI can access and what it can do with what it finds. Most enterprises fall short on at least two of these dimensions. The data lives in too many places, follows too many naming conventions, lags behind reality, or is governed by access models designed for human users rather than for AI agents that reason at machine speed. Closing those gaps is where the foundational AI work lives, and it is what determines whether AI deployment produces a meaningful return.

How does EverOps approach AI prerequisite work differently?

We embed engineers directly into customer environments and operate in production rather than only making recommendations. That commitment changes how we approach data, infrastructure, and platform modernization. We do the work that makes AI strategy executable, with accountability for outcomes that most consulting firms decline to take on. Our cloud modernization, observability, and FinOps engagements deliver immediate financial value and produce the substrate AI deployment will eventually require.

How long does an AI prerequisite typically take?

Timelines vary by the starting state. Partners with relatively consolidated infrastructure and reasonable data hygiene can be AI-ready in three to six months of focused foundation work. Customers with significant cloud account sprawl, fragmented telemetry, or inherited M&A complexity should expect six to eighteen months. For examaple, our Zendesk engagement delivered a 70% cost reduction across 2,500 EC2 instances through automation and modernization work that, by virtue of how it was done, also produced the kind of substrate AI deployment depends on.