Executive Analysis • Finance & Technology

By Eric Ford  |  Real World Leadership   •   Estimated read time: 18 minutes

Enterprise AI spending hit $252 billion in 2024. Over 80% of projects failed to deliver their intended business value. Those two facts belong in the same sentence, and most budget conversations don’t put them there.

What the Announcements Don’t Tell You

The press releases all sound the same. A company commits to a nine-figure AI investment. A board approves an accelerated AI roadmap. A CEO assures analysts that the company is “fully embracing” the technology. The announcements come fast. The results come slowly. And the gap between them is where real money goes to disappear.

Total corporate AI investment reached $252.3 billion in 2024, up 44.5% year-over-year.¹ Gartner projects worldwide AI spending will hit $1.5 trillion in 2025.² Meanwhile, McKinsey’s 2025 State of AI survey found that only 39% of organizations report any EBIT impact from AI at the enterprise level.³ Over 80% of respondents reported no meaningful bottom-line improvement despite adoption. BCG’s September 2025 update of the same question found that 60% of companies generate no material value from continued AI investment, and only 5% create substantial value at scale.⁴

Those numbers should give any CFO pause. Not because AI doesn’t work, but because the conditions required for it to work are more demanding than most organizations acknowledge when they’re approving budgets.

This article is not an argument against AI investment. It’s an argument for going into it with accurate information about what it actually costs, where the returns actually come from, and what separates the 5% that generate real value from everyone else spending real money to generate almost none.

Getting Started Costs More Than the Software License

The single most common mistake organizations make when budgeting AI initiatives is treating the vendor contract as the primary cost. It’s not. In many cases, it’s not even close to the largest cost. Here’s what the full picture looks like.

Infrastructure. Generative AI workloads are compute-intensive in ways that traditional enterprise software is not. A single NVIDIA H100 GPU draws 700 watts under load. Scale that across an on-premises cluster of 1,000 GPUs and you’re managing over a megawatt of continuous power demand, before factoring in cooling overhead.⁵ On cloud, hyperscale providers charge $3 to $5 per GPU-hour on-demand for mid-tier hardware, and hidden costs including data egress fees, storage for training datasets, and networking add another 20 to 40 percent on top of the listed compute rate.⁶ On-premises deployments carry lower variable costs over time but require capital investment in hardware (at $2 to $5 million for a modest cluster), facility upgrades, power contracts, and cooling infrastructure. Neither path is cheap, and neither is accurately modeled using estimates built from a vendor demo environment.

Licensing and API fees. SaaS AI tools and foundation model API access are increasingly priced on consumption: per token, per call, per query. That structure is economical during pilots. It can become alarming at production scale. A single agentic workflow running across hundreds of concurrent enterprise processes can generate token costs that scale faster than the productivity gains it produces. Some organizations that moved generative AI from pilot to production in 2024 and 2025 found their monthly cloud compute bills in the tens of millions of dollars — none of which was visible in the pilot-phase economics.⁷

Data readiness. This is the cost that consistently surprises organizations, and it’s the one most responsible for AI projects failing. Gartner found that 85% of AI projects fail due to poor data quality or a lack of relevant data.⁸ Informatica’s 2025 CDO Insights survey identified data quality and readiness as the top obstacle cited by 43% of respondents.⁹ Gartner predicts that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data.¹⁰

The problem is structural. Traditional data management architectures were built for reporting and compliance, not for training or running AI models. Making data AI-ready involves cleaning, labeling, governance, metadata management, and the construction of quality pipelines — work that is time-consuming, expensive, and often assigned to teams that are already overextended. Best-practice estimates suggest allocating 40 to 50% of total AI project resources to data readiness alone.¹¹ Most budgets don’t come close to that.

63%

of organizations either do not have, or are unsure whether they have, the right data management practices for AI. Source: Gartner, 2024.

Integration and customization. Off-the-shelf AI rarely maps cleanly to enterprise workflows. Custom integrations with ERP systems, CRMs, proprietary databases, and legacy platforms are the rule, not the exception. Large enterprises typically allocate 2 to 3 percent of annual revenue to integration and ERP systems.¹² AI deployments compound that demand.

Talent. AI skills carry a measurable wage premium. PwC’s analysis of nearly a billion job postings found that roles requiring AI skills command a 56% wage premium over comparable positions without that requirement.¹³ U.S. AI job postings in Q1 2025 carried a median salary of $156,998.¹⁴ Organizations that cannot attract or retain this talent pay twice: once in lost capability, and again in the cost of consulting and vendor dependency that fills the gap.

A realistic cost picture for a mid-market enterprise deploying AI in two or three business functions over 18 months would include infrastructure buildout or cloud commitment, data readiness remediation, licensing and API fees, integration engineering, change management, and talent. None of those line items is small. Together, they’re often two to three times what was presented to the board.

Cost Category	Benchmark Range	Notes
Cloud GPU compute	$3 – $5 / hr per GPU	Plus 20–40% in hidden egress, storage, and networking costs
On-prem GPU cluster	$2M – $5M+	Capital cost; excludes power, cooling, and ongoing staff
Data readiness	40 – 50% of project budget	Cleaning, labeling, governance, quality pipelines
AI talent (median U.S. salary)	$156,998 / yr	56% wage premium over comparable non-AI roles
Avg. sunk cost per abandoned project	$4.2M	Median time to abandonment: 11 months
Shadow AI breach premium	+$670K per incident	Above the global average breach cost of $4.9M

The Costs That Show Up After Go-Live

Getting an AI system into production is an achievement. It’s also when a different category of costs starts accruing, and where most organizations find their financial models were optimistic in ways they didn’t anticipate.

Model drift and retraining. AI models are not static assets. Their performance degrades as the world changes, as user behavior shifts, and as the data they were trained on grows stale. Retraining requires compute resources, engineering time, and fresh labeled data — all of which carry ongoing costs. For organizations that deployed generative AI features into customer-facing products in 2024, the retraining cycle is no longer theoretical. It’s a line item on the quarterly technology budget.

Inference costs at scale. Pilot economics and production economics are almost never the same. The compute cost structure of a pilot, which runs on controlled datasets with modest query volumes, bears little resemblance to what happens when a deployed model serves thousands of concurrent users, processes live transactions, or runs continuously inside agentic workflows. Some enterprises that moved systems into production found their monthly AI infrastructure costs in the tens of millions — a figure that was invisible in the original business case.⁷ Organizations that underbudget ongoing costs by 40 to 60% relative to initial estimates are common, not exceptional. That pattern is consistent enough that it should be treated as a planning assumption, not a risk.

Compliance and audit overhead. Regulators are catching up. The EU AI Act classifies AI systems by risk tier and carries fines of up to 35 million euros or 7% of global annual turnover for violations involving high-risk AI systems.¹⁵ In financial services, SR 11-7 guidance governs model risk management. In healthcare, HIPAA exposure extends to any system that processes protected health information, including AI tools. Compliance overhead includes documentation, audit trails, governance frameworks, risk assessments, and legal review. This cost is not optional, and it scales with the number of AI systems in production.

Shadow AI. This is the one most executive teams underestimate, and it may be the most financially significant. Shadow AI is ungoverned AI use by employees: personal accounts on free-tier AI tools, browser extensions that intercept page content, SaaS features that activate AI capabilities without IT awareness. According to Menlo Security, 68% of employees used personal accounts to access free AI tools in 2025, with 57% of them using sensitive data.¹⁶ IBM’s 2025 Cost of a Data Breach report introduced shadow AI as a formal breach factor for the first time. One in five organizations studied had experienced a data breach directly tied to a shadow AI incident. Those organizations faced an average of $670,000 in additional breach costs.¹⁷ Shadow AI breaches took an average of 247 days to identify, six days longer than standard incidents.¹⁷ Despite this, only 37% of organizations have AI governance policies in place.¹⁸

1 in 5

organizations studied by IBM in 2025 experienced a data breach directly tied to a shadow AI incident, adding an average of $670,000 in breach costs per incident.

Vendor lock-in and switching costs. The AI vendor landscape is moving fast. Organizations that tightly couple their workflows, data pipelines, and integrations to a single provider’s proprietary stack can find that switching costs are prohibitive when pricing changes, support degrades, or a better technical option emerges. The hidden cost of lock-in is not always visible until you try to leave. Evaluating portability before signing multi-year commitments is not just prudent procurement. It’s a financial risk management decision.

Why Most ROI Frameworks Give You the Wrong Answer

Most organizations that claim AI ROI are measuring the wrong things. The failure modes are consistent, and they show up across industries.

Measuring outputs instead of outcomes. A common proxy for AI ROI is something like “documents processed per hour” or “emails drafted per day.” These are outputs. The relevant question is whether those outputs translate into outcomes: lower operating costs, faster revenue cycles, reduced customer churn, higher margin per transaction. The jump from output to outcome is where most AI business cases break down, and most dashboards don’t measure it.

Attributing gains that would have happened anyway. If your customer service team’s satisfaction scores improve during the same period you deployed an AI chatbot, it’s tempting to credit the AI. But if you also hired more staff, restructured escalation protocols, or improved your product in ways that reduced inbound volume, the attribution is ambiguous. Sophisticated ROI frameworks use control groups, randomized assignment to AI-assisted and non-AI-assisted workflows, and pre/post measurement against stable baselines. Most enterprise AI ROI calculations do none of this.

Ignoring leadership and engineering opportunity cost. AI projects consume disproportionate amounts of organizational bandwidth. Engineering teams that are building AI systems are not building other things. Product leadership focused on AI integration is not focused on other product investments. When your most valuable engineers are spending 60% of their time on an AI initiative that delivers marginal returns, the opportunity cost of what they’re not building is real — and it almost never appears in the project’s financial model.

A more honest ROI framework separates two distinct categories of value. The first is efficiency gains: documented cost reductions from automating tasks, reducing error rates, or compressing cycle times. These should be measured against a specific baseline, net of all operational costs, and verified by the function that owns the process rather than the team that built the AI system. The second is revenue impact: documented changes in conversion rates, customer retention, product throughput, or revenue per employee. These require longer measurement windows, cleaner attribution methodology, and more organizational discipline to validate.

For both categories, time-to-value estimates should be treated skeptically. McKinsey’s research suggests ROI typically materializes within 12 to 24 months for successful implementations.¹⁹ The operative word is “successful.” For the majority of implementations, the timeline either extends significantly or never resolves into measurable value at all.

Where the Real Returns Actually Come From

AI does produce real, measurable ROI. The problem is not the technology. The problem is that organizations apply it to the wrong problems, or apply it to the right problems without the organizational conditions required to capture value.

Here’s what the evidence actually shows about where returns are generated.

Narrow, high-frequency, well-defined tasks. AI performs best when the task is specific, the success criteria are unambiguous, and the volume is high enough that small improvements compound. Loan document processing, insurance claims triage, code review, customer inquiry classification, contract abstraction: these are use cases where AI delivers consistent, measurable throughput gains. The operative feature is definition. Vague tasks produce vague results, and vague results don’t generate measurable ROI.

A concrete example: McKinsey case studies show 65% reductions in agent knowledge lookup time after deploying generative AI copilots in service teams.²⁰ A European bank reported its AI chatbot became roughly 20% more effective within seven weeks of deployment versus its prior rules-based system.²⁰ These are quantifiable, outcome-level gains. They came from narrow, well-scoped deployments — not from broad “AI transformation” mandates.

Processes where data already exists and is already governed. Organizations with mature data infrastructure consistently outperform those without it on AI outcomes. IDC research of more than 4,000 business leaders found that companies with strong data integration achieve an average $3.7x ROI from AI, with top performers reaching $10.3x.²¹ The correlation is not coincidental. AI models are only as good as the data they operate on. Organizations that skip data readiness pay more, wait longer, and get less.

Augmentation, not wholesale replacement. The deployments that produce the strongest sustained returns tend to be those that augment human judgment rather than attempt to replace it entirely. Financial fraud detection provides a clear illustration: models that incorporate small batches of analyst corrections fed back into graph-based models consistently lift recall by double digits while holding false positives flat.²² The human-AI collaboration model produces better outcomes than the AI-alone model and generates more organizational trust, which drives higher adoption rates, which compounds the return.

Workflow redesign before technology selection. McKinsey’s 2025 survey confirms that organizations reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting their modeling techniques.³ This is a critical finding. The organizations that start with “what AI tool should we buy” and then figure out how to use it are the ones that end up in the 60% generating no material value. The organizations that start with “what problem are we solving and how should the work actually flow” are the ones that generate real returns.

What Separates the 5% From Everyone Else

McKinsey defines AI high performers as organizations where AI contributes 5% or more to EBIT and leadership characterizes the impact as “significant.” That group represents roughly 6% of survey respondents.³ BCG’s 2025 analysis puts the comparable figure at 5%.⁴ Either way, the organizations generating real AI returns are a small minority of those investing real AI dollars.

What do they do differently? The research is consistent across sources.

They commit more resources, not less. More than one-third of AI high performers commit over 20% of their digital budgets to AI technologies.³ About three-quarters of high performers have scaled AI across their organizations, compared with one-third of others. This seems counterintuitive given the failure rate data, but the pattern holds: organizations that treat AI as a peripheral experiment get peripheral results. The ones that invest seriously and govern seriously are the ones that get serious returns.

They define success before they start. Projects with clear, pre-approved success metrics achieve a 54% success rate, compared with 12% for projects that lack them.¹¹ That is not a marginal difference. It is the difference between a coin flip and a near-certain failure. The organizations that cannot articulate what success looks like before the project kicks off are almost certainly going to be in the 80% that can’t show the return afterward.

They have sustained executive sponsorship. Sustained executive sponsorship correlates with a 68% project success rate. Projects that lose executive sponsorship at some point achieve an 11% success rate.¹¹ AI implementation requires organizational change, budget reallocation, workflow disruption, and sustained pressure on teams that are already operating at capacity. Without a senior leader who owns the outcome and runs interference on competing priorities, these initiatives stall. The data on this is unambiguous.

They invest in data infrastructure before AI infrastructure. The organizations that consistently generate AI returns are those that treated data readiness as a pre-condition rather than an afterthought. The winning programs earmark 50 to 70% of timeline and budget for data readiness work: extraction, normalization, governance metadata, quality dashboards, and retention controls.²² The ones that skip this step pay 2.8 times more in remediation costs when the problems surface in production.¹¹

They say no to low-value AI initiatives. Gartner warns that over 40% of agentic AI projects may be cancelled by 2027 if they lack clear value or governance.²³ High-performing organizations are more disciplined about project selection. They run fewer AI initiatives but run them better. The pressure to have an AI initiative in every function, to demonstrate AI activity to the board, and to keep up with the hype cycle produces exactly the kind of underfunded, poorly scoped, marginally governed projects that end up in the 80% failure category. Resisting that pressure is a competitive advantage.

Eight Questions to Ask Before You Approve the Budget

Before any C-suite leader or board member signs off on a significant AI investment, these are the questions that should have clear, honest answers. If they don’t, that’s the information.

1. What is the specific business problem this initiative is solving, and how will we know in 12 months whether it’s been solved? Not “improve efficiency” or “accelerate AI adoption.” A specific problem, a specific outcome, a specific measurement methodology defined before the project begins. If this cannot be answered in one clear paragraph, the project is not ready to be funded.
2. What is the data readiness score for this initiative, and has it been independently assessed? Not the data readiness score as characterized by the team proposing the project, or the vendor selling the solution. An independent assessment of whether the relevant data is clean, governed, accessible, and AI-ready in the specific context of the proposed use case.
3. What is the fully loaded cost model over three years, including infrastructure, data remediation, integration, talent, compliance, and ongoing operations? If the model does not include all of these categories, it is incomplete. If ongoing operational costs are not at least as large as the initial investment, treat that projection with skepticism.
4. What are the top three ways this project can fail, and what is the plan to mitigate each? Not risks buried in an appendix. A direct, honest discussion of the most likely failure modes: data that isn’t ready, integration that takes longer than modeled, user adoption that doesn’t materialize, regulatory exposure that wasn’t fully scoped. If the project team cannot name these clearly, they haven’t thought hard enough about them.
5. Who owns this initiative, and what authority do they have to make decisions, reallocate resources, and say no to scope changes? Diffuse ownership and committee-based governance are among the strongest predictors of AI project failure. A named senior leader with real authority and genuine accountability is a prerequisite, not a formality.
6. What is our shadow AI posture, and does this investment include governance infrastructure for ungoverned employee AI use? An organization that is investing in a major AI initiative while 68% of its employees are using personal AI accounts to process sensitive data has a serious exposure that the new investment alone will not close. The governance question and the investment question need to be answered together.
7. What is the pilot-to-production cost ratio, and has the production-scale cost been independently modeled? The inference cost structure at full production scale routinely bears no resemblance to the pilot-phase economics. Require that the production-scale cost be modeled by someone who is not selling the solution, using realistic usage volume projections rather than pilot-phase estimates.
8. If this project generates zero measurable ROI in 18 months, what is the exit protocol? Most AI projects that fail persist too long before being acknowledged as failures. The average sunk cost per abandoned AI project is $4.2 million, and the median time to abandonment is 11 months.¹¹ Organizations that define success and failure criteria upfront, including clear decision gates at 6 and 12 months, make better capital allocation decisions than those that let failing projects run on hope.

None of these questions should be uncomfortable to answer if the initiative is genuinely ready to be funded. The discomfort, if it comes, is information worth having before the check is signed.

---

References

1. Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2025. Stanford University, 2025. aiindex.stanford.edu/report/
2. Gartner. “Gartner Forecasts Worldwide AI Spending to Total $1.5 Trillion in 2025.” Gartner Press Release, 2025. gartner.com/en/newsroom
3. McKinsey & Company. The State of AI in 2025: Agents, Innovation, and Transformation. McKinsey Global Survey, November 2025. mckinsey.com
4. Boston Consulting Group. “The Widening AI Value Gap.” BCG Survey of 1,250 CxOs across 20+ sectors, September 2025.
5. Spheron Network. “AI Inference Power Consumption and GPU Electricity Costs: 2026 Guide.” Spheron Blog, 2026. spheron.network
6. GMI Cloud. “How Much Do GPU Cloud Platforms Cost for AI Startups in 2026?” GMI Cloud Blog, 2026. gmicloud.ai
7. Compute Forecast. “AI Inference Cost in Enterprise Infrastructure: Cloud Shift.” ComputeForecast.com, 2026. computeforecast.com
8. Gartner. “Gartner Survey Finds Generative AI Is Now the Most Frequently Deployed AI Solution in Organizations.” Gartner Press Release, May 2024. gartner.com
9. Informatica. CDO Insights 2025. Informatica Survey Report, 2025. informatica.com
10. Gartner. “Lack of AI-Ready Data Puts AI Projects at Risk.” Gartner Press Release, February 2025. gartner.com
11. Pertama Partners. “AI Project Failure Rate 2026: 80% Fail.” Pertama Partners Analysis, February 2026. pertamapartners.com
12. Integrate.io. “Data Integration Adoption Rates in Enterprises.” Integrate.io Blog, 2026. integrate.io
13. PwC. 2025 Global AI Jobs Barometer. PwC Research, 2025. pwc.co.uk
14. Veritone. AI Job Market Report Q1 2025. Veritone Blog, 2025. veritone.com
15. IP Consulting Inc. “Shadow AI Breaches Are Here: The $670,000 Problem Most Companies Can’t See.” IP Consulting Blog, 2026. ipconsultinginc.com
16. Menlo Security. Enterprise AI Usage Report, 2025. Referenced in Proofpoint Shadow AI Analysis. proofpoint.com
17. IBM. Cost of a Data Breach Report 2025. IBM Security, 2025.
18. IBM. “Only 24% of Generative AI Projects Include Built-in Security Measures.” IBM X-Force Research, 2025.
19. Second Talent. “AI Adoption in Enterprise Statistics & Trends 2025.” Second Talent Research, 2025. secondtalent.com
20. McKinsey & Company. “McKinsey Case Studies: Generative AI in Customer Service.” McKinsey QuantumBlack, 2025. mckinsey.com
21. IDC / Integrate.io. “Top AI Leaders Achieve $10.3x ROI Through Advanced Data Integration.” IDC Research of 4,000+ Business Leaders, 2024. integrate.io
22. WorkOS. “Why Most Enterprise AI Projects Fail and the Patterns That Actually Work.” WorkOS Blog, July 2025. workos.com
23. Gartner. “Gartner Top Strategic Predictions: 40% of Agentic AI Projects May Be Cancelled by 2027.” Gartner Research, 2024–2025. gartner.com

---

© 2025 Eric Ford. All rights reserved. Reproduction in whole or in part without written permission is prohibited. This article is intended for informational purposes only and does not constitute financial, legal, or investment advice.