Where Enterprise AI Budgets Actually Go: An Evidence Review

The Spend Is Real. The Returns Are Not.

Enterprise AI investment has become one of the defining capital allocation stories of the decade. Private AI investment in the United States reached $109.1 billion in 2024, according to Stanford University's 2025 AI Index — nearly twelve times China's $9.3 billion and twenty-four times the United Kingdom's $4.5 billion. Globally, generative AI alone attracted $33.9 billion in private investment in 2024, an 18.7% increase from 2023.

Those numbers suggest a market in full sprint. The operational reality in most organisations is more complicated.

IBM's Institute for Business Value found that 60% of organisations have not yet developed a consistent, enterprise-wide approach to generative AI implementation — despite the investment pressure. And Deloitte's research found that while two-thirds of organisations report productivity and efficiency gains from AI, only 20% are currently achieving revenue growth through AI investments, against a 74% aspiration rate. The gap between what executives hope AI will do and what their deployments actually produce is the most important story in enterprise technology today.

This article examines where enterprise AI budgets actually flow, why so many organisations are not capturing the value they expected, and what the evidence says about how to close that gap.

The Four Buckets AI Spending Falls Into

Enterprise AI expenditure can be grouped into four distinct cost categories. Understanding how spending distributes across them — and what each category tends to return — is the starting point for any honest budget review.

Cost Category	What it covers	Typical return horizon	Common mistake
Compute and infrastructure	GPU/cloud costs, inference, training runs	12–24 months	Over-provisioning for pilot workloads
Software and licensing	LLM API costs, platform licenses, tooling	6–12 months	Paying enterprise tier for low-volume use
Implementation and integration	Custom development, API work, workflow redesign	18–36 months	Underestimating integration complexity
Talent and capability building	AI engineering, prompt engineering, change management	24–36 months	Treating it as a one-time recruitment cost

The relative weight of these categories varies by deployment type. Organisations building proprietary models spend disproportionately on compute and talent. Organisations deploying third-party models spend the majority on software licensing, integration, and change management. The most common pattern in enterprises that report disappointing returns is a heavy concentration in compute and software — the categories with the most visible purchase orders — and systematic underinvestment in integration and capability building, where most of the actual value is created or destroyed.

Where the Money Actually Goes in Practice

Several patterns emerge consistently from research on enterprise AI spending:

Infrastructure claims the largest single share. Cloud GPU and inference costs are the most visible line item in AI budgets. The cost of inference has fallen dramatically — Stanford AI Index 2025 found that costs for GPT-3.5-level performance dropped over 280-fold between November 2022 and October 2024, while hardware costs declined approximately 30% annually. This creates a measurement problem: as unit costs fall, organisations often expand volume to compensate, holding total infrastructure spend constant while increasing the surface area of deployment. Software licensing undercuts expected savings. Enterprise licensing agreements with major model providers — OpenAI, Anthropic, Google, Microsoft Copilot — frequently cost more than expected once per-seat pricing, usage overages, and support tiers are factored in. Organisations that benchmark their AI costs against early-access pricing or research-tier agreements consistently find commercial deployment more expensive than projected. Integration is the budget line that surprises everyone. IBM's research finding that generative AI investment is expected to grow nearly 4x over two to three years comes with a critical qualifier: most of that growth is driven by the realisation that connecting AI to existing enterprise systems is more expensive and slower than initial estimates. Connecting an LLM to a legacy CRM, a 20-year-old ERP, or a patchwork of data warehouses requires engineering work that no vendor roadmap includes in its cost model. Change management is the budget line that almost no one plans for. PwC's research identified a telling pattern: technology delivers roughly 20% of an initiative's value. The other 80% comes from redesigning work. This ratio is not an argument against AI investment — it is an argument for funding the 80%. Organisations that budget generously for AI tools and minimally for training, process redesign, and change management are buying the 20% and leaving the 80% on the table.

The Adoption Numbers Tell a More Careful Story

78% of organisations reported using AI in 2024, up from 55% the year before, according to Stanford AI Index 2025. That number is often cited as evidence that AI adoption has crossed a critical threshold.

It is worth being precise about what it means.

Using AI is not the same as deploying AI at scale, deriving measurable returns from AI, or having an AI strategy that connects investment to business outcomes. Deloitte's research found that 37% of organisations are using AI at surface level with minimal process changes — the equivalent of issuing a Copilot license to every employee and counting the usage rate as adoption. Only 34% of organisations in the same survey were deeply transforming their businesses through AI.

The breakdown that matters:

Deep transformation (34%): AI is embedded in core processes, with workflow redesign, governance frameworks, and measurable outcome tracking. These organisations typically report multi-year implementation timelines and significant investment in integration and change management.
Selective deployment (30%): AI is redesigning key processes in specific functions. Returns are measurable but contained. These organisations are often mid-journey, with clear use cases but incomplete enterprise integration.
Surface adoption (37%): AI tools are in use but processes have not changed. Productivity gains are minimal or unmeasured. This is where most ROI disappointment lives.

The implication for budget planning is direct: moving from surface adoption to selective deployment requires a different kind of spending than moving from zero to surface adoption. The first step costs mostly in software. The second step costs mostly in people, process, and integration.

Why Revenue Growth Is the Hardest Return to Capture

Deloitte's finding that only 20% of organisations are currently achieving revenue growth through AI — against 74% that aspire to it — deserves detailed examination. It is the most consistent pattern in enterprise AI research, and it reflects something structural about how AI creates value.

AI creates value most reliably in cost reduction and productivity, and most erratically in revenue generation.

The reasons are not primarily technical. They are operational. Revenue growth from AI requires deploying AI in customer-facing contexts — sales, customer service, marketing personalisation — where performance is directly visible to customers, failure has reputational cost, and integration with CRM and customer data systems is non-negotiable. These are exactly the deployment categories where integration complexity is highest and change management requirements are most demanding.

Cost reduction and productivity gains, by contrast, accrue from internal deployments — document processing, code generation, meeting summarisation, report drafting — where failure is contained, iteration is fast, and the integration requirements are lower. These deployments work reliably. They also generate the kind of efficiency statistics that appear in AI investment cases but rarely show up as material items in an earnings call.

The path from cost reduction to revenue growth is not automatic. It requires the same investment in integration and change management that most organisations have already avoided in their internal deployments — at higher stakes.

The Pressure to Spend Is Real and Mostly Unhelpful

IBM found that 64% of CEOs report significant pressure from investors, creditors, and lenders to accelerate generative AI adoption. That number is one of the most telling data points in the enterprise AI landscape.

Pressure to accelerate adoption drives budget allocation decisions that prioritise visible spending over effective spending. The purchase of a major AI platform, a high-profile partnership announcement, or a rapid deployment of AI tools to thousands of employees all demonstrate investment and motion. None of them guarantees value.

Organisations responding to investor pressure by accelerating spending without accelerating the unglamorous infrastructure — data quality, integration work, process redesign, governance — are creating the conditions for the AI implementation failures that the research consistently documents. They are also funding the gap between the 74% that aspire to revenue growth and the 20% that achieve it.

The pattern IBM identified — that 60% of organisations lack a consistent enterprise-wide approach to generative AI — is not a technology failure. It is a governance failure that investment pressure makes worse, not better.

What the Evidence Says About Spending That Works

Across the research base, several characteristics consistently distinguish AI deployments that generate measurable returns from those that do not:

They define the outcome before they define the technology. Organisations that start with a specific business problem — reduce invoice processing time by X%, increase first-call resolution by Y% — consistently outperform organisations that start with a technology and look for applications. This is not a novel insight. It is the most frequently cited lesson in AI implementation research, and the least frequently applied in practice. They budget for integration as a first-class cost. The organisations in Deloitte's "deep transformation" tier universally treat integration work as a primary cost, not a contingency. They plan for it to take longer than estimated and cost more than estimated, and they treat that planning discipline as a competitive advantage. They measure containment, not deflection. The language applies beyond customer service. In every AI deployment category, the organisations generating returns measure whether AI actually resolved the problem (containment) rather than whether AI handled the interaction (deflection). The distinction is the difference between a deployment that reduces cost and one that merely shifts where cost accumulates. They treat change management as a capability, not a project. The PwC finding — that 80% of initiative value comes from redesigning work — is validated consistently across research traditions. The organisations that capture this value have in-house change management capability, not just change management consultants engaged for launch.

FAQ

What percentage of enterprise AI investment generates measurable ROI? Deloitte's research found that two-thirds of organisations report productivity and efficiency gains. Cost reduction is achieved by 40%. Revenue growth — the highest-value outcome — is currently achieved by only 20% of organisations, despite 74% aspiring to it. The gap is consistent across research sources and reflects underinvestment in integration and process redesign rather than technology failure. What is the biggest hidden cost in enterprise AI deployments? Integration consistently surprises organisations on cost and timeline. Connecting AI systems to legacy infrastructure — ERPs, CRMs, data warehouses — requires engineering work that is rarely priced into vendor proposals or initial business cases. Change management is the second most consistently underbudgeted category; PwC research found that technology delivers approximately 20% of initiative value, with the remaining 80% coming from redesigning work around the technology. How fast is enterprise AI spending growing? IBM's Institute for Business Value found that investment in generative AI is expected to grow nearly 4 times over the next two to three years from current levels. Stanford AI Index 2025 found that U.S. private AI investment reached $109.1 billion in 2024, and global generative AI private investment grew 18.7% year-over-year to $33.9 billion. Why are so few organisations achieving revenue growth from AI? Revenue-generating AI deployments require customer-facing integration, which carries higher integration complexity, greater change management demands, and direct reputational risk on failure. Most organisations have followed the path of least resistance to AI adoption — internal productivity tools — which generate efficiency gains but not revenue. Moving to revenue-generating deployments requires the same investments in integration and process redesign that most organisations have deferred. What separates organisations that capture AI value from those that do not? The consistent differentiator across research sources is not technology selection — it is operational discipline. Organisations that define business outcomes before technology, budget for integration as a first-class cost, and treat change management as a capability rather than a project consistently outperform those that invest in AI tools without the surrounding infrastructure.

Further reading: