Finance leaders no longer need convincing that AI matters. According to a study conducted by leading consulting firm, nearly nine in ten CFOs now say AI adoption in the finance function is extremely or very important, and the share describing their own posture toward AI as conservative has collapsed from 70% in 2020 to just 4% today. The debate has moved on from “should we adopt this” to “why isn’t it working the way we expected.”
That second question deserves more scrutiny than it usually gets. Across industries, Gartner’s research into AI infrastructure and operations found that only 28% of AI use cases fully succeed and meet ROI expectations. Another 20% fail outright. Among leaders who reported a failure, 57% said the root cause was simple: they expected too much, too fast.
“AI in finance rarely fails because the technology is broken—it fails because the operating model is.”
Furthermore, an independent analysis by the RAND Corporation of enterprise AI initiatives puts the broader enterprise AI failure rate even higher, finding that roughly 80% of projects never deliver their promised business value. Read through a finance lens, that failure pattern can be understood as spanning three broad outcomes: initiatives abandoned before production, initiatives that reach production but underdeliver, and initiatives that run but never recoup their cost.
Read these findings together, and a clear pattern emerges: AI in finance rarely fails because the technology is incapable. It fails because organizations run into four structural roadblocks that no amount of additional tooling can resolve on its own. This happens because each roadblock is treated as an isolated problem instead of a symptom of the same underlying gap: the absence of an operating model built for how AI actually needs to be deployed, governed, and budgeted for in a regulated function.
ROI fragmentation
“According to McKinsey research, large majority of organizations saw no measurable P&L impact from their initiatives”
The first roadblock shows up early. Generic, horizontal AI tools are built to be broadly competent, not deeply correct, in finance, ‘broadly competent’ is not the same as reliable. A model that drafts serviceable marketing copy will not, by default, reason correctly through a complex revenue recognition edge case or a non-standard credit memo. The judgment those tasks require is deeply domain-specific, not general-purpose.
The data bears this out In awidely cited report issued by MIT Nanda Initiative into enterprise generative AI deployments found that the large majority of organizations saw no measurable P&L impact from their initiatives, while a small minority extracted real, documented value. The differentiator was rarely the underlying model.
McKinsey research similarly finds that workflow redesign is one of the strongest predictors of financial value from generative AI. Among 25 organizational attributes examined, redesigning workflows had the largest effect on an organization’s likelihood of realizing EBIT impact from gen AI, reinforcing the idea that value comes less from the tool itself and more from reengineering the underlying process around it.
Runaway costs
“Gartner’s analysis found that organizations are rapidly moving from experimentation to scaled deployment, but are structurally underestimating the financial impact of rising token consumption”
The second roadblock is economic, and it is compounding faster than most finance functions have built the governance to track. Gartner’s analysis of consumption-based AI pricing found that organizations are “rapidly moving from experimentation to scaled deployment,” but are structurally underestimating the financial impact of rising token consumption. Because developers and business users naturally optimize for speed and convenience over cost, the bills pile up quickly.
Left ungoverned, this pattern is materially expensive at scale. Gartner projects that through 2028, at least half of generative AI projects will overrun their budgeted costs specifically due to poor architectural choices and a lack of operational cost discipline, not due to the underlying model pricing itself.
This is precisely the kind of risk a traditional software budget was never built to catch. Consumption-based AI spend behaves less like a fixed license fee and more like a utility bill that scales invisibly with usage until the invoice arrives. Without cost governance built in from day one—tracking cost per task, per workflow, per outcome—finance functions frequently discover the economics only after the pilot has already scaled past the point where redesign is cheap.
Vendor lock-in
“It is highly unlikely that a single AI vendor or model will meet every one of an organization’s requirements over time”
The third roadblock is strategic rather than financial, though it eventually becomes a cost problem too. Relying on a single frontier model is a structurally fragile position in a market moving this fast: model rankings shift within a single product cycle, providers deprecate versions with limited notice, and outages at a single vendor can take an entire dependent workflow offline.
Gartner’s own guidance on this point is direct: enterprises should adopt a multi-vendor approach. It is highly unlikely that a single AI vendor or model will meet every one of an organization’s requirements over time. Furthermore, hybrid strategies also give finance functions real leverage on token cost as more efficient, open-source, or specialized models enter the market.
The lock-in risk compounds specifically in agentic deployments, where every prompt, guardrail, and tool-calling pattern gets hand-tuned to one specific model’s behavior. This turn a later migration a substantial engineering effort rather than a configuration change. That is precisely why Gartner projects that by 2028, 70% of organizations building multi-LLM applications will route traffic through a model-agnostic abstraction layer, up from under 5% in 2024. The earlier that abstraction layer is architected in, the cheaper it is to keep future options open.
The talent gap
“The domain expertise required technical accounting judgment, audit standards, and risk advisory context cannot be substituted with general technical fluency alone.”
The fourth roadblock is the hardest to solve with a purchase order, because it isn’t really a technology gap, it’s a hiring problem for a role that barely exists in the labor market. Deloitte’s most recent finance workforce research found that 64% of CFOs and finance leaders have already named at least one technical skill a development priority for their teams, and identified some of the hardest capabilities to find and hire for as cash flow management, financial analysis, budgeting and forecasting, and generative AI experience – in combination, not in isolation.
That combination is the actual bottleneck. A strong data scientist without technical accounting depth, or a strong controller without AI fluency, can each get partway to a working solution. Getting the rest of the way requires both perspectives in the same room, wrapped into the same workflow.
This is consistent with what Deloitte’s broader enterprise research finds across functions: the AI skills gap is now cited as the single biggest barrier to AI integration, ahead of data quality, budget, or governance concerns. In a regulated function like finance, that gap is sharper still, because the domain expertise required technical accounting judgment, audit standards, and risk advisory context cannot be substituted with general technical fluency alone.
The common pattern underneath the four roadblocks
Treated individually, these four problems look like four separate procurement decisions: pick a more specialized tool, negotiate a better pricing tier, add a backup vendor, or hire a few more tech specialists.
Treated as a system, however, they point to the same root cause: Most finance functions are trying to run an AI capability without the operating model that AI capability actually requires. They lack a structured way to match use cases to the right level of specialization, govern consumption cost before they compound, keep architecture portable by design, and pair technical execution with real domain judgment from day one.
This is also why maturity, not ambition, is the better lens for thinking about where a finance function actually stands. An organization does not move from “no AI” to “AI-mature” in a single deployment. It moves through a curve-experimentation, first production use case, cost and governance discipline and sustained multi-process advantage. Each of the four roadblocks tends to appear at a different point on that journey:
- Fragmented ROI is usually an experimentation-stage failure.
- Runaway cost is a scaling-stage failure.
- Vendor lock-in is a maturity-stage failure that early-stage teams rarely see coming.
- The talent gap sits underneath all three, acting as the ultimate governor of how fast an organization can move along the curve at all.
The strategic takeaway
None of these four walls get solved by adding another point solution to the stack. They get solved by building deliberately, process by process – the operating discipline to match the right level of AI sophistication to each usecase, govern its cost before it compounds, keep the underlying architecture portable, and pair every deployment with the domain expertise the process actually demands.
That discipline is not a tool. It is a capability, and it has to be built as one.
Pierag’s AI & Digital practice works with finance and technology leaders to build exactly this kind of capability: combining technological agility with deep financial subject-matter expertise to guide custom R&D, govern AI consumption cost before it becomes a budget problem, architect for model portability from day one, and close the domain-technical talent gap that generic vendors and generic hires cannot close on their own so that scaling AI in finance is a structured journey, not four separate fires to fight.
Sources
Gartner/Salesforce: only 28% of AI use cases in infrastructure and operations fully succeed and meet ROI expectations; 20% fail outright; 57% of leaders who reported failure cited overly ambitious, poorly scoped expectations —https://www.salesforce.com/news/stories/cfos-invest-ai-for-growth/
RAND Corporation, “Why AI Projects Fail” (2025): approximately 80% of enterprise AI projects fail to deliver promised business value — https://www.rand.org/pubs/research_reports/RRA2680-1.html
McKinsey & Company (2025 State of AI research, as cited by Talyx): Among 25 organizational attributes examined, redesigning workflows had the largest effect on an organization’s likelihood of realizing EBIT impact from gen AI — the-state-of-ai-how-organizations-are-rewiring-to-capture-value_final.pdf
Gartner: “Gartner Predicts AI Coding Costs Will Surpass Average Developer’s Salary by 2028 as Token Consumption Surges” — https://www.gartner.com/en/newsroom/press-releases/2026-06-24-gartner-predicts-ai-coding-costs-will-surpass-average-developer-salary-by-2028-as-token-consumption-surges
Gartner, “10 Best Practices for Optimizing Generative and Agentic AI Costs” (March 2026): through 2028, at least 50% of GenAI projects will overrun budgeted costs due to poor architectural choices and lack of operational discipline (as cited by TrueFoundry) — https://www.truefoundry.com/blog/the-real-cost-of-generative-ai
Gartner analyst Max Goss, on avoiding single-vendor AI dependency and adopting a multi-vendor approach — https://www.computerworld.com/article/4188012/too-good-to-be-true-avoid-free-ai-token-offers-or-risk-vendor-lock-in.html
Gartner: by 2028, 70% of organizations building multi-LLM applications will use AI gateway/abstraction-layer capabilities, up from under 5% in 2024 (as cited by Swfte AI) — https://www.swfte.com/blog/avoid-ai-vendor-lock-in-enterprise-guide
Deloitte, “Finance Workforce Strategy in the AI Era” — https://www.deloitte.com/us/en/what-we-do/capabilities/finance-transformation/articles/finance-workforce-strategy-ai-era.html
Deloitte, “CFO Guide to Human Capital Trends: AI & Human Collaboration” — 87% of CFOs predict AI adoption in finance will be extremely or very important; 64% named at least one technical skill a development priority — https://www.deloitte.com/us/en/what-we-do/capabilities/finance-transformation/articles/cfo-guide-human-capital-trends-ai-human-collaboration.html
