Enterprise AI costs spiral because they're invisible — per-token pricing, usage-based compute, and shadow tool sprawl break budgeting. Larridin measures AI cost against workflow-level value.
That's the core problem. Not that AI is expensive — it's that nobody knows what they're spending on or whether it's working.
Global AI spending hit $223 billion in 2025 and is projected to reach $301 billion in 2026, per Gartner. AI workloads now consume 24% of public cloud compute, up from 8% in 2023. And yet the Forbes AI Study 2025 found less than 1% of global executives report achieving significant ROI from those investments. The money is flowing. The accountability isn't.
Traditional IT spending follows predictable patterns. You buy licenses. You provision servers. You budget for seats. AI breaks every one of those assumptions.
Per-token pricing creates variable costs with no ceiling. A single OpenAI GPT-5.2 call costs $1.75 per million input tokens and $14 per million output tokens. GPT-5.2 Pro? $21 input, $168 output. DeepSeek V3.2 charges $0.28. The same "AI" line item on a P&L could mean a 60x cost difference depending on which model an engineer chose on a Tuesday afternoon. Finance teams built for annual license renewals have no framework for this.
Usage-based models compound the problem. Cloud compute auto-scales. An ML training job that runs overnight can burn through a month's budget before anyone checks a dashboard. And 89% of IT leaders increased their cloud budgets in 2025 specifically because AI workloads made costs unpredictable — not because they planned to spend more, per a nops.io FinOps survey.
Then there's shadow AI. A Harmonic Security analysis of 22.4 million enterprise prompts found employees using 665 different AI tools, with over 90% of organizations seeing personal AI accounts bypass IT entirely. BlackFog's survey puts it at 49% of employees using unsanctioned AI tools weekly. Every one of those is an untracked cost — $20/month subscriptions multiplied across thousands of employees adds up fast, and it's invisible to finance.
Here's what typically happens inside an enterprise trying to manage AI costs: finance sees a growing Azure or AWS bill. Engineering sees a collection of API keys and model endpoints. Nobody sees how those map to each other.
The Forbes AI Study found only 33% of organizations report regular cross-functional collaboration between finance and AI teams. Finance tracks cost. Engineering tracks accuracy, latency, throughput. The intersection — what did this specific AI capability cost, and what value did it produce — lives in nobody's spreadsheet.
This isn't a tooling problem you can solve with a better dashboard. It's a measurement architecture problem. FinOps platforms can tell you which Azure resource group is expensive. They can't tell you whether the $47,000 your team spent on GPT-4 API calls last month accelerated deal cycles or just generated emails nobody read.
We've watched this play out in every enterprise conversation we've had. One large consulting firm came to us specifically because their Azure cloud spend was climbing and they had zero visibility into whether any of it was producing returns. They could see the bill. They couldn't connect it to outcomes. That gap — between cost and value — is where budgets spiral.
Spending $50,000 a month on AI is fine if it saves $200,000. The problem is not knowing.
Most enterprises try to solve this with surveys. They ask managers: "Is AI making your team more productive?" Managers say yes, because nobody admits their expensive AI initiative isn't working. We've written about why survey-based ROI measurement fails — the short version is that self-reported productivity gains don't survive scrutiny.
Connecting cost to value requires measuring at the task level. Not "are people using AI" but "how long does this specific workflow take with AI versus without, and what does each path cost?"
Consider a concrete example. A development team spends $8,000/month on Copilot licenses and $3,200/month on Claude API calls for code review automation. That's $11,200 in AI costs. If those tools reduce average PR cycle time from 4.2 hours to 1.8 hours across 400 PRs monthly — that's 960 hours saved. At a blended engineering rate of $95/hour, that's $91,200 in recovered capacity. The ROI is obvious. But only if you can measure both sides.
Without task-level measurement, you're stuck defending AI budgets with vibes. And vibes don't survive a CFO review.
Governance doesn't mean locking things down. It means knowing what you're spending and why.
A functional AI cost governance framework has four layers:
1. Inventory and attribution. You need a complete picture of every AI tool, API, and model in use — sanctioned and unsanctioned. This includes SaaS subscriptions, API keys, cloud compute tied to AI workloads, and employee-purchased tools. The workflow mapping gap applies directly here: you can't govern what you can't see.
2. Cost allocation by workflow. Raw cloud bills tell you what resources cost. They don't tell you what work those resources supported. Allocating AI costs to specific workflows, teams, and business outcomes is where most organizations stall. FinOps handles infrastructure attribution. What's missing is the layer above: which business process consumed this compute, and did it produce value?
3. Value benchmarking. Once costs are allocated to workflows, you can benchmark: what does this workflow cost with AI? Without AI? What's the delta in time, quality, and throughput? This is where task-level before/after measurement becomes essential — not as a one-time study, but as continuous instrumentation.
4. Spend controls with flexibility. Hard spending caps kill innovation. Budget alerts without context create alarm fatigue. Effective controls set thresholds by workflow or team, flag anomalies against expected usage patterns, and give engineering autonomy within guardrails. The goal is informed spending, not restricted spending.
FinOps practices have matured rapidly. Deloitte estimated $21 billion in US cloud savings from FinOps adoption in 2025, with organizations achieving up to 40% cost reductions through rightsizing, reserved instances, and waste elimination.
Those gains are real. But they address infrastructure efficiency, not AI effectiveness.
Shutting down idle GPU instances saves money. Switching from on-demand to reserved pricing saves money. But neither tells you whether the AI workloads running on that infrastructure are producing business value. You can have a perfectly optimized cloud bill powering AI tools that nobody uses productively.
The missing layer is connecting infrastructure costs downward to workflows and outcomes. AWS launched Q for Cost Optimization at FinOps X 2025 to surface savings recommendations. Azure introduced AI Foundry reservations for more predictable AI pricing. Google Cloud improved forecasting with AI-specific outlier detection. These are meaningful improvements to cost visibility. They still don't answer: "Was this spend worth it?"
That answer requires measuring what people actually do with AI tools — at the workflow and task level, continuously, not through quarterly surveys or annual audits.
Shadow AI deserves its own accounting because the numbers are staggering.
Reco.ai's 2025 report found smaller firms averaging 269 shadow AI tools per 1,000 employees. IDC's Global Employee Survey reported 39% of enterprise workers using free AI tools and another 17% paying for their own subscriptions — meaning over half of AI tool usage sits outside IT's budget and governance.
The financial exposure isn't just subscription costs. Harmonic Security detected 579,113 sensitive data exposures across the 22.4 million prompts they analyzed. The average cost of a data breach at high-shadow-AI organizations runs around $670,000 per incident.
But here's what makes shadow AI a cost management problem rather than just a security problem: that unsanctioned spending often represents real productivity gains. Gartner predicts 75% of employees will use shadow IT tools by 2027. Employees aren't using rogue AI tools because they're rebels. They're doing it because the sanctioned tools are slower, worse, or nonexistent.
The cost management question isn't "how do we stop shadow AI" — it's "how do we get visibility into what's working so we can fund it properly and cut what isn't."
The shift from anxiety to intelligence is fundamentally about measurement granularity.
Cost anxiety says: "Our AI spend doubled and I don't know why." Cost intelligence says: "Our AI spend doubled because three teams adopted Claude for contract analysis, reducing review cycles from 5 days to 6 hours, saving approximately $340,000 quarterly in outside counsel fees."
Same spend increase. Completely different conversation with the CFO.
Building this intelligence requires three capabilities that most organizations lack today: passive workflow discovery (understanding what people actually do, not what they say they do), task-level before/after measurement (quantifying the impact of AI on specific work), and continuous cost-to-outcome attribution (connecting dollars to results, not just resources). These are the capabilities that turn AI budgets from a liability discussion into an investment discussion.
The enterprises that figure this out first won't just control costs better. They'll invest more aggressively — because they'll have the data to justify it.
Calculate AI API ROI by measuring the cost of API calls attributed to a specific workflow against the measurable output improvement — time saved, throughput increased, or errors reduced. Divide net value gained by total AI cost (including compute, API fees, and integration overhead). Surveys don't work; you need task-level before/after measurement.
Harmonic Security's 2025 analysis found over 90% of organizations have employees using personal AI accounts that bypass IT. IDC reports 56% of workers use free or self-purchased AI tools. The gap between sanctioned and actual AI usage is the single largest blind spot in enterprise cost management.
AI workloads use per-token and usage-based pricing that scales with consumption, not seats. A single model choice can create a 60x cost difference. Auto-scaling compute means overnight training jobs can consume monthly budgets. And shadow AI subscriptions add untracked costs that don't appear in cloud bills at all.
FinOps provides infrastructure-level cost visibility — identifying waste, rightsizing resources, and optimizing pricing tiers. For AI, it's necessary but insufficient. FinOps tells you what resources cost; it doesn't tell you whether the AI workloads on those resources are producing business value. You need a value measurement layer on top.
Start with four layers: complete AI tool inventory (including shadow AI), cost allocation by workflow rather than just resource group, value benchmarking with task-level before/after metrics, and flexible spend controls with anomaly detection. The goal is informed spending, not restricted spending.
Deloitte estimated $21 billion in US cloud savings from FinOps in 2025, with up to 40% reductions through infrastructure optimization alone. But the bigger savings come from redirecting AI spend toward high-value workflows and away from low-impact ones — which requires connecting costs to outcomes, not just reducing compute waste.
Stop guessing where to deploy AI next.
Larridin's AI Opportunity Discovery finds high-impact automation opportunities hiding in your workflows — in minutes, not months.
Discover AI Opportunities →