Reuters this week reported that banks worldwide are facing unexpectedly high costs from their AI agents. The McKinsey Global Banking Annual Review pointed out that token bills are running wild and could hit 20% of their total labor expenses.
Banks have already worked to control these costs by routing tasks based on complexity to different models.
But the problem’s back again due to AI agents, who are racking up costs, in part, while communicating with each other.
Now these organizations are reserving their agentic AI use for areas of most productivity. They’re also shifting from just measuring productivity to looking at the actual revenue growth from AI to better assess value.
Taking charge of these tasks are new teams that some banks are calling AI FinOps.
I recently wrote an article about tokenomics and AI ROI. Today, I return to AI ROI with a look at how FinOps is taking on AI.
I consider FinOps best practices, AI use of the FOCUS specification, and the ongoing struggle enterprises are having proving AI business outcomes.
The state of AI technology ROI optimization
I first referenced the nonprofit, industry-led Tokenomics Foundation in the article above. It exists to establish standards and best practices for enterprise AI cost management.
In September, it released a report focused on the emerging economic picture of enterprise AI, spanning 11 industries. I recommend checking out the full report.
Some key revelations for my purposes here:
- 87% of enterprise cloud tokens among those surveyed came from just the three hyperscalers (AWS Bedrock, Google Vertex, and Azure Foundry).
- 86% of the enterprises are already using or evaluating AI model routing.
- All use frontier models, with around 1 in 5 also using open-weights models from at least one of the following providers: DeepSeek, Qwen, Kimi, or GLM.
- Almost all plan to shift to more open-weights model usage over the next year.
- Only 20% were not confident in their visibility on AI spending.
- But 73% could not confidently prove AI business results to their CFO. (Only 9% of all surveyed were very confident in their ability to do so! I look at their common traits below.)
Hence, proving AI value or ROI is unsurprisingly the most commonly cited challenge enterprises are facing in this arena.
(Only 7% cited cost or pricing complexity.)
Token costs are at an all-time low, which is probably why only 4% indicated they would ask model and token providers for cheaper prices.
Instead, the thing most wanted from AI model and token providers is more transparency and granular data, with the use of standards like FOCUS (see below) not far behind.
What is FinOps and how is it evolving beyond cloud cost management?
FinOps is the financial answer to DevOps, and the FinOps Foundation defines it like this:
“FinOps is an operational framework and cultural practice which maximizes the business value of technology, enables timely data-driven decision making, and creates financial accountability through collaboration between engineering, finance, and business teams.”
The goal of FinOps is to optimize the value of technology, whether this means cloud cost optimization, SaaS spend management, or, increasingly, AI.
Borne from the era when enterprises faced sobering cloud cost and waste problems, FinOps focuses on bringing teams together from across the business to understand and control costs.
The FinOps Foundation’s own 2026 report (it surveyed 1,192 organizations with more than $83 billion in annual cloud spending) found that 98% of these teams are also managing AI spending.
This is a dramatic rise from 2024 (when it was less than a third).
Today, FinOps combines AI tools spending, SaaS (90%), software licensing (64%), and often even data center spending (48%).
In short, it’s a discipline that’s shifted from the cloud to managing the value of technology overall, which can be very hard to separate cleanly into its component parts.
In addition to working for transparency and understanding, it is also evolving to help businesses plan their future investments.
How FOCUS cloud cost management fits in
FOCUS (FinOps Open Cost and Usage Specification) is a FinOps framework or standard with the goal of reducing cost complexity.
FOCUS normalizes tech billing and usage data across providers (cloud, SaaS, data centers, AI tools, internal). This ensures that the billing data that filters into a business throughout the year aligns across providers and services on things like terminology, taxonomy, and included metrics.
It’s a separate project from the Linux Foundation-managed FinOps Foundation (which launched FOCUS in 2023, but with its own steering committee and support from Google and Microsoft).
Today all the major cloud providers participate, and among surveyed businesses that spend at least $100 million a year on cloud, 68% are using or experimenting with FOCUS (with another 18% planning to).
The standard’s next release (1.5) is scheduled for December, and it plans to add a number of new features that are highly relevant to AI.
This includes model and developer information, token-level visibility (between input and output, cached and not), requester ID (who is responsible for charges where possible), and potentially even a price comparison dataset (between published and negotiated prices).
FOCUS is a tool for FinOps that makes it easier for businesses to slice and dice charges and analyze usage across the board.
But it’s still just a standard way of being billed. Using this data effectively remains the responsibility of the businesses themselves.
How can enterprises improve technology ROI through FinOps and unified cost management?
With greater visibility on charges and a clearer understanding of who is in charge, the challenge shifts from identifying spending to figuring out what the spending is worth.
In terms of pure AI cost management and reduction, there are many steps businesses can take to get a handle on charges.
As Gouri Sankar Dash wrote for his Forbes piece, businesses should be tackling five critical dimensions:
- Consumption across workloads: Monitor and control token use, model costs, agent interactions, and infrastructure consumption.
- Workload optimization: Align AI model capability to the task at hand, with classification going to smaller models and complex analysis to more advanced reasoning systems.
- Knowledge and efficiency: Optimize and cache so that businesses don’t pay to repeatedly process unnecessary things (aka context management).
- Governance and controlling autonomy: Reserve access for real need, with approval and auditability. (I wrote in my prior article about an out-of-office responder that cost one company $10,000 a day.) Only then expand autonomy where the evidence supports its value.
- Measure outcomes: Track costs by results like resolved (and repeated) incidents, resolution times, and operational reliability.
Unit economics, like measuring the cost per transaction in cloud consumption, charge for resolved issues in customer service (including labor and quality), or the cost per accepted software change in AI-powered development (including testing and verification) is critical for measuring AI’s real impact in a way that can be compared with costs.
But this also requires good operational data and system integration, as well as having standardized (and accurate) definitions of what success means in each case.
ROI isn’t just about cost reduction.
It also is about accelerating and scaling work for more revenue, improving quality, and creating all-new products or services that weren’t possible before.
The bottom line: Better technology spend management for 2027
The enterprises that were very confident in their ability to connect AI spend to measurable business outcomes in the Tokenomics Foundation report had the following things in common regarding their IT cost optimization:
- Their AI spending is metered and can be attributed across teams and workloads.
- They track concrete business outputs (like tickets, PRs, and revenue).
- They have defined AI FinOps/Tokenomics ownership (these are 3.7x more likely to show AI value to their CFO).
- They have sufficient workflow understanding to use model routers (these are 4x more likely to show AI value).
At PTP, we’ve worked extensively in this area. Our AI VOICE Framework was built from hands-on experience helping enterprise clients secure effective, reliable, measurable AI investments.
It’s based on a series of foundational pillars, and drawing from it in the area of proving AI value, I recommend the following:
- Unify the spending visibility. FinOps is all about this practice, and the FOCUS standard makes this process far easier. Consolidate and clearly allocate cloud, SaaS, AI, licensing, and infrastructure spending.
- Clearly define ownership. Critical to our VOICE Framework, the Tokenomics Foundation study also makes clear this has a clear relationship to being able to demonstrate AI value. Among companies that had no clear enterprise AI economics owner, not a single one could show AI value.
- Measure outcomes over activity. This is also a core component of VOICE. Tracking tokens, usage, and hours saved are meaningless in a vacuum. (Our most recent PTP Report demonstrated how many companies are finding coding hours saved don’t always translate back to additional productivity, and the same is true in sales.)
As the FinOps discipline continues to clarify AI economics, it will get easier for businesses that utilize the discipline to make more informed decisions on things like models, vendors, and the necessary infrastructure.
The more standardized one’s economic data is, the more accessible it can be for business decision-making overall.
And this drives smarter investments and better partnerships.
References
Banks seek to rein in soaring AI agent costs, McKinsey says, Reuters
State of Tokenomics; September 23, 2026 Release Candidate 1.0, Tokenomics Foundation
From FinOps To TokenOps: The New Economics Of Enterprise AI Infrastructure, Forbes
State of FinOps Survey: AI Value and Skills Top Priorities as FinOps Matures Across Technology Value (98% Manage AI, 90% SaaS, 64% Licensing, 48% Data Center), The Linux Foundation
Enterprise AI’s Real Problem Isn’t Spending. It’s Proving What That Spending Is Worth, ValueOps by Broadcom
What’s coming in FOCUS 1.5, and what is not, FOCUS
FAQs
What is the difference between traditional cloud cost management and enterprise-wide technology spend management?
FinOps began with cloud cost management. This was focused on controlling expenses, eliminating waste, and optimizing your resources. Enterprise-wide spending management applies these same principles but across the cloud, AI, SaaS, licensing, and more. The boundaries between these areas are fuzzy and overlapping, but this expanded focus also shifts the goal from just reducing costs to better understanding (and maximizing) the total economic impact.
How can businesses identify hidden costs and reduce waste across cloud, AI, and SaaS platforms?
The FOCUS standard can help here, as it normalizes billing data across providers. Paired with FinOps practices that connect spending to specific units and consider business outcomes, this helps identify idle resources, unused licenses, redundancies, and excessive areas of spending. It should also consider human oversight time required for automation.
What are the best FinOps practices for optimizing cloud, AI, and SaaS spending?
Having unified cost visibility, as discussed, allows better comparison. But FinOps is also about having clear accountability and tracking the actual business outcomes related to costs. With standardized reporting and ownership tied to spending, budgets are easier to make and track, and measurements based on outcomes at the unit level help pair spending with returns rather than just focus on cutting costs across the board.
