The FinOps Foundation's State of FinOps 2026 report surveyed 1,192 practitioners representing more than US$83 billion in annual technology spend, and it contains a number that should reframe how you think about your own AI programme. 98% of organisations now manage AI spend as a formal discipline, up from 63% in 2025 and 31% in 2024.
Three years ago almost nobody governed AI cost. Today almost everybody does. The organisations that moved early are not the ones spending the least. They are the ones who can answer a CFO's question inside a single meeting.
That matters more this year than last. Forrester's 2026 Technology & Security Predictions forecast that enterprises will defer 25% of planned AI spend to 2027 as the gap between vendor promises and delivered value widens. Budget is not disappearing. It is moving to the teams that can account for it.
What is FinOps for AI?
FinOps for AI is the operating discipline that makes AI spending visible, attributable and defensible. It extends cloud financial management into tokens, inference, GPU hours, model licences and agent activity, so that every dollar of AI cost can be traced to a team, a workload and a business outcome before the invoice arrives.
The FinOps Foundation defines FinOps as a practice for managing the value of technology, not merely the cost of cloud. In February 2026 the Foundation formally updated its mission from "Advancing the People who manage the Value of Cloud" to "Advancing the People who manage the Value of Technology".
That change of wording is the whole story. Cost reduction was the old brief. Investment governance is the new one.
For a Hong Kong enterprise, the practical translation is simple. FinOps for AI is what stands between an AI programme and the awkward board meeting where nobody can explain what HK$4 million bought.
Why did AI cost management become a board-level issue in 2026?
Because the numbers stopped being rounding errors. Gartner forecasts worldwide AI spending will total US$2.52 trillion in 2026, a 44% year-on-year increase. At that scale, AI stops being a line item inside an IT budget and starts being a capital allocation decision that boards are expected to interrogate.
Three forces converged. The first is absolute spend. Gartner attributes more than half of the 2026 total to AI infrastructure, with AI-optimised servers alone growing 49%.
The second is accountability. Forrester notes that fewer than one-third of decision-makers can tie AI value to their organisation's financial growth, which is why CEOs are increasingly routing AI approvals through the CFO.
The third is internal funding pressure. The State of FinOps 2026 report found that many organisations are being asked to self-fund AI investment through optimisation savings elsewhere. In other words, your AI budget may now depend on how well you can cut something else.
For a Hong Kong department head, that last point changes the conversation entirely. You are no longer asking for new money. You are competing for recycled money, against colleagues making the same argument.
How is AI spend different from ordinary cloud spend?
AI spend is variable, opaque and agent-driven in ways that traditional cloud cost tooling was never designed to handle. A virtual machine costs a predictable amount per hour. An AI agent consumes tokens, invokes tools, queries a vector database and runs multi-step workflows autonomously, so its cost is a function of behaviour rather than provisioning.
The State of FinOps 2026 respondents named three specific challenges when applying FinOps to AI.
--- Visibility. Pricing models vary widely across providers and services, making like-for-like comparison difficult.
--- Allocation. Attributing AI cost to a business unit is materially harder than attributing infrastructure cost.
--- Value. Investments are often exploratory, so returns are hard to define early. One practitioner quoted in the report put it bluntly: nobody can yet answer whether their AI is providing value.
There is also a speed problem. A misconfigured agent can generate in hours the kind of spend that traditional cloud provisioning would take months to accumulate. Monthly invoice review is too slow a control loop for that failure mode.
This is also why the top tooling capability requested by practitioners in 2026 was granular monitoring of AI spend across tokens, LLM requests and GPU utilisation. The market is asking for instruments that do not fully exist yet.
What does a FinOps for AI operating model look like?
The dominant model is a small central enablement team with federated champions embedded in business units. The State of FinOps 2026 data shows 81% of practices operate either centralised enablement (60%) or hub-and-spoke (21%), and even organisations managing over US$100 million in spend average only 8 to 10 practitioners plus 3 to 10 contractors.
The structural detail worth copying is reporting line. 78% of FinOps practices now report into the CTO or CIO organisation, up 18 percentage points versus 2023, while those reporting to the CFO fell to 8%.
The reason is influence. Practitioners with VP, SVP or C-suite engagement showed roughly two to four times more influence over technology selection than those engaging only at director level, including cloud service selection at 53% versus 24% and provider selection at 47% versus 16%.
Read that as a governance instruction. If your AI cost function sits inside finance reporting, it will describe the past. If it sits inside technology leadership with executive sponsorship, it will shape decisions before commitments are signed.
For a 200-person Hong Kong professional services firm, this does not mean hiring a team. It means naming one accountable owner, giving that person executive air cover, and embedding a champion in each business unit that runs AI workloads.
How do you allocate AI costs to a business unit?
You allocate by first standardising the cost data, then tagging at the workload level, then agreeing a shared-cost rule before anyone disputes it. The order matters. Teams that attempt allocation before standardisation end up arguing about the numbers rather than the decisions.
The industry mechanism for standardisation is FOCUS, the FinOps Open Cost and Usage Specification, which normalises billing data across providers. The State of FinOps 2026 report identifies AI workloads as the single most requested area for FOCUS to expand into next, ahead of data centre and broader SaaS support.
A workable allocation sequence for a mid-market enterprise looks like this.
--- Standardise. Pull every AI-related invoice into one normalised format, including model APIs, GPU compute, vector databases and AI features bundled inside SaaS licences.
--- Tag at source. Require every AI workload to carry an owner, a cost centre and a use case before it reaches production. Retrofitting tags is the most expensive form of cleanup.
--- Split shared costs by rule, not by argument. Agree in advance how a shared knowledge base or a common embedding index is apportioned.
--- Report in business units, not tokens. A department head cares about cost per resolved enquiry, not cost per million tokens.
That last translation step is where most programmes fail. Technical unit economics only becomes persuasive once it is expressed in the operational unit the business already manages. If you are still building the underlying cost picture, start from our companion piece on what enterprise AI actually costs and how to attribute it, then apply the governance layer described here.
How do you prove AI value when the return is not obvious yet?
You define the unit of value before deployment, not after. The State of FinOps 2026 report identifies determining AI value and ROI as one of the top three challenges practitioners face, precisely because exploratory projects are measured retrospectively against goals that were never written down.
A defensible value case has four components, and each can be agreed before a single model is called.
--- The baseline. What does this process cost today, measured in hours, headcount, error rate or cycle time?
--- The unit. What single operational unit will you price? Cost per invoice processed, per enquiry resolved, per document reviewed.
--- The threshold. At what unit cost does this become worth continuing, and who decides?
--- The kill criteria. What result, by what date, ends the project without a debate?
Writing kill criteria feels pessimistic. It is the opposite. A programme with a documented exit is far easier to fund, because the CFO's downside is bounded and visible.
This connects directly to a pattern covered elsewhere in our library on why the majority of enterprise AI pilots never reach production. The failure is rarely technical. It is almost always a missing definition of success.
What goes wrong when enterprises skip AI cost governance?
Four failure patterns recur, and none of them announce themselves early. Each one is survivable in isolation. Together they are how an organisation reaches the end of a financial year with significant AI spend and no defensible account of what it produced.
The invisible run rate. AI features arrive bundled inside SaaS renewals, so spend grows without ever appearing as an AI decision. The State of FinOps 2026 data shows 90% of practices now manage SaaS precisely because of this sprawl, up from 65% a year earlier.
The unallocated pilot. A proof of concept runs on a central account with no cost centre attached. Twelve months later it is production-critical and nobody owns the bill.
The runaway agent. An autonomous workflow loops, retries or over-retrieves, and cost accumulates faster than any monthly review can catch.
The value vacuum. The system works, adoption is real, and yet nobody can state the unit economics. When budgets tighten, unquantified programmes are cut first, regardless of merit.
Hong Kong enterprises carry an additional layer here. AI cost governance and data governance overlap, because where a workload runs affects both its price and its compliance position. Enterprises weighing that trade-off should read our explainer on what data residency actually means for enterprise AI alongside their cost model.
What should a Hong Kong enterprise do in the next 90 days?
Build visibility first, governance second, optimisation last. The State of FinOps 2026 data is unambiguous on sequencing: across every technology category, the most prioritised capabilities are allocation, forecasting, budgeting and reporting, because teams consistently find they must understand cost before they can reduce it.
A realistic 90-day sequence for a 200 to 500 person organisation looks like this.
--- Days 1 to 30. Inventory. List every AI cost the organisation carries, including bundled SaaS AI features, departmental subscriptions bought on expense cards, model API keys and GPU capacity. Expect the list to be longer than anyone predicted.
--- Days 31 to 60. Attribute. Assign an owner and a cost centre to each item. Publish the allocation rule for anything shared. Set spend alerts at the workload level rather than the account level.
--- Days 61 to 90. Quantify. Pick your two highest-spend workloads and define their unit economics, baseline and kill criteria. Present these two, not all of them, at the next budget review.
Two workloads presented properly will earn more credibility than twelve presented vaguely. The point of the exercise is not completeness. It is demonstrating that AI spend in your organisation is governed rather than merely incurred.
Organisations building the wider control layer around this should also look at how an AI gateway centralises model access and usage limits, since routing decisions and cost controls are increasingly the same set of levers.
The strategic takeaway
The shift captured in the State of FinOps 2026 data is not about accounting. It is about who gets to decide. FinOps practitioners engaged at executive level now influence technology selection two to four times more often than those who are not, which means cost discipline has quietly become a seat at the architecture table.
Forrester expects a quarter of planned AI spend to slip into 2027. Some of that deferral is prudent. Much of it will simply be budget moving away from leaders who could not answer the question, and towards leaders who could.
You do not need a large team, expensive tooling or a mature practice to start. You need an inventory, an owner, and two workloads you can describe in the language your CFO already uses.
We understand AI. We understand you. With UD by your side, AI never feels cold. That is what 28 years alongside Hong Kong enterprises has taught us: the hardest part of technology is rarely the technology.
Reviewed by the UD enterprise AI team.
Where to Start
A cost discipline is only useful once it is pointed at a real workload. If you are not yet sure which AI initiative in your organisation deserves budget first, start with a readiness view rather than a spreadsheet. We'll walk you through every step, from readiness assessment and use-case shortlisting to deployment, cost allocation and performance tracking, with 28 years of Hong Kong enterprise experience behind every recommendation.
