AI Cost Management: FinOps for AI, TBM, and Tying Spend to Business Results
Mark aistrategy.guide as a preferred source Our AI strategy guides show up more often, with a Preferred badgeHow to control AI spend and connect it to business outcomes the CFO will accept. FinOps for AI handles the token and GPU bill; Technology Business Management (TBM) handles the allocation to departments and results. A CIO's view on budget splits, showback, and why transparency is what brings departments together.
The CFO conversation about AI has changed in the last twelve months. A year ago the question was “how much are we spending on AI?” and nobody could answer it, because the spend was spread across a cloud bill, a dozen API accounts on corporate cards, AI add-ons buried in SaaS renewals, and engineering salaries. Today the question is sharper: “which of these AI line items is producing a business result, and who owns it?” The answer decides whether next year’s AI budget grows or gets frozen.
I have spent most of my career on the technology side of that table, reporting to CFOs as a CIO and global engineering leader. The pattern repeats. Technology spend that cannot be explained by department and by outcome gets cut first in a downturn, whatever its actual value, and AI spend is now large and volatile enough to trigger that reflex. Technology Business Management (TBM) was built for exactly this problem. FinOps adds the real-time layer that usage-priced AI needs.
This page is the operating model I would put in place: what to measure, how to split the AI budget across departments, how FinOps and TBM divide the work, and how to report it so the CFO sees each department’s AI spend next to the result it bought. For the vendor landscape (IBM Apptio and its alternatives), see the companion TBM and Apptio alternatives roundup on We The Flywheel. For where cost control sits in the wider governance programme, start at the AI governance hub.
Why AI spend breaks the usual IT budget
Classic IT budgeting assumes costs are mostly fixed and planned once a year: licences, headcount, data center capacity, multi-year contracts. AI spend behaves differently in three ways that matter to finance.
First, it is usage-priced and spiky. Model API spend scales with tokens, which scale with traffic, prompt length, agent loops and retries. One agent workflow stuck calling a model in a loop can multiply a monthly bill in days. The budget was set in a planning cycle; the cost is set by user behaviour.
Second, much of it sits inside other budgets. A lot of enterprise AI spend never arrives on an invoice labelled AI. It shows up as a price uplift on a CRM, office suite or support platform renewal, as a GPU line inside the cloud bill, or as an expense report for a team’s API key. Gartner has revised its worldwide IT spending forecast upward several times in 2026 and named AI infrastructure as the driver each time; the numbers are in the Intelligence Supercycle reference. Most of that growth lands in existing categories.
Third, the value lands outside IT. The tokens are paid from a technology budget, while the result (faster claims handling, more qualified leads, fewer support escalations) shows up in an operating department’s KPIs. When nobody connects the two, IT carries a growing cost and the business reports a benefit with no cost attached, and the budget fight starts there.
The FinOps Foundation’s State of FinOps 2026 survey of over 1,000 practitioners shows how fast this moved. 98% of respondents now manage AI spend within their FinOps practice, up from 63% in 2025 and 31% two years earlier, and AI cost management is the top skill those teams are hiring for. The same survey has 78% of FinOps teams reporting to the CTO or CIO rather than to finance. With that reporting line, somebody has to build the bridge to the CFO on purpose.
What TBM is
Technology Business Management is a standard taxonomy for classifying technology spend, plus a management practice for using it in budget decisions. The TBM Council stewards it. Apptio, whose platform became the reference implementation, was a founding force behind the Council, and IBM bought Apptio for $4.6 billion in cash in 2023. The product now sells as IBM Apptio.
The core is the TBM taxonomy, now at version 5.0.1 (July 2025). It maps spend through four layers:
| Layer | Question it answers | Where AI shows up |
|---|---|---|
| Technology Cost Pools | What did we buy? | Cloud Services, Software & SaaS, Staffing, Outside Services |
| Technology Resource Towers | What function does it serve? | AI Compute (under Compute), AI Storage, AI Network, AI Models (under Application) |
| Technology Solutions | What do we deliver with it? | AI solution type: AI/ML model development, AI platform services, generative AI solutions |
| Technology Consumers | Who uses it, and for which business value? | Customer service, marketing, finance, engineering |
The last layer is why TBM matters for AI. It forces the question “who consumed this and what did they get?” for every dollar, which is the question the CFO is asking. Since version 5.0 the taxonomy has explicit AI sub-towers and an AI solution type, so AI spend can be tracked in the same structure as the rest of the technology budget. A separate innovation bucket is the easiest line to cut; a cost the CFO can trace to a department’s KPI is not.
TBM has real critics, and the companion roundup covers them in more detail. Two criticisms are worth taking seriously. A full TBM programme can turn into an exercise in allocation precision that costs more to run than the decisions it informs. And a platform implementation can take a year before it produces a number anyone trusts. Both are avoidable: start with a small number of solutions and consumers, and accept approximate allocation early.
FinOps for AI: the real-time layer
The FinOps Foundation defines FinOps as an operational framework and cultural practice that maximizes the business value of cloud and technology. The 2025 Framework made Scopes a core element (Public Cloud, SaaS and Data Center first, with AI as a practitioner-defined scope) and added “and technology” to that definition. Its open billing-data standard, FOCUS (FinOps Open Cost and Usage Specification), reached version 1.3 in December 2025 with allocation columns for shared resources and a contract commitment dataset. The Foundation says future versions will extend to data center and AI scopes.
For AI, FinOps covers the work that moves too fast for a monthly TBM cycle. Mapped to the Framework’s own capability names:
- Allocation. Every model call carries a team, product and use-case tag, enforced through an internal AI gateway rather than left to each team’s API key.
- Unit Economics. Cost per thousand tokens is the vendor’s metric. Cost per resolved ticket, per processed claim or per generated campaign is the business’s metric, and it is the one FinOps for AI should report.
- Anomaly Management and Budgeting. Budgets and alerts per use case, rate limits on agent loops, and routing to cheaper models where quality allows. Model choice is now the largest single cost lever: the same workload can differ by an order of magnitude in cost depending on which model tier handles it. The per-request data comes from LLM observability tooling, and the capacity side from GPU inference platforms.
- Rate Optimization. Provisioned throughput, reserved GPU capacity and enterprise agreements with model vendors are financial commitments and need the same discipline as cloud reservations. The strategic trade-off is in the vendor optionality piece.
How FinOps and TBM fit together
FinOps runs weekly or daily, owned by engineering and platform teams, and checks whether each workload is efficient. TBM runs monthly and quarterly, owned jointly by the CIO and CFO organisations, and checks whether total spend is allocated correctly and producing the results the business expected. The FinOps Framework names IT Financial Management (ITFM), the discipline TBM formalises, as one of its allied personas, so the two were designed to connect.
| FinOps for AI | TBM | |
|---|---|---|
| Cadence | Daily to weekly | Monthly to quarterly |
| Primary owner | Engineering, platform, FinOps team | CIO and CFO jointly |
| Unit of analysis | Workload, model, request | Solution, department, business capability |
| Core question | Is this efficient? | Is this worth it, and who pays? |
| Data standard | FOCUS billing data | TBM taxonomy |
| Output | Optimisation actions, guardrails | Budget allocation, showback, business-value reporting |
They meet at the solution and consumer mapping. FinOps tags each AI request with use case and department; TBM rolls those tags up into Technology Solutions and Technology Consumers, so finance sees the whole picture, including the staffing and SaaS costs that never appear in a token bill.
Splitting the AI budget across departments
The output the CFO wants is a table: each department’s share of AI spend, what it is driving toward, and whether it is working. Here is the structure I use. The numbers are an illustrative example for a mid-size company spending $10M a year on AI across all cost pools, not data from a specific client.
| Consumer (department) | Share of AI spend | Main AI solutions | Business result it should move | Unit cost reported monthly |
|---|---|---|---|---|
| Customer service | 25% | Agent assist, self-service automation | Cost per contact, first-contact resolution | Cost per resolved contact |
| Engineering and product | 25% | Coding assistants, AI features in product | Delivery throughput, feature revenue | Cost per merged change, AI feature gross margin |
| Sales and marketing | 20% | Content generation, lead scoring, personalisation | Pipeline, conversion rate | Cost per qualified lead |
| Operations and supply chain | 10% | Forecasting, document processing | Forecast accuracy, cycle time | Cost per document or forecast |
| Finance, HR, legal | 5% | Contract review, close automation, recruiting | Close time, time to hire | Cost per contract reviewed |
| Shared AI platform (central) | 15% | Gateway, observability, evaluation, governance | Reliability, compliance, reuse | Platform cost as % of total AI spend |
Three rules keep the table honest.
Every row has a business owner as well as a cost centre. The VP of customer service owns the cost-per-contact number; the CIO does not. If no business leader will put their name against the result, the spend is an experiment and should be budgeted and reported as one.
The shared platform stays central and visible. Gateways, evaluation, security and governance tooling benefit everyone, and allocating them too finely gets departments arguing about their share of the logging bill. Keep the platform as one line with its own efficiency metric.
The business metric is agreed before the spend is approved and reported every month next to the unit cost. A CFO will accept an ROI claim built that way, because the metric was fixed before anyone knew the answer. It connects directly to how I measure AI adoption and to the per-cluster failure data in the cost of failed AI projects: the use cases most likely to fail are the ones that never agreed a result. In a multi-business-unit group the same table repeats per unit, which is one reason a single enterprise AI strategy document is usually the wrong shape.
Showback first, chargeback second
Once the allocation exists, the question is whether to show departments their AI costs (showback) or move those costs into their P&L (chargeback). I do both, in that order.
Showback runs first, for at least one or two quarters. Publish each department’s AI consumption, cost and unit cost without moving money, so every leader can challenge the allocation method while the stakes are low. Disputes in this phase are useful, because each one fixes a tagging rule or a mapping.
Chargeback comes second, and only for consumption a department actually controls. Seat-based AI tools and department-specific use cases move to chargeback cleanly. Shared platform costs and early experiments stay central, because charging them back discourages the reuse and experimentation the platform exists to support.
What transparency does to the room
In the organisations where I have seen TBM take hold, the biggest change was in the budget conversation itself. The savings came later and mattered less.
Without a shared taxonomy, annual planning is a negotiation between departments that each see technology as a cost someone else is causing. Finance sees a growing IT line, IT sees business units that request more and fund less, and the business units see IT as slow and expensive. Everyone brings their own numbers and none of them reconcile.
With a shared taxonomy and a showback report everyone has read for months, the same meeting becomes a portfolio discussion. The CFO can see which departments’ AI spend is moving their KPIs. Department heads can see what their peers spend and get. IT stops defending one large number and presents a set of business investments, each with an owner. That shift, from “IT’s budget” to the company’s technology portfolio by department and outcome, is what puts finance, technology and the business units on the same side of the table.
AI raises the stakes. Its spend is variable, it is growing faster than the rest of the technology budget, and its benefits land in departments that do not pay the bill. Left opaque, it is the first thing cut when margins tighten. Reported by department and outcome, it becomes the easiest technology spend to defend.
A 90-day starting plan
- Inventory AI spend across the cost pools that carry it (Cloud Services, Software & SaaS, API accounts on expenses, Staffing). Expect the first total to come in well above anyone’s estimate.
- Route model calls through a gateway that enforces tags for department, product and use case. This is the single most useful technical control.
- Map spend to the TBM taxonomy at a coarse level: the AI sub-towers, a handful of AI solutions, and departments as consumers. Use FOCUS exports where providers support them.
- Agree one business metric and one unit cost per use case with each business owner.
- Publish a monthly showback report to the CFO and department heads, one page long: the department table above, with trend lines.
- Choose tooling after the first two cycles, once the questions are clear. The Apptio alternatives roundup compares IBM Apptio, Flexera, Finout, Nicus, MagicOrange, SAP LeanIX and Vendr. None of them is open source; the open parts of the stack are FOCUS and the published TBM taxonomy.
My verdict: AI cost control is solved by transparency, and TBM plus FinOps is the most proven way to give a CFO that transparency.
Sources
- TBM Council, TBM Taxonomy, version 5.0.1, July 2025 (layers, cost pools, AI sub-towers and solution type).
- FinOps Foundation, FinOps Framework 2025 (definition, Scopes, domains, capabilities, allied personas).
- Linux Foundation / FinOps Foundation, FOCUS 1.3 release, December 2025.
- State of FinOps 2026 key findings, summarised by CloudKeeper.
- GeekWire, IBM completes $4.6B acquisition of Apptio, 2023.
- Department budget table: illustrative structure, not client data. If a figure here looks wrong, send it and I will publish the correction.
