AI FinOps: taking control of what AI really costs your company.
Most companies deploying AI do not know what it costs them. Not "roughly": not at all. Spend is scattered across token-metered APIs, licences billed per user, prepaid credits that expire, and subscriptions paid on expense reports that IT never sees.
AI FinOps is the discipline that pulls all of that into one place, so someone can actually decide something. This guide covers what that means, why cloud FinOps methods do not transfer directly, which metrics to track, and how to get started in 90 days.
Updated August 2026.
What AI FinOps is, and what it is not.
FinOps is an operational framework born in the cloud, maintained by the FinOps Foundation, itself hosted by the Linux Foundation. Historically, it existed to make variable cloud spend visible, allocate it to the teams generating it, and create shared accountability between finance and engineering.
In March 2026, the FinOps Foundation updated its mission from "the value of Cloud" to "the value of Technology", and formally added AI as a technology category covered by its framework, alongside public cloud, SaaS, data centre, and data platforms. There is now a dedicated "FinOps for AI" page within the official framework. In other words, AI FinOps is not a marketing neologism: it is a documented extension of a framework that has existed since 2018.
- It is not cost-cutting
- The goal is informed decision-making, not economising at all costs. The 2026 framework goes further: it states that an AI scope focused on innovation should tolerate higher waste in order to move fast. Applying cloud efficiency standards to AI experimentation means either under-investing or manufacturing the appearance of governance without the substance.
- It is not employee surveillance
- Measuring volumes and costs does not require reading conversations. That is the subject of the "managing without surveilling" section below.
- It is not a tool
- It is a practice. A tool accelerates it; it does not replace it.
Why cloud FinOps does not transfer directly to AI.
This is the core of the subject, and it is what most articles on AI FinOps skip over. Six structural differences.
1. The unit of cost is not deterministic
An EC2 instance costs the same every hour. An LLM request never costs the same twice: cost depends on input tokens, output tokens, which model was called, cache state, and sometimes on silent retries after errors. Practical consequence: you cannot forecast an AI bill from a request count. You have to measure, not extrapolate.
2. Prompt quality is a cost driver
A badly written prompt, an oversized RAG context, a conversation history that is never truncated: these are decisions that show up directly on the bill. In the cloud, optimisation is infrastructural. In AI, it is twofold.
Applicative first, and it belongs to developers: context size sent, history truncation, caching. Human second, for any direct use of a chat interface: a user who pastes an entire document instead of an excerpt, or reopens a conversation for every question instead of keeping the thread, burns more without a single line of code being at fault. Training teams on lean usage is a cost lever in its own right, not just a productivity question.
3. Pricing levers come last, not first
In the cloud, the first optimisation reflex is pricing: reserved instances, savings plans, multi-year commitments. Nothing changes technically, you commit to a volume, and the bill drops.
On LLM APIs, equivalent mechanisms exist (provisioned throughput on Azure OpenAI or AWS Bedrock, volume discounts on large enterprise contracts), but they only become relevant at high, stable usage, which most organisations do not have yet. Until that threshold, the only available levers are usage, application architecture, and model choice. The order is reversed compared to the cloud: you start with the hard work, not the negotiation.
One thing worth watching
Many organisations built their budgets assuming a continuous drop in token prices. Structural supply constraints, GPU availability and energy costs, make that bet fragile. Budget on today's prices, not the ones you are hoping for. The LLM API price comparison gives current rates, vendor by vendor.
4. Models get deprecated fast
A model released eighteen months ago is often deprecated, replaced, or absurdly expensive compared to its successor. This creates recurring work that does not exist in the cloud: monitoring deprecation announcements, simulating the impact of a migration, and weighing the cost of migrating against the cost of staying put. It is a simulation exercise, not an optimisation one: if I switch this use case from model A to model B, how much do I save, and what do I lose in quality?
5. Licences and consumption coexist
This is the most confusing part. Enterprise AI spend mixes two incompatible pricing models.
- The named licence
- Microsoft 365 Copilot, GitHub Copilot, ChatGPT Enterprise, Salesforce Agentforce. Billed whether it is used or not.
- Usage-based consumption
- OpenAI, Anthropic, Mistral, Gemini APIs. Billed per token consumed.
A dashboard that only handles one of the two is useless. The first is managed through adoption, the second through architecture. Same euros, two different jobs.
6. Shadow IT is massive and immediate
Subscribing to an AI tool costs 20 € a month on a card and takes three minutes. Nobody has ever provisioned a Kubernetes cluster on an expense report. That gap in friction explains why shadow AI is far more widespread than cloud shadow IT ever was.
Where the money actually goes: the five buckets.
Before you measure, you need to know where to look. Five buckets, in decreasing order of visibility.
- 1. Named AI licences
- Copilot, Gemini for Workspace, Agentforce, ChatGPT Enterprise. Visible on vendor invoices. The easiest bucket to measure, and almost always the one holding the most pure waste.
- 2. Production APIs
- What your applications consume when they call LLMs. Visible in vendor consoles, rarely broken down by team or use case.
- 3. Associated infrastructure
- Vector databases, GPUs, orchestration, observability, embedding storage. Usually booked under the cloud budget, never tied back to the AI use case it supports. The most consistently underestimated bucket.
- 4. Prepaid credits
- Pools purchased in advance, from Salesforce, Azure or others, with an expiration date. Paid once, forgotten, expires.
- 5. Shadow AI
- Individual subscriptions paid on expense reports or company cards. Invisible by design, unless you go looking in expense data and SSO logs.
A useful exercise before any tooling: try adding up these five buckets for last month. How long it takes you is already a diagnosis.
The Inform, Optimize, Operate cycle applied to AI.
The FinOps framework runs on a continuous three-phase loop. Here is what each phase means when the subject is AI.
Phase 1. Inform: know
The goal is visibility, not savings. Concretely:
- Inventory
- Every vendor, every licence, every contract. Including the ones you will discover along the way.
- Centralise
- One source, one currency, reconciled against real invoices. A dashboard that does not reconcile with accounting will never be trusted by finance.
- Allocate
- Tie every euro to a team, a use case, a cost centre. It is the most thankless and the most decisive work: without allocation, no tradeoff conversation is possible.
- Detect anomalies
- An alert on unusual variation beats a monthly report. An agent loop with no bound can burn a month's budget in a weekend.
Do not move to phase 2 until phase 1 is honest. The most common mistake is optimising numbers nobody believes.
Phase 2. Optimize: decide
In decreasing order of return for the effort involved:
- Cut what is not used. Licences assigned but never activated, former employees still billed, duplicates, credits about to expire. Immediate gain, zero risk, no technical tradeoff. Always start here.
- Right-size the fleet. Reassign rather than cancel when the usage exists elsewhere.
- Optimise application architecture. Model cascading (a small model handles 80 % of cases, the large one handles the rest), prompt caching, batch APIs, context compression, tighter RAG.
- Weigh model choices. Simulate before migrating, on real euros and a representative quality sample.
- Renegotiate. Last, once you finally know what you actually consume. A negotiation without data is the vendor writing the contract.
Phase 3. Operate: sustain
This is the phase everyone skips, and it is the one that decides whether the whole effort survives past six months.
- Budgets by team and by use case, with alert thresholds at 50 %, 80 %, 100 %.
- A short monthly review with the same people: IT, finance, one point person per consuming team.
- A clear policy on who can subscribe to what, and through which channel.
- An offboarding rule that explicitly includes AI licences. Most offboarding checklists never mention them.
AI FinOps KPIs.
Do not track all twelve. Pick four, and stick with them.
Cost
| KPI | What it reveals |
|---|---|
| Total monthly AI spend | The starting point. All buckets combined. |
| Cost per team | Makes tradeoffs possible. Without it, no conversation happens. |
| Average cost per request | Applicative efficiency signal. Track the trend, not the absolute number. |
| Cost per active user | On a named licence: ties cost to real adoption, not licences purchased. On usage-based consumption: reveals how concentrated usage is across users. |
Waste
| KPI | What it reveals |
|---|---|
| Dormant licence rate | The most profitable of all. Converts to euros immediately. |
| Cost of dormant licences (€/year) | The version that speaks to finance. |
| Unused prepaid credits | Money already spent, about to expire. |
| Budget vs. actual variance | The quality of your forecasting, and therefore your control. |
Efficiency and forecasting
| KPI | What it reveals |
|---|---|
| Activation rate (activated / assigned licences) | The truth about adoption, before any narrative. |
| Daily burn rate | Enables end-of-month projection, and therefore a useful alert. |
| End-of-month projection | Turns observation into anticipation. |
| Output/input token ratio | Leading indicator of API cost drift. |
The 2026 framework makes a point of this: KPIs should be defined by technology category and business context, not carried over from cloud-specific defaults. An R&D scope and a production scope should not share the same targets, and that is fine.
Licence waste: the most immediate opportunity of all.
If you only do one thing, do this.
A dormant AI licence is a licence assigned to someone who never activated it, or has not opened it in months. It is billed in full. It shows up nowhere as a problem, because nothing in the system flags the absence of use: the invoice is correct, the assignment is correct, only usage is missing, and nobody is watching for that.
- The optimistic rollout
- 100 licences purchased for a company-wide deployment. 40 people actually use it.
- The former employee
- The account is deactivated in the directory; the AI licence stays attached and billed. Offboarding never accounted for it.
- The duplicate
- The same person has a Copilot licence and a ChatGPT Enterprise subscription for the same use case.
- The expiring credit
- A pool purchased in advance, never consumed, that expires on its anniversary date.
What makes this opportunity particular: fixing it requires no technical tradeoff, no migration, no risk. It is inventory work. And the return is immediate and quantifiable, which makes it the best entry point for building credibility with finance before tackling anything else.
Managing without surveilling: the constraint that changes everything.
Measuring employees' AI usage means processing personal data. Under GDPR, that implies a legal basis, a minimisation principle, and information to the people concerned, and in some jurisdictions, consultation with employee representative bodies before rollout. A tool that reads conversation content to calculate cost is disproportionate under GDPR: cost is computed from volumes, not content.
- Measure volumes, not content
- Token count, model called, timestamp: that is enough to calculate a cost. Prompt text is never necessary.
- Pseudonymise by default
- Budget management happens at the team level. Individual identity is only useful in one specific case, reassigning a licence, and can be handled separately, behind its own access control.
- Document before you deploy
- Data processing record, retention periods, subprocessors, hosting location. A data protection officer who discovers the tool after the fact will block it, and they will be right to.
A useful side effect: a system designed this way lands much better with teams. A tool perceived as surveillance gets worked around, and you end up measuring nothing. That is the principle behind optidome's architecture.
Where to start: 90 days.
A realistic starting plan for an organisation beginning from zero.
Inventory
- List every AI vendor the company pays for. Check accounts payable, not just IT.
- Run the last six months of expense reports through a software-subscription filter.
- Pull the list of assigned licences from the directory, vendor by vendor.
- Identify who in the organisation owns this. If nobody does, the effort stops here.
Deliverable: one table with vendor, pricing model, monthly amount, owner, renewal date.
Measurement
- Cross-reference assigned licences against activation data. This is where the first real numbers come out.
- Break down API spend by team or application.
- Set a baseline: last month's spend, bucket by bucket.
- Reconcile the total against actual invoices. A gap above 5 % means you are missing a bucket.
Deliverable: an annualised waste figure, defensible in front of finance.
Decision
- Cancel or reassign the dormant licences identified.
- Set budgets by team and alert thresholds.
- Establish a 30-minute monthly review with a fixed agenda.
- Add AI licences to the offboarding checklist.
Deliverable: a realised saving, and a rhythm that runs without you.
The first three months are not about optimising. They are about producing numbers everyone believes. Everything else follows from that.
Do you need a tool?
Honest answer: not at the start.
The first 30 days run on a spreadsheet and invoice exports from each vendor's admin console. If your AI spend is under a few thousand euros a month across two or three vendors, a spreadsheet updated monthly is a rational choice.
A tool becomes worth it once three conditions stack up.
- More than three or four AI cost sources
- Not just LLM APIs: count named licences (Copilot, Agentforce, ChatGPT Enterprise) and the agentic features embedded in your SaaS tools too. Most reasonably equipped companies cross this threshold without noticing. Past it, manual reconciliation becomes most of the work.
- A need for continuity
- Waste regenerates constantly, at the pace of hires and departures. A one-off audit does not hold for six months.
- A need for alerting
- Catching an overrun at month-end means you are just observing it. Catching it on day 12 means you can act.
That is exactly what optidome does.
Pulling together costs through eleven read-only connectors, continuously detecting dormant licences and pricing them in euros, setting budgets and alerts, without ever storing conversation content, and hosted in France.
If you would rather start with a spreadsheet, the 90-day plan above works perfectly well without us.
AI FinOps: the questions that keep coming up.
- Is AI FinOps just cloud FinOps with a new name?
- No. The methodological framework is the same. It is literally the same framework, formally extended to AI in 2026. But the levers differ: pricing levers (commitment, volume) only come into play last and at high volume, cost per request is not deterministic, optimisation is applicative rather than infrastructural, and the mix of named licences and usage-based consumption has no real equivalent in the cloud.
- At what spend level does this become worth doing?
- Less a question of amount than of dispersion. Whether a company spends 5,000 € or 200,000 € a month, one that splits that spend across six vendors needs visibility more than one that concentrates it in a single contract. The usual trigger is the number of sources, not the total. A mid-market or enterprise company that has rolled out Copilot at scale, runs production APIs, and has a couple of agentic tools hits that threshold fast, regardless of budget.
- Who should own this inside the organisation?
- Usually IT for the technical data, finance for the budget tradeoffs. In organisations that already run a cloud FinOps practice, that team typically absorbs the AI side: the FinOps Foundation measures that 98 % of FinOps teams now manage AI spend, up from 31 % in 2024. Elsewhere, it often lands with IT Asset Management.
- Can you measure AI usage without violating privacy regulations?
- Yes, provided you measure volumes rather than content, pseudonymise by default, and document the processing before deployment.
- Should we wait for the market to settle before starting?
- It will not settle for a while: billing models remain inconsistent across vendors, and the FinOps and Linux Foundations have just announced an initiative to create open, unified standards for AI billing across providers. In the meantime, the one thing that costs nothing is the inventory.
Sources.
- FinOps Foundation, FinOps Framework 2026: Executive Strategy, Technology Categories, and Converging Disciplines, March 2026. finops.org
- FinOps Foundation, FinOps for AI (Technology Category). finops.org
- FinOps Foundation, State of FinOps 2026. data.finops.org
- FinOps Foundation, FinOps X 2026 Day 1 Keynote. finops.org