The AI FinOps Category Map: Three Approaches to a New Problem
Enterprise AI spend has grown faster than the tooling built to govern it. In 2024, most organisations ran one or two AI tools and managed costs manually. In 2026, the average enterprise runs six to eight AI models across three to four vendors simultaneously — each with different billing models, different token counting methodologies, and different pricing change cadences.
A new category of platform has emerged to help. But "AI FinOps tool" now describes at least three distinct approaches, each built for a different layer of the problem. Understanding which layer your organisation needs to solve determines which approach fits.
Category One: Cloud-Native FinOps Extended into AI
These platforms were built to manage cloud infrastructure costs — compute, storage, networking across AWS, Azure, and GCP. They are mature, enterprise-proven, and increasingly extending their coverage to include AI API costs.
What they do well: unified cloud + AI API visibility for organisations running AI workloads on managed cloud infrastructure. Strong on tagging, attribution, and chargeback for engineering teams.
What they were built for: infrastructure spend — billed by the compute second, the storage gigabyte, the API call. The mental model is a cloud bill.
The buyer profile that fits: organisations whose primary AI cost question is "which cloud infrastructure is our AI running on, and what does it cost?" Teams running AI training workloads, GPU clusters, or large-scale inference infrastructure on cloud compute.
Category Two: LLM Observability and Token Analytics
These platforms were built to instrument AI model calls — tracking tokens, latency, model versions, and prompt patterns at the API request level. They emerged from the developer tooling ecosystem and have grown upward into cost analytics.
What they do well: granular token-level visibility, prompt performance tracking, cost per feature or cost per customer attribution for teams building AI-native products.
What they were built for: API-billed workloads where the primary question is "how many tokens did this feature consume, and what did it cost per user?"
The buyer profile that fits: engineering and product teams building AI-powered products on top of LLM APIs — where understanding model behaviour and cost per inference is the core problem.
Category Three: AI Subscription FinOps
This is the newest layer — and the one most directly tied to how enterprises actually deploy AI in 2026. Most enterprise AI spend doesn't start with infrastructure or API calls. It starts with a seat licence: a Claude Max subscription, a Microsoft 365 Copilot licence, a GitHub Copilot Business plan.
The problem at this layer is different. It's not "how many tokens did our infrastructure consume?" It's "which of our 200 licenced seats are actually being used, at what plan tier, and are we paying for the right plan for each person's actual behaviour?"
The questions this category answers:
Which seats are ghost seats — licenced but idle?
Which users are on Max plans consuming Pro-level usage?
When a vendor silently changes pricing or billing models, which seats are most exposed?
What is our recoverable budget — the spend we could eliminate tomorrow without any loss of capability?
What this layer requires: per-seat behavioural data, not invoice aggregation. Plan utilisation rates, not just total spend. Rightsizing recommendations with confidence scores, not just cost reports.
The buyer profile that fits: IT Directors and CFOs managing multi-vendor AI licence portfolios across knowledge workers, developers, and operations teams — where the primary waste is in the subscription layer, not the infrastructure layer.
Which layer is your problem?
Most organisations have costs at more than one layer. The question is where the primary waste lives — and therefore which layer to instrument first.
A practical diagnostic:
If your AI cost question starts with "which cloud region is this running in" → you have an infrastructure layer problem.
If your AI cost question starts with "which feature or customer is driving these API costs" → you have a token analytics problem.
If your AI cost question starts with "we have 200 AI licences and I don't know which ones are being used" → you have a subscription FinOps problem.
Most enterprise AI waste in 2026 lives in the third category. Industry data shows 35% of AI seats are effectively idle — licenced, billed monthly, and generating near-zero usage. At $100–$200 per seat per month for premium AI plans, that waste compounds quickly.
The AI FinOps category is still being defined. The organisations building the discipline now — mapping their vendor landscape, instrumenting their seat layer, and establishing the metrics that matter — will set the standard for how this field works. The tools exist. The framework exists. The gap is execution.
See your organization's AI spend data
PromptKing connects to your AI vendors and surfaces exactly this analysis — for your seats, your vendors, your budget.