AI FINOPS
AI FinOps: get AI spend under control before it becomes the biggest line on the bill
Cloud FinOps gave you visibility over provisioned compute. AI gave you a cost shape it was never designed for: consumption billed by the token, no tags, several vendors, and a hidden cost stack that runs well beyond the model invoice. I have run this inside a live enterprise programme, where US$400,000 a month of AI spend sat alongside a US$500M cloud estate. I help you see it, attribute it, and defend it.
AI FinOps engagements delivered through Straitworks (Param Webeek Pte. Ltd.), Singapore.
98% of FinOps practices now manage AI spend, up from 31% two years ago. It is the FinOps Foundation's top forward-looking priority, and it arrived faster than most operating models could absorb.
The problem is not that AI spend is unmeasured. It is that it is measured at the wrong resolution. A finance team can see a model API invoice, a GPU reservation and an AI platform charge. It cannot see what a single claims triage agent, onboarding copilot or content generator costs to run. That gap is not a reporting problem. It is a data and governance problem, and it needs someone who has already built allocation across a large multi-cloud estate.
Method
What I do
AI spend visibility at the workload level
Most companies can produce a vendor invoice in five minutes. Almost none can produce a fully loaded cost per AI use case. I build the second thing: every production model, embedding job, retrieval call, guardrail hop, evaluation pipeline and human review loop attributed to one workload with one accountable owner. If a number cannot be produced in five minutes, it is not a cost model, it is accounts payable.
Unit economics instead of aggregate spend
"How much did we spend on AI this month" is a board slide. "What does it cost us to resolve one ticket, triage one claim, generate one qualified lead" is a management tool. I build the cost per outcome model, then use it to turn model selection, caching, batching and prompt architecture from engineering preferences into business decisions.
Model rightsizing as a running discipline
FinOps Foundation's 2026 guidance documents savings of up to 90% moving specialised tasks to small language models, and up to 50% shifting production inference onto purpose-built silicon. Those numbers are real and they decay without a process. I run the review: every production workload benchmarked quarterly against a cheaper option on accuracy, latency and cost, with the burden of proof sitting on the more expensive model.
Governance with a working kill switch
A budget envelope per workload, an alert at 70%, a circuit breaker at 100%, and a documented escalation path. Agent loops and viral internal features can burn a quarter of budget over a weekend. This is the cheapest item on the list and the one that prevents the incident that makes the whole programme political.
The data layer underneath
This is where my cloud background matters. AI cost is largely determined by the data the models reason over. Dirty or unmodelled data produces longer prompts, deeper retrieval, more reranking, more retries and more human review. I have spent three years building normalised cost and usage data across GCP, AWS and Azure, including FOCUS Specification normalisation, and the same discipline is what makes AI cost attributable at all.
Proof
AI cost analytics inside a running enterprise programme
Alongside the cloud estate, the same programme produced recurring analytics on AI spend running at approximately US$400,000 per month. Model API consumption, AI platform services and the associated data storage were brought under the same allocation and reporting discipline as everything else: attributed to owners, forecast, and reviewed on a monthly cadence that executives already relied on. Not a pilot, and not a one-off cost review. A live monthly line that had to reconcile.
The cloud estate underneath it
Three years, US$500M annual cloud spend across GCP, AWS and Azure brought under full visibility. Fourteen business units, 9,000+ cloud projects, a 13 person team spanning solution architecture, data engineering, analytics and application development. 60 to 65% savings against list price on commitment purchases, renegotiated against real usage rather than vendor projections. Built on the FOCUS Specification, which is the same data model AI cost reporting now has to extend.
Why that matters for you
AI spend does not arrive on its own invoice. It arrives inside the cloud bill, the data warehouse, the AI platform subscription and three vendor accounts that do not reconcile. Getting it attributed requires cloud cost engineering first. That is the part I have done at scale.
Working With Me
How an engagement works
AI Spend Visibility Audit (fixed fee, 3 weeks)
Inventory every AI workload in production and pilot. Map each to one owner and one business outcome. Produce a consolidated cost view across model APIs, AI platform services, storage and orchestration. Identify the top five workloads by spend and the top five by unbudgeted spend, which are rarely the same list.
Fractional AI FinOps lead (monthly retainer)
I run the cadence: monthly cost review, quarterly rightsizing, forecast tied to business volume drivers rather than vendor commitments, and a standing slot in your operating rhythm. Delivered alongside your engineering and finance leads, not in place of them.
Governance and instrumentation build (project)
Budget envelopes, circuit breakers, gateway level attribution, tagging and allocation rules, and the dashboards your executives and engineers will actually open. Two audiences, two views.
All work is delivered on a B2B basis through Param Webeek Pte. Ltd., based in Singapore, working across APAC, UK and US time zones.
Who this is for
AI and platform leaders at mid market and enterprise organisations who have moved from AI pilots into production and just received the first invoice that does not reconcile.
Particularly relevant if you are in Singapore or APAC and AI spend is now a line item your leadership is asking about quarterly rather than annually.
Not the right fit if your total AI spend is under a few thousand dollars a month. At that scale the answer is a budget alert, not a programme.
Frequently asked questions
How is AI FinOps different from cloud FinOps?
Cloud FinOps allocates provisioned capacity billed by the hour. AI FinOps allocates consumption billed by the token, across vendors that do not share a billing model, where one user request can trigger ten downstream calls. The framework transfers. The instrumentation does not.
Have you actually done this, or is it adjacent work?
At the reporting and governance layer, inside a live programme: roughly US$400,000 a month of AI spend attributed to owners, forecast and reviewed monthly, sitting inside a US$500M cloud estate. That is the layer where most enterprises are stuck today. Prompt level optimisation inside an application team is a joint engagement with your engineers, and I will say so rather than sell around it.
Do we need new tooling?
Usually not a new platform. Attribution normally comes from an AI gateway plus prompt level observability, joined to your existing cloud cost data. I am tool agnostic and will work with what you already own before recommending anything new.
Can you work with our existing FinOps team?
That is the normal shape. I extend the team into the AI cost surface rather than replacing anyone.
How quickly can this start?
The audit is three weeks from kickoff. A scoping call is enough to tell you whether it is worth doing.
Next Step
If AI spend just became a line item nobody can explain
That is the conversation I have most weeks. Book a 30 minute scoping call.
Book a 30 minute scoping call© 2026 Harshavardhan Shinde. All rights reserved.