There’s a conversation happening in boardrooms right now that goes something like this:
“We need to invest more in AI. Where’s the budget coming from?”
“Find it in your existing spend.”
If that sounds like being asked to fund your kitchen renovation by finding loose change in the couch, you’re not alone. But the data suggests something more interesting is going on and the companies that figure this out will have a structural advantage over those that don’t.
The State of FinOps 2026 report, published by the FinOps Foundation, surveyed hundreds of practitioners across the global FinOps community. One finding stood out above the rest: organisations are increasingly being asked to self-fund AI investments through optimisation savings. They’re squeezing existing technology budgets to create headroom for AI and it’s working, but only for teams that have the right visibility.
The Circular Logic Problem
Here’s the paradox at the centre of this strategy, you need to optimise AI spend to fund more AI investment but you can’t optimise AI spend because you can’t see it.
The State of FinOps report confirms both sides of this coin. 98% of organisations now manage AI spend, up from just 31% two years ago. That’s an extraordinary acceleration. But at the same time, AI cost management is the #1 skillset gap that FinOps teams are trying to fill, more than any other capability.
Almost everyone is trying to manage AI costs, and almost nobody has the tools or expertise to do it well.
This creates a specific, frustrating dynamic. The CFO says “find savings to fund AI.” The FinOps team looks at their cloud optimisation dashboard and sees diminishing returns, the report quotes practitioners who’ve reached 97% optimisation on their existing cloud footprint. The big rocks are gone. What’s left is a high volume of smaller opportunities that require more effort to capture.
Meanwhile, the AI budget line item is growing 40–60% quarter over quarter, and nobody has task-level visibility into where that money is going.
The big rocks haven’t been found yet in AI. They’re sitting right there, waiting.
Where the AI Savings Actually Are
If you’ve read our piece on model arbitrage, you know the headline: most enterprises route everything to frontier models regardless of task complexity. Classification tasks running on Opus. Sentiment analysis on GPT-4o. Data extraction on models that cost 20–50x more than what the task actually requires.
Let’s put real numbers to this. Here’s what a typical self-funding cycle looks like for a company spending $25K/month on LLM APIs:
Phase 1: Discover ($0 effort, immediate visibility)
| Task Category | Current Model & Cost | Optimal Model & Cost |
| Classification (35% of calls) | GPT-4o -> $8,750/mo | Haiku 4.5 -> $525/mo |
| Summarisation (25% of calls) | GPT-4o -> $6,250/mo | Sonnet 4.5 -> $1,875/mo |
| Extraction (20% of calls) | GPT-4o -> $5,000/mo | Sonnet 4.5 -> $1,500/mo |
| Complex reasoning (20%) | GPT-4o -> $5,000/mo | Keep GPT-4o -> $5,000/mo |
| Total | $25,000/mo | $8,900/mo |
Monthly savings identified: $16,100. Annual: $193,200.
Note what happened here, we didn’t touch the complex reasoning tasks. We didn’t degrade quality. We simply identified that 80% of API calls were using a model that was dramatically overqualified for the task. This is the AI equivalent of the “grossly misconfigured instance” that cloud FinOps teams found five years ago.
Phase 2: Reinvest (the self-funding loop)
Now it gets interesting. You’ve freed up $16,100/month. That’s not theoretical, it’s real budget that’s already been approved and allocated. It’s just being spent badly.
| Reinvestment | Monthly | Annual |
| New agentic workflows (customer success, internal ops) | $6,000 | $72,000 |
| AI tooling & platform investment (PromptLeash, observability) | $2,000 | $24,000 |
| Experimentation budget (new use cases, POCs) | $3,000 | $36,000 |
| Returned to business (net cost reduction) | $5,100 | $61,200 |
The company now has 3 new AI-powered workflows, a proper AI governance platform, and an experimentation budget, all funded by optimising spend that was already approved. And they’re still returning $61K/year to the bottom line.
This is the “squeeze more from existing footprint to create space for AI spend” dynamic the State of FinOps report describes. But it only works if you can actually see where the waste is.
Why Cloud FinOps Tools Can’t Do This
The State of FinOps report shows that FinOps teams are applying the same maturity path to AI that they used for cloud: first gain visibility, then build planning discipline, then optimise for value. Allocation, forecasting, and reporting are the top-priority capabilities across every new technology category.
| AI Cost Dimension | Why Cloud Tools Miss It |
| Token-level pricing | Input vs output tokens, prompt caching, variable pricing per model, none of this maps to cloud billing constructs |
| Model-task mismatch | Cloud tools see total API spend. They can’t tell you that 60% of your calls are classification tasks running on a $15/M token model when a $0.25/M model would do |
| Agentic loop costs | Agents spawn sub-calls, retry, and loop. A single user request can trigger 50+ API calls. Cloud billing shows aggregate spend, not per-agent economics |
| Cross-provider routing | Most enterprises now use 3-4 LLM providers. Cloud tools give single-provider dashboards. There’s no unified task-level view across OpenAI + Anthropic + Google |
| Cost per business outcome | The metric that matters is “what does it cost to classify one ticket” or “what does it cost to summarise one report” not total monthly API spend |
The problem is that existing FinOps tooling was built for cloud infrastructure. It understands compute hours, storage GB, and network egress. It does not understand token-level cost attribution, model-task mismatch, or agent loop economics.
The Self-Funding Flywheel
The companies doing this well aren’t treating it as a one-time cost reduction exercise. They’re building a continuous flywheel where AI spend visibility creates savings, savings fund new AI capabilities, and new capabilities generate business value that justifies further investment.
Quarter 1 — Visibility: Connect LLM providers to a cost analytics layer. Surface cost-per-task, model utilisation, and routing opportunities. Identify the low-hanging fruit (usually 30–50% of spend is immediately optimisable).
Quarter 2 — Optimise: Implement smart routing. Move classification, extraction, and summarisation to cheaper models. Set budget guardrails and circuit breakers. Capture $10–20K/month in savings for a company at $25K/month spend.
Quarter 3 — Reinvest: Deploy savings into new use cases. Fund 2–3 new agentic workflows. Expand AI adoption across business units that were previously blocked by budget constraints.
Quarter 4 — Compound: New workflows generate new optimisation opportunities. More data flowing through the system means better routing decisions. Cache hit rates improve. Model prices continue falling (10–30% every 6 months). Margins expand automatically.
The Numbers the Board Wants to See
If you’re in the room when AI investment decisions are made, here are the five numbers that make the self-funding case: current monthly AI spend, percentage of spend that is optimisable, monthly savings achievable, reinvestment allocation plan, and projected ROI on reinvested savings.
| Metric | What It Proves |
| Model-Task Mismatch Rate | % of API calls using overqualified models. Typically 60-80%. This is the size of your savings opportunity. |
| Cost Per Successful Completion | What each task actually costs, not blended averages. Shows exactly where routing saves money. |
| Optimised vs Actual Spend | Side-by-side comparison: what you’re paying vs what you would pay with intelligent routing. The gap is your self-funding budget. |
| Net New Capabilities Funded | # of new AI workflows, agents, or use cases funded directly from optimisation savings. Proves the flywheel is turning. |
| AI ROI Per Dollar Optimised | For every $1 saved through routing, how much incremental business value was generated by reinvesting it? This is the compound metric. |
Why This Matters More for Lean Teams
One more finding from the report worth highlighting: team sizes remain lean. Even organisations managing $100M+ in cloud spend average just 8–10 FinOps practitioners. They scale through automation and federated champions, not headcount.
This means the self-funding approach needs to be automated. You can’t have a team of three manually auditing LLM API logs to find routing opportunities. The tooling needs to do the heavy lifting: automatically analyse traffic, identify mismatch, recommend routes, and surface savings opportunities without human intervention.
The AI Budget Isn’t Coming From Somewhere Else
The State of FinOps 2026 report makes one thing clear: the era of AI as an experimental line item is over. 98% of organisations are now managing AI spend. It’s a real budget category that needs real governance.
But the report also reveals the uncomfortable truth that most teams aren’t being given new budget for AI. They’re being told to find it. And the cloud optimisation well is running dry.
The irony is that the biggest optimisation opportunity in most organisations’ technology stack right now is their AI spend. It’s the newest, fastest-growing, and least-governed category. It’s full of big rocks that nobody’s picked up yet because they don’t have the visibility to see them.
The companies that build the self-funding flywheel will outpace their competitors not because they spend more on AI, but because every dollar they spend works harder.
That’s not a FinOps strategy. That’s a competitive advantage.