This year, Yarken ran a closed-room roundtable at ITC Europe in Barcelona. No slides, no pitches. Just senior technology and finance leaders from insurance and financial services talking honestly about what's actually going on in their organizations.

We focused on AI economics: where the spend is going, why it's so hard to manage, and what needs to change before it becomes a problem that's much harder to solve. We opened with a simple question.

"Who here can tell me, within ten percent, what their organization is spending on AI this month?"

Not one hand went up. These were Heads of, Directors, and C-suite leaders from some of Europe's largest insurance and financial services firms. People who own real budgets, real P&Ls, and real accountability to the board. And not one of them could answer the question.

That moment set the tone for everything that followed. We were not here to talk about AI strategy. We were here to talk about money: where it's going, why it behaves differently from anything these organizations have managed before, and what to do about it before the bill becomes structural. Here's what came out of the room.

The numbers nobody is tracking, but should be

Let's start with the market reality, because the scale of what's coming is still not landing for most organizations. Banking and financial services GenAI spend is forecast to reach $85 billion by 2030, a 1,400% increase from where it sits today. Token and inference spend alone is expected to grow from roughly $1-3 billion to somewhere between $20-35 billion over the same period. That's a CAGR of 55-60% across banking AI spend, compounding every year.

The dynamic that makes this so difficult to manage: token unit prices are falling sharply, projected down 90% by 2030, but token consumption is growing 50 to 100 times faster. The bill keeps rising even as the cost per token drops. Volume is winning, by a long way. Most organisations are not equipped for this. The governance frameworks, the attribution models, the forecasting tools have not kept pace with the spend. And the spend is not waiting.

Two doors. Most organisations are only watching one.

We introduced a framing in the room that landed hard: the AI cost problem has two doors.

The experimentation trap

One of the clearest tensions to surface in the discussion was between the mandate and the reality. "AI first" is the directive coming down from boards and leadership teams across the sector. Move fast, experiment, adopt. The pressure is real. But the frameworks to manage what that adoption actually costs—in tokens, in compute, in embedded SaaS charges—have not kept pace with the mandate. As one participant put it: organizations are pushing inefficient workflows and calling it innovation.

Token consumption during the experimentation phase is high by design. You're running prompts that aren't optimized, using frontier models when a cheaper one would do, generating output that gets thrown away. That's fine. That's how you learn. The problem is when there's no mechanism to track what the experiment cost, no way to compare actual spend to the business case assumption, and no clear owner of the number when the CFO asks.

Several participants acknowledged this directly: business cases are being built with token cost assumptions that don't hold post-deployment. The POC is constrained. The production rollout isn't. Without tracking in place, the gap between the two is invisible.

Agentic AI: the cost you really can't predict

First-generation AI has a manageable cost structure. A prompt, a response, a token count. You can model it. Agentic AI is different. An agent plans. It calls tools. It spawns sub-agents. It retries failures and expands its context window as it works. The same business task can cost three times more on a bad run than a good one, and the variance is driven by factors that are hard to forecast and sometimes impossible to control.

In financial services, the highest token-intensity use cases are exactly the ones this sector is scaling fastest: AI software engineering, compliance and legal review, KYC and AML workflows, claims triage. By 2027, agentic workflows are expected to be the primary driver of sustained inference load across the industry. The cost management problem is not coming. It is here.

When we asked the room whether anyone had built cost guardrails into AI workflows before they hit production, meaningful silence again. The idea of a "kill switch"—a mechanism to halt a runaway AI process before it burns through budget—came up multiple times. It isn't exotic. It's basic financial control. Almost nobody has it yet.

The governance gap nobody has filled

The question of who owns the AI cost produced one of the most revealing exchanges of the session. Finance? The CTO? Engineering? The business unit that requested the feature? In most organizations the answer is functionally nobody, or worse, everybody, which amounts to the same thing. AI spend sits in a governance gap. It's technically visible somewhere, usually split across several cost centers, billing reports, and P-card line items, but practically unmanaged as a single number.

Three management questions are already forming in the market, and they will define AI financial governance for the next decade:

By 2027-2028, these will be standard questions in every major FS firm's technology economics model. By 2030, token spend will be governed the way cloud consumption is today: forecast, allocated, optimized, and linked directly to business value. The organisations building that capability now will have options. The ones waiting will be managing it reactively, at scale, under pressure.

The question that matters now

We closed every session the same way: one commitment, one owner, one timeframe. Something real, something doable, something you'd actually follow through on.

Most of the room left with a version of the same action. Get a number. Any number. A starting point for visibility that doesn't currently exist.

That's where this begins. AI spend is already material in most of these organisations, and it is growing faster than the controls around it. The governance gap is not a future problem to solve when things settle down. Things are not going to settle down. The spend will keep climbing, the vendor contracts will keep repricing, and the board will eventually ask for an answer that nobody has prepared.

The organizations that build financial controls around AI now are the ones with options later. Whether you're building agents at scale or managing what your SaaS vendors are quietly embedding into your stack, the question is the same: