How Do I Actually Lower My AI Spend? A Practical Guide for Finance Teams

There’s a number floating around that should make every CFO pay attention. Visa — a payments company, not an AI lab — is now consuming 1.9 trillion AI tokens per month. That’s double what they were using just one month earlier. Eighty-nine per cent of their employees are active AI users, and nearly half qualify as power users, averaging 25 prompts a day.

Now zoom out. If a payments company is scaling AI this aggressively, what’s happening across financial services, fintech, and the thousands of companies building on open banking infrastructure? The answer: AI spend is exploding — and most organisations have no framework for controlling it. They have cloud budgets. They have procurement processes for SaaS tools. But AI consumption doesn’t fit neatly into either category. It’s usage-based, unpredictable, and growing faster than anyone forecasted.

This article is for the people who are staring at a cloud invoice that’s 40% higher than last quarter and trying to figure out what happened — and more importantly, what to do about it. Not the theoretical “AI governance framework” stuff. The practical stuff. What actually works.

When you buy a SaaS product, you know what it costs. Ten seats at $50 per month. Done. Predictable. AI doesn’t work this way. API-based AI services charge per token — per input and per output. The cost of a single API call depends on which model you use, how long the prompt is, and how long the response is. A simple classification task might cost fractions of a cent. A complex document analysis task using a frontier model could cost several dollars per request.

This means your AI bill is a function of adoption — and adoption is the one thing your organisation is actively trying to increase. Every time a team builds a new automation, connects a new workflow, or rolls out an AI feature to customers, consumption goes up. The more successful your AI strategy is, the higher the bill gets. That’s the paradox finance teams are wrestling with right now.

These aren’t hypothetical. These are the levers that actually move the number.

This is the single biggest cost lever most companies ignore. Not every task requires a frontier model. Classifying a support ticket? A smaller, cheaper model handles that just fine. Summarising a short email? Same. Generating a complex financial analysis from raw data? That’s where you want the expensive model. The mistake most teams make is defaulting everything to the most powerful (and most expensive) model because it was the first one they integrated. Build a routing layer that sends simple tasks to cheap models and complex tasks to premium models. The cost difference can be 10–50x per request.

If the same prompt produces the same output, you’re paying twice for the same answer. This is surprisingly common in production AI systems — especially in customer support, document processing, and data enrichment pipelines where the same types of queries repeat frequently. Implement a semantic caching layer that stores responses and serves them again when a sufficiently similar prompt arrives. Some teams report 20–40% cost reductions from caching alone.

Longer prompts cost more — both on the input side and typically on the output side too, since verbose prompts tend to produce verbose responses. Review your production prompts and strip out unnecessary context, redundant instructions, and padding. A well-engineered prompt that’s 40% shorter produces the same quality output at 40% lower cost. This sounds trivial but across millions of API calls per month, the savings compound fast.

Most AI providers offer batch APIs at significantly reduced rates — sometimes 50% cheaper than real-time inference. If your workload doesn’t need an instant response (think: overnight document processing, daily report generation, weekly data classification), batch it. The results are identical; you just wait a few hours instead of a few seconds. For back-office financial operations, this is often a trivial change that cuts the bill in half for those specific workloads.

This sounds obvious but a shocking number of organisations don’t do it. Every major AI and cloud provider allows you to set monthly spend caps and usage alerts. Set them. Set them per team, per project, and per environment. A runaway loop in a development environment or an unanticipated spike in production traffic can burn through thousands of dollars in hours. The alert doesn’t reduce your spend — but it stops a manageable bill from becoming an emergency.

Here’s something that doesn’t get talked about enough: many companies are sitting on cloud and AI credits they’re never going to use. Startup grants from Google Cloud, Azure, or AWS. Enterprise commitments that were over-provisioned. Promotional credits from provider partnerships. These credits expire — and when they do, the value is gone. If your organisation has unused Google Cloud capacity, you can sell Google Cloud credits through brokers who connect you with buyers looking for discounted capacity. It’s a straightforward way to recover cash from credits you’ll never consume instead of watching them expire on your balance sheet.

If your monthly AI spend exceeds $5,000 with any single provider, you almost certainly have room to negotiate. Contact your account representative and ask about volume pricing, committed use discounts, or custom rate cards. Most providers have these programmes but won’t proactively offer them unless you ask. And if your current provider won’t negotiate, explore alternatives — the model landscape is increasingly competitive and switching costs are lower than most teams assume.

Most organisations don’t have clear visibility into what’s driving their AI costs. Here’s a typical breakdown for a mid-sized company running production AI workloads:

The insight here is that production inference is usually the largest line item — but internal tools and development environments are often the fastest-growing categories, because they scale with employee adoption rather than customer demand. Visa’s numbers bear this out: their 44% power-user rate means almost half the company is generating significant token consumption every working day.

Let’s be honest about something. If your AI spend is going up, it probably means your organisation is doing the right thing. The companies that will struggle in two years aren’t the ones spending too much on AI — they’re the ones spending too little. Visa isn’t tracking token consumption because they want to cut it. They’re tracking it because they want to reward the teams that are using it most effectively.

The goal isn’t to minimise AI costs in absolute terms. The goal is to maximise the value you get per dollar spent. That means eliminating waste (duplicate calls, oversized models for simple tasks, uncached repeated queries) while continuing to invest in the AI capabilities that actually move your business forward.

For finance and operations leaders in financial services and fintech, this is a familiar challenge dressed in new clothing. You’ve been optimising infrastructure spend for years — cloud migration, API gateway costs, payment processing fees. AI is just the newest line item. The tools are different but the discipline is the same: measure it, allocate it, and relentlessly cut the waste while protecting the investments that generate returns.