Measuring AI ROI: A CFO's Challenge

Is your organization's AI spend wasted improving transition sentences for internal memos or is it driving real value for your customers? Justin Moore is spot on. The companies that can tell the difference will have a huge advantage.

When just 14% of CFOs say they can see the outcome of their investment, that doesn't necessarily mean they're investing in the wrong places. It's just about tying spend to outcomes – which is what CloudZero can do for the trillions of dollars of AI spend we're seeing worldwide. Read on and see what CloudZero's CEO has to say!

The invisible invoice: understanding the true cost of AI The AI bill is larger than most product teams realise. And we are not even seeing the full invoice yet. The tools we use daily are still operating in an industry where the economics are far from stable. Reports suggest OpenAI spent approximately $1.69 for every dollar it generated in revenue in 2025. Investor reports around Anthropic pointed to deeply negative gross margins in 2024. Much of today’s AI pricing appears optimised for rapid adoption while the long-term economics are still maturing. We have seen this playbook before. Cheap cloud. Subsidised rides. Discounted deliveries. The product becomes embedded into workflows long before the business model fully stabilises. What companies pay today may only be part of the long-term cost picture. Here is what that picture actually looks like. - Token costs scale invisibly. Every request, workflow, and agent loop consumes tokens. As usage grows, so does the bill — quietly, in the background, often with no one watching. Enterprise AI cloud spend tripled in a single year despite token prices falling sharply. Cheaper tokens did not reduce spending. They accelerated consumption. - Agent costs are unpredictable. An AI agent does not complete a task in one step. It loops. It retries. It feeds outputs back into inputs. A poorly scoped workflow can run dozens of unnecessary cycles before resolving — each one burning budget, each one invisible to the team that approved the feature. - Infrastructure costs surprise teams at scale. An AI feature that feels affordable in a pilot can behave very differently once real users, larger datasets, and higher request volumes enter the system. More users mean more inference calls. More automation means more background processing. More AI features often mean more storage, monitoring, logging, and compute running continuously behind the scenes. The challenge is not just that AI is expensive. It is that the cost curve often scales faster than teams initially expect. - Operational Cost There are the operational costs most business cases barely account for. AI engineers. Evaluation workflows. Human review systems. Monitoring pipelines. Governance layers. A support team manually verifying AI-generated responses before they reach customers. An agent looping endlessly because an edge case was never scoped properly. A workflow becoming slower, not faster, because human review is now required at every critical step. These are not edge cases. They are increasingly becoming part of the operational reality of AI at scale. AI is genuinely powerful. The opportunity is real. But the operational reality is also more complex, more expensive, and more fragile than most pitch decks suggest. As PMs, the question worth sitting with is not whether to use AI. It is whether we fully understand the cost structure of how we are using it right now. Does anyone on your team actually own the AI cost number?

Most AI revenue coverage is asking the wrong question. Yes, the headline numbers are real, the growth is real. What's almost entirely missing is the question any seed investor asks a founder in the first ten minutes: what color is this revenue? A few data points worth sitting with: - OpenAI: roughly 70-75% of revenue from consumer subscriptions. 800M+ weekly users, ~5% paying. Roughly $25 per weekly user. - Anthropic: roughly 80% enterprise. 500+ companies spending $1M+ annually on Claude. Most of the Fortune 10. Claude Code from zero to $2.5B annualized in nine months. Roughly $211 per monthly user. Same industry and headlines but structurally different businesses: One is built on attention while the other is built on integration. When pressure hits a business, attention evaporates and integration doesn't. Consumer subscriptions cancel quietly and fast, but enterprise contracts wired into compliance or codebases require an engineering project to unwind. The 2022 SaaS correction was a 'rehearsal'. Multiples fell 67% on average and inside that average, some companies fell 90% while infrastructure and security held. It's important to remember that the crushed ones weren't necessarily bad businesses; they had the wrong color revenue for the moment. AI will produce a sharper version of the same split, because the hype cycle is bigger and the consumer share is larger. Worth noting: in the few days since I published this piece (https://lnkd.in/eJBtXUgk) in ComputerWeekly.com, Anthropic disclosed it had hit $30B in annualized revenue, up roughly 80x year-over-year. The gap I'm describing is only widening...

IT companies using artificial intelligence models at scale are likely to benefit as AI firms move from the per-seat to usage-based model of pricing, though they may face some near-term margin concerns, industry experts said. Demand for productivity tools such as Anthropic’s Claude Enterprise and GitHub has increased over the past year. The higher demand increases inference or compute cost for AI firms, which say it is not sustainable in the long run for them to absorb this. To address this issue, Anthropic and GitHub have moved from fixed price to consumption-based models. For large IT companies, which are big consumers of AI technologies as they restructure their business, this triggers additional margin pressure in the short term. But in the long term, IT companies are expected to benefit from economies of scale, said industry experts. Read this, and more in today's Eye on AI page on ET https://lnkd.in/gKGuiEiB

I have been following the conversation around AI costs, deployment and ROI for months, and I think we are asking the wrong question. Every executive panel and board meeting is obsessed with: Is AI worth it? The real question we should be asking is: Who is actually keeping the money? The recent data by tech firms is a wake-up call. - Uber burned through its entire 2026 AI budget by April because engineers were using agentic coding tools. - Microsoft quietly began winding down most internal Claude Code licenses this month after token-based billing proved unsustainable at scale. - Bryan Catanzaro, NVIDIA's VP of Applied Deep Learning, recently said in an interview with Axios, "For my team, the cost of compute is far beyond the costs of the employees." This is not a technology problem. It is a governance problem that was always coming. ➡️ The pattern holds across industries: - 80-85% of enterprises miss AI cost forecasts by 25% or more - Fewer than 1% of executives report ROI of 20% or greater - 60% of organizations see minimal or no measurable value from AI - 30% of GenAI projects are abandoned after proof of concept This is not a random variance. It reflects a structural failure in how AI has been deployed: open access, token-based consumption pricing, and no enterprise-grade financial controls applied to rollouts that moved fast in late 2025. The same costs straining #CFO budgets are fueling one of the fastest revenue ascents in enterprise technology history. Anthropic is getting close to a $1 trillion valuation with about $45 billion in yearly revenue run rate, mainly from companies spending heavily on AI. However, at this stage, much of the financial value appears to be captured by AI vendors rather than the companies adopting the technology. While AI clearly has the potential to create value, organizations need the right discipline and management practices to realize and capture those benefits themselves. ➡️ If you want to flip that dynamic, you have to stop treating AI as an IT experiment and start managing it like a major capital deployment. That requires three things: 1. Do you have consumption governance: quotas, model routing, unit economics tracked by function? 2. Can you articulate AI ROI at the business unit level and not just cite headline productivity claims? 3. Are your AI investments structured around measurable outcomes or around access and adoption metrics? The enterprises that lead this decade will not be the ones that spent the most. They will be the ones that applied the same financial discipline to AI that they apply to any major capital deployment: Clear ownership, Clear accountability, and Measurable results. What are you seeing in your organization? Are #AI investments delivering actual results, or are you still waiting to see proof of the value?

How much will you pay for AI tomorrow? Today the answer still feels comfortable. The subscription is cheap enough, the API is fast enough, the model quality is good enough, and the whole thing looks like obvious productivity gain. So companies keep replacing people, processes, junior training, documentation and internal knowledge with AI calls. In many cases, this really does make teams faster. I use AI every day, and I am not interested in pretending that the old way was somehow nobler. But there is a difference between using AI as leverage and rebuilding your operating model around a rented dependency you do not control. The uncomfortable part is that nobody really knows what AI should cost. A large part of the market is still living on investor money, strategic subsidies and the belief that scale will eventually make the economics work. Maybe it will. Maybe inference gets cheaper, hardware improves, smaller models become good enough, and we all get a stable, affordable automation layer. But there is another scenario. The model your product depends on is deprecated. The new one is slower. Or more expensive. Or behaves differently enough that your prompts, QA and user experience need to be rebuilt. The context window gets larger, your product quietly burns more tokens, and what used to look like cheap automation becomes an enterprise-grade invoice. By that point, the real damage may not be technical. The API can be replaced. The harder thing to replace is the engineering muscle you already optimized out. You may no longer have the people who understood the boring parts of the system. You may no longer train juniors to debug without assistance. You may no longer keep enough documentation because “the model can explain it”. Your workflows, hiring plans, product assumptions and quality habits may already be shaped around the idea that AI is permanently available, cheap and good enough. That is the lock-in people underestimate. It is not only vendor lock-in. It is capability lock-in. The dependency moves from your codebase into your organization. So the question is not whether companies should use AI. Of course they should. Refusing AI because it creates risk would be like refusing electricity because power cuts exist. The real question is whether your company can survive AI becoming slower, more expensive, restricted, degraded, or unavailable. That means designing for optionality from the beginning. Keep critical logic outside model-specific prompts. Test model changes like database migrations. Track AI cost per operation, not only monthly spend. Preserve engineering skill, especially in juniors. Evaluate smaller models for routine tasks. Consider local, private or self-hosted inference where it makes business sense. And do not let one provider become the invisible foundation of your whole delivery process. AI is not magic. It is infrastructure without an exit strategy. Just a beautiful demo slowly turning into a very expensive trap.

If your team built an internal AI tool this year, you're probably burning 60%+ of your inference budget on the same text — over and over. The model isn't reading it any closer the second time. Here's what's happening. Every call re-sends the system prompt. Re-sends the tool definitions. Re-sends the first 80% of the conversation. None of that changed. You paid for it last call. You're paying for it again right now. Both Anthropic and OpenAI offer prompt caching. After the first call, the cached portion costs ~10% of the normal token price. The math adds up fast: a chatbot with a 4,000-token system prompt, called 1,000 times a day, is paying for 4M tokens of repeated text every single day. Think of it like ordering the same dish from your favorite restaurant every day, and getting charged for the recipe every single time. The recipe didn't change. The bill should. The fix isn't a rewrite. It's a flag. You mark the stable portion of the prompt as cacheable, the API does the rest. ProjectDiscovery published a production case study showing a 59% cost reduction, growing to 70% after a few rounds of optimization. We've seen similar numbers on agent workflows we've audited. Three things to audit this week: Your system prompt — same on every call? Cache it. Your tool/function definitions — change per call? If not, cache them. Your conversation context — first half stable? Cache it. If you're shipping AI in production and the cloud bill is starting to feel weird, my DMs are open. Curious what other teams are catching in the wild.

It's not just the AI Era. We're in the AI ROI Era and CloudZero is enabling CFOs, CEOs, Engineering leaders and others to answer the question: "What are we getting for all this spend? What's our ROI?"

The world will spend $2.6 trillion on AI this year. Only 14% of CFOs can measure the return. AI bills show what you're consuming, not what you're producing. Every board is asking whether their AI investment is paying off. None can answer for sure. Today, that changes. The financial control plane for AI is here. The technology isn't new. We spent ten years building the allocation engine that gives hundreds of enterprise customers a clear view of their cloud economics. We process 14 trillion billing events a year. What's new is connecting it to AI. It took three months of the most concentrated work I've seen in my career to bring it to market. The goal isn't to spend less on AI. It's to put more behind what works. Welcome to the AI ROI Company. Welcome to the new CloudZero. Justin makes the case → https://lnkd.in/dTP68-_D #AIROI #FinancialControlPlane #CFO