Generative AI has moved from a few experiments in data scientists' notebooks to a major line item on most organizations' cloud bills. As a data engineer who has built the systems behind that spend, I have observed this shift happen firsthand. The costs for GPU hours, model inference, vector storage, and agent orchestration add up very quickly. Yet AI spend is often unpredictable and hard to attribute to teams, and even harder to tie to the value those dollars create. Most organizations are using GenAI this way, yet few can answer a simple question: what do all these dollars buy, and is the spend even worth it?

Why GenAI Breaks Traditional FinOps

Traditional FinOps for cloud infrastructure is to tag resources, aggregate costs, and present them in a dashboard. This approach is not sufficient for GenAI, because:

• Attribution is fuzzy: A shared model endpoint or vector database serves many teams, so a tag tells you what the platform cost, not which project or initiative consumed it.

• Spend is probabilistic: Because GenAI services scale with usage, the same initiative can cost very differently from one month to the next. Spend therefore needs to be viewed in probabilistic terms. Those costs are driven by GenAI-specific factors: token consumption, model selection and routing, inference frequency, context-window size, embedding generation, vector database utilization, agent loops, retrieval volume, GPU utilization, fine-tuning, and caching. Each moves the economics directly: a modest change in context size or an extra agent loop per request can shift an initiative from clearly worthwhile to barely justified.

• Value is uncertain: Most initiatives funded with GenAI compute spend are bets on whether they will generate value, in the near term or the long term. Some will generate cash soon, while others build capability that pays off later. Treating them all as equal cost centers hides their true economics.

Traditional FinOps asks whether cloud resources are being consumed efficiently. GenAI requires an additional question: whether the AI workloads consuming those resources are economically justified. The reframing that worked was to treat GenAI spend as an investment portfolio, with each initiative an investment carrying a predicted return, an associated compute cost, and a confidence level. FinOps then becomes the discipline of directing compute toward the initiatives with the best risk-adjusted return per dollar.

A Framework Leaders Can Govern

Here is a practical framework organizations can use, built around design choices that matter to decision makers rather than to engineers.

• Metadata-driven intake: Financial fields and formulas live in a simple mapping file. Finance can modify the model in the mapping file instead of in code, and changes to the form then take effect automatically.

• A single comparable number: Each initiative is reduced to risk-adjusted Net Present Value (NPV) per dollar of compute. It is computed per return type (for example, subscription or efficiency), with all returns discounted to a common present value so initiatives can be ranked on the same scale.

As a hypothetical example, consider a customer service agent initiative. Suppose it is expected to generate $500K in net present value (NPV), consumes $100K of compute, and the team's confidence in the projection is 80%. The risk-adjusted NPV is the expected NPV multiplied by that confidence: $500K x 80% = $400K. Dividing by the $100K compute cost gives a risk-adjusted NPV of $4.00 per compute dollar. That single number lets this initiative be ranked directly against any other, however each earns its return.

• Confidence scoring: Rules check whether the financial inputs are complete, while a language model reads the written justification and scores the quality of the reasoning. That score drives a risk multiplier, so an optimistic projection with thin evidence cannot outrank a well-supported one.

Below is a high-level architecture diagram representing the framework:

Governance and Data Security Are the Foundation

A FinOps system contains sensitive information, namely forward-looking financial projections, so data security is essential. I think of security in two parts: who can log in to the system, and what data that user can see or edit once logged in. All users sign in through Single Sign-On (SSO), and an access file determines which organizations a user can see or edit based on their username. All filtering happens on the server before any data is returned to the browser, because hiding figures in a browser is not security; restricting what the system returns is.

Create, update, and delete actions are logged in an append-only history, so every change can be traced back to its original creator and subsequent editors with a full history of changes. For finance functions subject to strict regulation, that audit trail is the minimum required for an organization's leaders to trust the tool.

Lessons for Leaders Adopting AI

Three core lessons apply to any organization putting this framework into practice. First, externalize the parts that change, keeping financial formulas and field definitions in configuration your business team can edit rather than buried in code. Second, make uncertainty explicit, so every projection carries a confidence level that can be weighed. Third, decide where human judgment belongs: deterministic checks in rules, and nuanced assessment with a model and a person reviewing it.

Five questions every AI investment committee should ask:

• What business outcome is this AI initiative expected to produce?

• What will it cost at today's usage, and at 5x or 10x today's usage?

• What assumptions drive the expected return?

• How confident are we in those assumptions?

• What evidence would cause us to increase, reduce, or terminate funding?

Conclusion

GenAI has increased compute spend and made it harder to track. By reframing the problem in terms of portfolio management, organizations gain a shared language for allocating expensive AI dollars. Most importantly, the whole organization shifts from a reactive 'that AI expense went up again' mindset to managing a portfolio of AI initiatives for the best return.

About the Author

Rohit Nagpal is a Senior Data Engineer specializing in AI-powered data platforms and serverless architectures at enterprise scale and recently joined ISSIP as an Ambassador. His work focuses on integrating AI agents into production data pipelines for automation and decision support. He holds a Master of Science in Business Analytics from California State University, East Bay.

FinOps for GenAI: From Unpredictable AI Costs to Managed Investment Decisions