Based on the webinar: How to Scale AI Without Surprising Your CFO
Panelists:
Eric Lam — Head of Cloud FinOps, Google · Chase Platon — Senior Staff Technical Program Manager, Shopify · Roi Ravhon — Co-Founder & CEO, Finout
Scaling AI without surprising your CFO means building financial visibility, cost attribution, and unit economics into AI programs before production costs become unmanageable — not after. It is the organizational discipline that separates AI programs that earn continued investment from those that get shut down by finance.
For most organizations, the first phase of AI adoption was defined by a single question: can we build it? Pilots were funded, experiments were run, and the engineering teams delivered. The infrastructure bills were small enough to absorb, and the CFO was not in the room.
That phase is over. As AI moves from isolated experiments into production workloads — powering customer-facing features, internal tools, automated pipelines, and agent frameworks — the question has fundamentally shifted. It is no longer "can we build it?" It is "what does it actually cost, and where is the ROI?" The CFO is now in the room, and the organizations that scale AI successfully are the ones that anticipated that transition rather than being caught by it.
This is the core theme that emerged from a recent webinar featuring Eric Lam (Head of Cloud FinOps, Google), Chase Platon (Senior Staff Technical Program Manager, Shopify), and Roi Ravhon (Co-Founder and CEO, Finout): AI at scale is an economics problem as much as an engineering problem, and the teams that treat it that way are the ones that earn the organizational trust to keep investing.
The gap between AI pilot costs and production costs is structural, not incremental. Pilots are scoped, controlled, and easy to absorb. Production is none of those things — it is dozens of features running simultaneously, used by real customers, with costs that compound across every request, every model call, and every team that ships something new.
A pilot running on a single use case with a small user base can appear cost-trivial. The same architecture running across a dozen features, processing millions of requests per day, looks completely different on an invoice. What makes this particularly difficult is that the growth is non-linear. AI infrastructure costs do not scale cleanly with user count or feature count. They scale with usage patterns, model choices, and architectural decisions that engineering teams make for performance reasons — unless cost has been made a visible design criterion from the start.
This is the trap most organizations fall into: by the time costs become visible, the architectural decisions that drove them are already embedded in production systems.
Cloud compute costs have native observability. AWS Cost Explorer, GCP Billing, and Azure Cost Management provide resource-level breakdowns with tagging, allocation, and anomaly detection built in. AI spend — spread across multiple models, providers, and use cases — does not come with the same visibility by default. A provider invoice tells you what you spent. It does not tell you which product, team, or decision drove that spend.
This is the visibility gap that makes AI costs surprising. Spend accumulates across dozens of services and features with no shared attribution layer. By the time the invoice arrives, the spend is already incurred — and tracing it back to specific decisions or features is a forensic exercise, not an operational one.
The solution is deliberate cost attribution: ensuring every AI workload is tagged, tracked, and tied to the team and product that generated it. This is not a complex technical problem. It is an organizational one — it requires that engineering teams treat cost attribution as a first-class concern when building AI features, not an afterthought.
The pattern that works: Organizations that avoid CFO surprises enforce a simple rule — no AI feature ships to production without cost attribution. Every team owns its spend. The invoice is never a mystery.
Granular AI cost visibility for business units means giving product managers, engineering leads, and business owners real-time access to the costs their decisions generate — expressed as unit economics tied to outcomes, not abstract infrastructure spend. A team that knows what their feature costs per successful user interaction can reason about that number. A team that receives a monthly infrastructure bill cannot.
The principle mirrors what mature cloud FinOps programs have learned: when teams can see their own spend in real time, they make better decisions. They question whether the most expensive model is actually necessary for a given task. They notice when a new feature is running far above expected cost. They start treating cost efficiency as a design criterion, not a finance team concern.
Getting there requires a cost model that maps AI spend to business outcomes — cost per completed session, cost per resolved ticket, cost per generated output — so that teams are working with unit economics that connect to the value they are delivering.
Making the CFO an AI infrastructure ally means giving them visibility, predictability, and a clear link between AI spend and business outcomes — the same three things they need to support any infrastructure investment. When those conditions are met, the CFO shifts from budget gatekeeper to AI investment advocate.
The framing of the CFO as a stakeholder to manage — or worse, to avoid — is the wrong mental model. The organizations scaling AI most effectively have made the CFO an active participant in AI infrastructure strategy, not a gatekeeper to route around. They need to see that AI costs are instrumented and attributed. They need confidence that anomalies will be caught before they become invoice surprises. And they need a model that connects AI investment to measurable returns — not as a one-time business case, but as an ongoing operational metric.
Teams that build this relationship early earn something valuable: organizational permission to scale. When the CFO trusts that AI costs are managed with rigor, investment decisions move faster. New use cases get funded. The conversation shifts from "how do we control this cost?" to "where should we invest more?"
Scaling AI responsibly means treating cost attribution, unit economics, and financial governance as engineering requirements — not post-launch remediation tasks. The organizations that do this from the start scale furthest, with the most organizational support.
That means cost visibility from day one — not retrofitted after the bills arrive. It means unit economics metrics that connect AI spend to business value. It means cost review as a standard part of the process for new AI features. And it means the FinOps function expanding its scope to cover AI as a first-class cost category alongside cloud compute and Kubernetes.
The organizations represented in this panel — operating AI at the scale of Google and Shopify — have learned these lessons through direct experience. The consistent message is that the teams that build financial discipline into their AI programs from the start are the ones that scale furthest, fastest, and with the most organizational support.
Finout is built for exactly this moment. As AI workloads become a permanent, growing layer of enterprise infrastructure, the need for unified cost visibility — across cloud, Kubernetes, and AI spend — in a single system of record that engineering and finance both trust is no longer optional. It is the foundation that makes responsible AI scale possible.