Case Study: AI FinOps Program for Enterprise GenAI Spend

Insurance

AI FinOps Program for Enterprise GenAI Spend

The Challenge: Fast-Growing AI Spend That Nobody Could Explain

An insurer had moved quickly on generative AI and had around forty use cases in production, spread across claims, underwriting, customer service and internal operations. Spend on models and supporting infrastructure had grown ninefold in twelve months. Finance received one consolidated invoice from each provider and could not see which use cases or business lines were behind the cost, or whether any of it was paying for itself. After several months of missed forecasts, the CFO put new AI initiatives on hold until the organization could show where the money was going.

Our Solution: Attribute the Spend, Then Reduce It

Lydatum set up an AI FinOps practice with the insurer's finance and technology teams, built in four parts:

• Cost Visibility & Attribution: All model traffic was routed through a central gateway that tagged every request by use case and business unit, and dashboards showed spend next to volume and outcomes.
• Workload Optimization: Each task was matched to the least expensive model that met its quality bar. Prompt caching and batch processing were added for work that was not time-sensitive, and every change was checked against an evaluation suite.
• Unit Economics: We calculated cost per claim and per customer interaction for every use case and compared it with the value delivered.
• Budgets & Governance: Budgets and alerts were set by business unit, and new use cases now go through a short cost review before approval. A forecast links expected spend to business volumes.

The unit economics review led the insurer to retire six use cases that cost more than they returned and to redesign four others. Finance and engineering review the same figures each month, and an internal team we trained during the engagement now runs the practice.

The Impact: Lower Spend and a Forecast Finance Trusts

Within five months the insurer had a cost base it could explain and forecast.

Monthly spend fell by 47% while transaction volumes kept growing, and forecasts came within five percent of actuals, which was enough for the CFO to lift the hold on new initiatives. Each new use case now comes forward with a cost model and a target cost per transaction.

Services Delivered: AI Cost Visibility & Attribution, Model Routing & Cost Optimization, AI Unit Economics & ROI, AI Budget Governance & Forecasting

Technologies Used: Amazon Bedrock, Azure OpenAI, LiteLLM, OpenTelemetry, Snowflake, Power BI, Python