Ten times cheaper per session — how four governance decisions brought AI costs under control

The challenge

Three questions without answers

The company was launching successive AI-based features and planning to sell them to its clients. It couldn't provide a clear answer to three basic questions. How much does it cost — individual teams were using different tools and different accounts. There was no single budget line covering everything, nor a way to link costs to specific applications. Who has access to what — in the internal tool being deployed, only administrators had access to the knowledge base and features. Regular users saw the interface but couldn't use it, which nobody knew until team testing. What happens to the data — the question became critical the moment features were meant to cover client data, not just the company's own.

What we did

Four decisions instead of an elaborate policy

• Consolidated model access — one company account with a shared pool of funds instead of scattered individual subscriptions, enabling ongoing spending control and cross-application cost comparison.
• Matched the model to the task instead of using one for everything — switching the model in correspondence processing cut the session cost roughly tenfold while preserving quality.
• Established a data processing policy: no data retention and EU-based processing for models handling client data. For the most sensitive applications, a variant running in the client's own infrastructure was provided.
• Sorted out permissions — created a user group with explicitly defined access scope to the model, knowledge base, and features, instead of leaving full access only to administrators.
• Introduced measurement before selling — each client-facing feature is first launched in test mode with a single recipient to measure actual consumption and choose the billing model.
• Approved the AI budget as a single line item with a six-month horizon.

The result

What changed

The cost of using AI is known, assigned to applications, and subject to ongoing control. Data processing rules are formulated in writing and ready to present to a client or auditor. The team has role-appropriate access without needing admin privileges. Features sold to clients have their profitability calculated before entering the offering.