Why should you care?

Azure OpenAI invoice grows linearly with tenants — finance expects unit economics per copilot interaction.

See it live — copy this example

Paste into an ASP.NET Core 8+ / AgentNest project, then run with dotnet run (set your API keys in user-secrets).

// AgentNest.Cost/ModelRouter.cs
public sealed class ModelRouter(IChatClient mini, IChatClient full)
{
    public IChatClient Pick(string feature, int complexityScore) =>
        feature switch
        {
            "crm.tag" => mini,
            "erp.variance" when complexityScore < 3 => mini,
            _ => full
        };
}

What happened?

• ModelRouter sends low-risk CRM tagging to gpt-4o-mini and reserves full model for complex ERP variance analysis.
• Follow the steps below — typing the code yourself is the fastest way to learn.

Practice next

1. Tag each agent call with feature name for cost allocation.
2. Cache embedding vectors for unchanged KB chunks.
3. Set MaxOutputTokens per endpoint aggressively.
4. Add semantic cache keyed by question embedding similarity.
5. Batch CRM tagging jobs off-peak with reserved capacity.

Remember

Route requests to right-sized models per feature. Cache embeddings and frequent RAG answers. Track cost per tenant from usage tables.

FinOps review

OpenAI bill doubles after CRM launch.

Outcome: ModelRouter and embedding cache cut blended cost per interaction 40%.