How do you reconcile usage across OpenAI, AWS Bedrock, and Vertex AI?

If you run models across OpenAI, Claude on AWS Bedrock, and Gemini on Vertex AI, how do you reconcile your own usage logs with each provider’s billing data? I’m curious what teams do in practice. Do you track requests and tokens by model in your own logs, or start from provider billing exports? When the two disagree, how do you tell whether it’s a reporting delay, a retry, or a difference in how usage is counted? If you’ve built a reporting workflow for this, what would you track from day one?