The new bill nobody owns
The pattern is familiar. A team ships an AI feature on OpenAI, Anthropic or Bedrock, usage grows, and six months later there's a five-figure monthly line that nobody can map to customers or features. Every request quietly carries maximum context, nothing is cached, and the most expensive model handles work a cheap one could do. It's the AWS bill story again, faster.
What I do
- Spend audit across your AI providers, mapped to features and, where possible, per-customer unit economics
- The cheap wins: caching, prompt-size discipline, batching, and routing simple work to smaller models
- Platform choices — direct APIs versus Bedrock, and what your data-handling promises actually require
- Guardrails: budgets, alerts and per-feature limits so growth can't ambush you
- Governance basics — logging, access and what customer data leaves your boundary
Scope, honestly
I do the platform and the bill: infrastructure, cost and controls. I don't tune your prompts for quality or build your models — you know your product better than I do. Where the fix is "use a smaller model here", I'll show you the numbers and your team makes the call.