One page · from running AI in production

The AI Spend Checklist

On most clouds, the infrastructure bill is not the AI bill. The money goes to tokens, and most of it can be saved without changing what customers see. Tick these off in order.

1Know where the money goes

  • Tag every AI call by feature, team and customer, so spend has an owner.
  • Measure cost per task (per summary, per conversation, per document), not per month.
  • Compare price per finished task, not price per token. A cheaper model that needs two tries is not cheaper.

2Cache the repeated part

  • Keep the front of every prompt fixed: instructions, policies, examples.
  • Put per-request details at the end: dates, IDs, customer data. One timestamp at the top breaks the cache.
  • Providers that support caching discount the repeated part heavily (up to about 90%). Check your cache-hit rate weekly.

3Route by task, not by vendor

  • Small, fast models for classifying, tagging, extracting and routing.
  • Top-tier models only for writing, reasoning and anything a customer reads.
  • One routing layer in front of many providers, so changing a model is a setting, not a project.
  • Send work that can wait to batch APIs, typically about half price.

4Put a ceiling on every call

  • Daily caps per user, team and customer, plus a budget alarm before the bill arrives.
  • Max tokens and retry limits on every call. A runaway loop can cost more in one day than a normal month.
  • A kill switch someone can flip without a deploy.

5Plan for capacity and failure

  • A second provider for every critical path. Capacity limits usually show up as rate limits, not outages.
  • The same model is now sold on more than one cloud, so you can shop for price and capacity.
  • Test failover with a drill, not a status page. A switch is only as good as what moves when you flip it.

6Govern what the AI can touch

  • Never log raw requests; they carry keys and customer data.
  • Only run a tool call the model was offered for that request. It stops most prompt-injection damage.
  • Set retention and no-training terms with every provider, in writing.

7Contact center: cost per interaction

  • After-call summaries, QA scoring and agent assist are where per-conversation costs pile up. Measure each one per interaction.
  • Real-time paths (agent assist, routing) on small, fast models; after-call work on batch, at about half price.
  • Same fixed instructions on every call means caching pays on every interaction.
  • A person reviews anything a customer will read until the quality numbers say otherwise.

Want a second pair of eyes? I'll do a 30-minute review of where your AI and contact-center spend is going, and what I'd change first. Vendor-agnostic: I work for the buyer.

Book a 30-minute review

Rob Enslow · AI & CX Advisory · rob@cx.robenslow.com · cx.robenslow.com