An AI assistant costs a few hundred dollars in its first month. In the third month, the bill is three times higher, but usage has barely changed. Nobody can say why until someone finds that a new document set doubled the text sent with every request, and a retry setting quietly ran some requests twice.

The monthly invoice tells you how much you spent. It does not tell you which task, change, or team caused it. A little monitoring after launch answers that question while it is still cheap to fix.

Measure cost per completed task

Pick a unit the business understands: one answered support question, one processed invoice, one drafted proposal. Then track:

  • Cost per task: total model and tool cost divided by tasks completed.
  • Tasks per day: to separate "more usage" from "more expensive usage."
  • Cost of failed or abandoned tasks: work that was paid for but produced nothing useful.

A rising total with a steady cost per task usually means adoption is growing. A rising cost per task means something changed.

Record what drives the cost

For each request, log the model used, the amount of text sent and returned, the number of tool calls, and whether it was a retry. Keep the identifiers, not the customer content. This is the same timing and step data that helps when you diagnose a slow workflow, so collect both at once.

Common causes of creeping costs:

  • More or longer documents attached to each request.
  • Conversation history that grows with every message.
  • Retries after timeouts that actually succeeded the first time.
  • A switch to a more capable model for every task, not just the hard ones.

Set a budget and an alert

Agree on a normal range for cost per task and a monthly ceiling. Alert the owner when either is exceeded, using the principles from our automation alerts guide. Many providers also let you set a hard spending limit; use it as a safety net, not as the only check.

Reduce cost without lowering quality

Test each change against your quality test set before keeping it:

  • Send only the parts of documents the task needs.
  • Summarise long conversations instead of resending them in full.
  • Route simple tasks to a smaller model. Our guide to choosing a model covers the trade-offs.
  • Cache answers to questions that repeat word for word.

Report value, not just spend

Once a month, put cost per task next to the time or money each task saves. That turns the bill into a business decision. Our ROI guide explains how to set up the comparison.

To set up cost monitoring for an AI workflow you already run, email Storm Nova Labs.