Fast answer
AI costs run away through five holes: automations calling premium models for tasks a cheap model handles, verbose prompts and bloated context sent on every request, retry loops and error storms multiplying calls silently, per-seat licenses (Copilot) assigned broadly but used narrowly, and zero budget alerts on pay-per-use services. The fix is the same discipline as cloud cost governance: model-tier matching, spend caps and alerts, usage audits, and someone accountable for reading them.
The five holes in the bucket
- Premium models for commodity tasks: classifying an email doesn't need the most expensive model — tier-matching each workflow to the cheapest model that passes quality testing routinely cuts API spend by half or more.
- Context bloat: automations that resend entire document histories with every request pay for the same tokens repeatedly. Prompt and context hygiene is real money at volume.
- Loops and error storms: a failed call that retries aggressively, or two automations triggering each other, burns spend at machine speed. This is the classic weekend-bill-shock scenario.
- Shelf-ware seats: Copilot licensed for everyone, used by a fraction — the same zombie-license pattern as any SaaS, at a premium price point.
- No alerts: pay-per-use AI with no budget caps or anomaly alerts means the invoice is your monitoring system. It's the most expensive monitoring system available.
Governance that takes an afternoon to set up
None of this requires an AI economist. Budget caps and anomaly alerts on every AI service account. A model-tier policy: which workflows may use which models, tested for quality at the cheaper tier first. Rate limits on every automation so a loop hits a wall instead of your credit card. And tagging, so every dollar of AI spend traces to a workflow with an owner.
If this sounds like the checklist from our cloud cost guide, that's because it is the same discipline — AI spend is cloud spend with faster failure modes.
The monthly review: cost per outcome, not cost per token
The number that matters isn't the API bill — it's cost per outcome: what does one processed invoice, one triaged ticket, one generated report cost, and what did the manual version cost? Reviewed monthly, this makes AI spend defensible line by line, kills automations whose economics stopped working, and surfaces the loops and drift while they're cheap.
This review is a standing section of the monthly report in our managed AI service — the same way backup verification is a standing task in managed IT. Boring on purpose: the goal is an AI line item as predictable as your rent.