Key takeaways
- Seat bills rarely jump on their own. A jump almost always comes from usage-based spend.
- Break usage down by day, model, project and user. The cause is usually obvious once you do.
- The most common cause is an expensive model doing cheap work.
- The fix that stops it recurring is a hard cap and a separate budget for each automation.
An AI bill that doubles is usually not a mystery. It feels like one because nobody was watching the usage-based part, and the person who changed something did not connect it to cost. This is a diagnostic: the seven usual causes, where the evidence shows up, and what to do about each.
Step one: break the spend down
Open your provider's usage or billing dashboard and view spend by:
- Day, to find when it started.
- Model, to see if an expensive model is doing most of the work.
- Project or workspace, to find which automation or team.
- User or API key, to find who.
Then ask the simplest question: what changed on or just before the day the line went up?
The seven usual causes
1. A premium model doing routine work
Evidence: most spend sits on the most expensive model, on tasks like classification, extraction or short replies.
Fix: move routine tasks to a smaller, cheaper model. See matching AI models to jobs.
2. An automation running more often than anyone thought
Evidence: a steady, high number of requests from one project or key, often around the clock.
Fix: check the trigger. "Every new email" or "every record update" can mean thousands of runs a day. Batch, filter or schedule it.
3. A loop or retry storm
Evidence: a sharp spike over hours or a weekend, from one source.
Fix: stop it, then add step and retry limits to the agent and a per-project cap.
4. Too much context on every request
Evidence: input token counts far higher than output, with large average request sizes.
Fix: send the relevant extract rather than the whole document or mailbox, and use caching where the provider offers it.
5. A new agent nobody costed
Evidence: a new project, environment or key that appeared around the increase.
Fix: give it an owner, a budget and a review date. See AI agent spending limits.
6. Seats nobody uses
Evidence: seat count higher than weekly active users, often by a third or more.
Fix: reclaim seats with no activity in 60 days, and buy new seats on request rather than in bulk.
7. Duplicate tools
Evidence: several AI subscriptions on expense claims or cards, often for similar purposes.
Fix: consolidate onto your governed provider. See shadow AI in the workplace.
Quick reference
| Pattern in the data | Likely cause |
|---|---|
| Gradual climb, one expensive model | Premium model on routine work |
| Flat but high, around the clock | Automation trigger too broad |
| Sudden spike, then flat | Loop or retry storm |
| High input, low output | Too much context |
| New line from a new project | Uncosted agent |
| Seat line up, usage flat | Unused seats |
Stop it happening again
- Set a hard organisation cap. See hard AI spending limits.
- Put every automation and agent in its own project with its own cap.
- Name an owner for each.
- Review usage monthly for the next quarter, then quarterly.
Frequently asked questions
Why is our AI usage bill so high?
The most common causes are a premium model being used for routine work, an automation or agent running far more often than expected, large documents being sent with every request, a loop or retry problem, and unused seats still being billed. Usage reports broken down by project, model and day usually show which one it is.
How do I find what is driving AI costs?
Open the provider's usage dashboard and break spend down by day, model, project or workspace, and user. Look for the day the increase started and the project or model where it is concentrated. Then ask who changed what around that date.
How can we reduce AI costs without reducing usage?
Map routine tasks to cheaper models, trim the context sent with each request, reclaim unused seats, and consolidate duplicate subscriptions. These usually cut spend noticeably without anyone losing capability.
How do we stop it happening again?
Set a hard organisation cap, give each automation or agent its own project with its own budget, name an owner for each, and review usage monthly for the first quarter.
Get a straight answer for your business
The Clarity Package gives you a written business case and a detailed implementation plan for 500 USD. The Implementation Package does the whole build, live in 30 days, for a fixed 5,000 USD.