Seven places your AI and cloud bill quietly leaks money

Nobody sets out to waste money on cloud and AI. It happens one sensible decision at a time. A warehouse sized up for a backfill. A test environment that never sleeps. A seat for someone who left in spring.

None of these show up as a line called "waste" on an invoice. Here's where they hide, and how to find them in your own bills.

1. Warehouses that wait for work

Snowflake bills a running warehouse whether it's busy or not. A warehouse that wakes for a query every ten minutes and stays up for ten more is idle most of the day, and you pay for all of it.

How to spot it: compare the credits a warehouse metered with the credits its queries used. The gap is idle time.

How to fix it: lower the auto suspend setting, usually to 60 seconds for batch work. Dashboards may want a few minutes so their cache stays warm. Our guide to idle warehouses goes deeper.

2. Data nobody reads anymore

Tables, datasets and buckets keep costing money long after the project that made them ended.

How to spot it: look for storage that is billed every day but hasn't been queried in months.

How to fix it: ask the owning team if anything still depends on it. Export a final copy to cold storage if it must be kept, then delete the original.

3. AI work on a bigger model than it needs

The biggest models are often used for jobs a smaller one does just as well: sorting tickets, pulling fields out of text, short summaries.

How to spot it: sort AI spend by model. If a top tier model carries a large or fast growing share, check what it's being used for.

How to fix it: try the smaller model on a sample of real requests and compare the results. Switching routine work can cut that model's spend by half or more. Caching repeated prompts helps too.

4. Seats nobody uses

Copilot, ChatGPT and Claude seats are easy to hand out and easy to forget.

How to spot it: list seats with no activity in the last 30 days. GitHub reports this for Copilot.

How to fix it: remove the seat, or move it to someone who asked for one. Check with the person first. Some people use a tool in bursts.

5. Test and development that never sleeps

Development environments often run around the clock, even though people work a third of that.

How to spot it: tag spend by environment and look at the share that isn't production.

How to fix it: schedule test resources to stop at night and on weekends, and make the smallest size the default.

6. Steady spend at on demand prices

On demand pricing is the most expensive way to pay for something you use every single day.

How to spot it: find services that run almost every day with little change from one day to the next.

How to fix it: price a commitment or reservation sized to the lowest steady level, not the peaks. Keep on demand pricing for anything above it.

7. A spike nobody explained

A runaway job, a retry loop or a test left on can double a service's bill in a day.

How to spot it: compare each day with a normal day for the same service. A jump that doesn't match growth elsewhere is worth a question.

How to fix it: find the change behind it, then set an alert, so next time you hear about it the same day instead of at month end.

Letting RootSpend do the looking

RootSpend runs these checks on your connected tools. Quick checks look for spikes, idle warehouses, unused data, idle Copilot seats, oversized models and development spend. The AI analyst looks across tools for the rest, like steady spend that a commitment would make cheaper.

Every finding comes with the monthly saving, the steps to fix it, and how risky it is. RootSpend never makes the change itself. And 14 days after you mark a fix done, it checks your real spend to confirm the saving.