18 ways teams waste money on AI
Most AI waste comes from ordinary habits. 18 examples, what each one costs, and the fix for each.
Picture a 20-page competitor report that an AI agent writes every morning. It's thorough. It costs about $24 a day. Nobody has opened one in 63 days.
That's $1,486 for nothing, and nobody did anything wrong. Someone set it up for a good reason, the reason went away, and the job kept running.
Most AI waste looks like this. It doesn't come from one reckless decision. It comes from ordinary habits that nobody notices, because the bill arrives as one total at the end of the month.
We put together 18 examples like it. They're illustrative, not drawn from any one company, but each is a pattern that's easy to fall into. Together they cost $14,076. The work they did should have cost about $1,560.
About 89% of the spend in these examples bought nothing.
Almost every fix is a habit or a setting. The examples fall into 5 groups.
1. Work that never stopped
AI tools don't get tired, and they don't notice when they're going in circles.
The agent that never stopped. A background research agent kept searching and calling tools for 19 hours. It spent $2,714 on a job worth about $94. Every agent that runs on its own needs a limit on steps, time, and spend.
127 calls to the same tool. A connector got stuck searching for the same information and the agent kept asking. $624 spent, $18 needed. When a tool fails the same way twice, stop and change the approach.
Kept trying to save a broken session. The agent was clearly confused. The user kept typing "try again" for two more hours. $511 spent, $42 needed. A confused session rarely recovers. Start a clean one with the right context.
The daily report nobody reads. The one from the top. Scheduled work keeps going long after anyone needs it. Check what's running on a schedule every so often, and turn off what nobody uses.
2. Too much context
Every file, log, and old message in a session gets sent to the model again with each request. You pay for all of it, every time.
One coding session, three weeks. The same Claude Code thread was used for features, bugs, and pull requests for three weeks. By the end, every small question carried weeks of unrelated history. $1,842 spent, $286 needed. Start a fresh session when the task changes, and compact long ones.
Loaded the whole repo to change one button. An agent read hundreds of files to change "Save" to "Save changes." $186 spent, about a dollar needed. Point it at the component.
50,000 lines of logs. Entire production logs were pasted in again and again to debug one error. Send the error and the lines around it.
Searched everything for one known document. Every question about a policy searched thousands of files, when the answer was always in the same handbook. Point the tool at the handbook.
3. The same work, done more than once
Four agents, one bug. Four agents each read the whole repo and investigated the same issue on their own. $842 spent, $176 needed. Start with one agent. Run several only when the work actually splits into separate pieces.
Rebuilt something that already existed. An agent spent hours building a component that was already in the design system. Ask it to check what exists before it builds.
The summary of the summary of the summary. A meeting transcript became a summary, then an exec summary, then a Slack summary, then part of a weekly summary. Each step paid to reread the last one. Write the final format directly from the source.
The same task in four AI tools. One research request went into four different assistants. Comparing answers can be worth it on purpose. Out of habit, it's paying four times for one answer.
4. Paying for more than the job needs
Used the best model for everything. A premium model handled formatting, tiny edits, classification, and basic questions for a month. $540 spent, $112 needed. Keep the top model for work that needs it and route the simple work to smaller ones. We wrote more about this in how to pick the right AI model for each task.
Paid extra for speed nobody needed. The fastest, priciest processing option was used for overnight and background jobs. $1,240 spent, $490 needed. Use standard or lower-cost processing when nobody is waiting on the result.
Ran it on 50,000 rows first. A classification job went straight to the full customer dataset without a test run. If the prompt needs work, you find out after paying for all 50,000. $2,318 spent, $185 needed. Test on a small sample, tune it, then scale.
5. Not work at all
The last three are smaller, and a little funnier. Someone planned a two-week trip to Italy on the company account: hotels, trains, restaurants, packing lists, and a lot of itinerary revisions. Someone else spent three days comparing TVs. A third used it to write dating messages, including more than one request for "20 more."
Together that's $272. Nobody's budget breaks over it, and most people don't think twice, because the company tool is the one that's open.
The fix is a sentence in your AI policy saying what the company account is for, and an easy personal option if you'd rather people use one.
Why nobody notices
Look at the list again. Almost none of it would show up as a problem on an invoice. It shows up as "AI spend went up again this month," with no way to tell which part was a feature that shipped and which part was a report nobody read.
Three things make the difference:
- Limits that hold. Agents and scheduled jobs need step, time, and spend limits. People need budgets that apply in the tools themselves. With Tokenize budgets, each person gets one monthly limit that covers Claude, ChatGPT, and Cursor.
- Seeing spend while it happens. You want to hear about a runaway agent the same day it starts. Tokenize shows spend for every API key and agent and sends an alert when usage jumps.
- Better habits, without a lecture. Most of the fixes above are habits: starting fresh, pointing the tool at the right file, picking a smaller model. Tokenize sends each person private tips based on how they use AI, like picking a better model for the task or giving clearer instructions up front.
All 18, in one table
| Example | Spent | Needed | Wasted | Fix |
|---|---|---|---|---|
| The agent that never stopped | $2,714 | $94 | $2,620 | Add step, time, and spend limits |
| Ran it on 50,000 rows first | $2,318 | $185 | $2,133 | Test a small sample, tune it, then scale |
| One coding session, three weeks | $1,842 | $286 | $1,556 | Start fresh when the task changes and compact long sessions |
| Daily report nobody reads | $1,486 | $0 | $1,486 | Turn it off |
| Paid extra for speed nobody needed | $1,240 | $490 | $750 | Use standard processing when nobody is waiting |
| Four agents, one bug | $842 | $176 | $666 | Start with one agent; split only when the work does |
| 127 calls to the same tool | $624 | $18 | $606 | Stop repeated failures and change the approach |
| Kept trying to save a broken session | $511 | $42 | $469 | Start a clean session with the right context |
| Used the best model for everything | $540 | $112 | $428 | Route simple work to smaller models |
| Searched everything for one known document | $438 | $38 | $400 | Query the known handbook directly |
| Rebuilt something that already existed | $364 | $24 | $340 | Check existing components first |
| 50,000 lines of logs | $287 | $9 | $278 | Send the error and the lines around it |
| Loaded the whole repo to change one button | $186 | $1.12 | $184.88 | Point it at the component |
| The same task in four AI tools | $218 | $54 | $164 | Use one tool unless comparing answers has a purpose |
| The summary of the summary of the summary | $194 | $31 | $163 | Write the final format directly from the source |
| Planned a two-week Italy vacation | $143 | $0 | $143 | Use a personal account |
| Researched a new TV for three days | $91 | $0 | $91 | Use a personal account |
| Used work AI for dating messages | $38 | $0 | $38 | Use a personal account |
| Total | $14,076 | $1,560.12 | $12,515.88 |
These examples are illustrative. Costs are for each example as described and will vary with the model, tool, and pricing plan.
None of these take long to fix once you can see them. If you'd like to find out which ones are happening on your teams, book a demo and we'll look at your own spend with you.