Will AI Token Usage Grow 1,000x? Agents and Budgeting
Box CEO Aaron Levie expects far more agents than people, and token bills that are hard to predict. Here is what that means for AI budgets.
For Finance, Executives and IT

Will AI token usage really grow 1,000x?
Nobody can give you a reliable multiple, and the interview behind this question doesn't offer one either. What Box CEO Aaron Levie did say is that companies are heading toward many more agents than people (he cited Jensen Huang's range of 10 to 100 times more), each processing information all day. He also said you "can't really predict in advance, at least not today" how much you will spend on tokens. The direction is up. The size is unknown.
TL;DR
- Token spend is a usage bill, not a license fee. It rises with every document, event and task an agent touches.
- Levie, citing Jensen Huang, described a future with roughly 10 to 100 times more agents than people. More agents means more tokens.
- Levie said you can't predict token spend in advance today, and joked he could hear the "shivers of a thousand CFOs."
- "1,000x" is a way to frame the scale of the change. No source here backs it as a forecast, so don't put it in a budget.
- The practical answer is unit costs: what one contract review, one invoice or one code check costs, multiplied by volume.
- Workflows change often as models improve, so budgets need a review cycle, not a single annual number.
What is a token, and why do agents use so many?
A token is a small chunk of text, roughly a word or part of a word, that an AI model reads or writes. Most providers charge per token, with different rates for what goes in and what comes out.
A single chat question uses few tokens. An agent is different. An agent is an AI system that takes a goal, breaks it into steps, calls tools, reads documents and checks its own work. Each step uses tokens. A task that looks like "review this contract" might involve reading the whole document, looking up related files, drafting a summary, checking it, and writing results back to a system.
So spend scales with three things: how many agents you run, how often they run, and how much material they read each time. A person asking a chatbot ten questions a day is one thing. An always-on agent that reviews every incoming file is another.
Why would every piece of information touch an agent?
Levie's picture is of organizations where agents roam through company information, doing the processing. In the interview he said something has to hold the "ground truth" for those agents: which data they can reach, whether that is CRM data, ERP data, data lakehouse data or files, and under what rules.
If agents become the default way work gets processed, then the volume of material they read grows with the volume of the business. Some illustrations of where that could show up:
- Go-to-market: an agent reads every inbound lead, call note and contract.
- Cybersecurity: an agent reviews every security event or alert instead of a person sampling a few.
- Engineering: an agent reviews every code push.
- Supply chain: an agent checks every shipping, customs or supplier document.
Those are examples of the pattern, not figures from the interview. Levie's own examples were workflows like contract management, supply chain document review, invoice processing, customer onboarding, loan processing and client intake. The shared feature is that each one has a steady stream of documents or events, and every item in that stream can cost tokens.
Why is token spend so hard to budget?
Traditional software has a seat price. Ten people, ten licenses, a number you can put in a spreadsheet. Token spend has no such anchor. Several things make it slippery:
- Usage drives cost. A busy month costs more than a quiet one, and success makes it worse. If a workflow works well, teams run it on more items.
- Agents vary run to run. One task may take three steps and the next may take thirty, depending on what the agent finds.
- Models change. Levie said his team throws things away "all the time" because a new model can make a piece of scaffolding unnecessary. A new model may be cheaper per task, or smarter and more expensive, or able to take on tasks you never ran before.
- Workflows get redesigned. Levie described Box Architect, a system that uses frontier models to design agent workflows. He said an early internal version built in 30 seconds what customers spend two days designing on a whiteboard. When design gets that cheap, people build more of it, and more workflows mean more spend.
That is the CFO problem in one line. The cost is real, it grows with success, and last quarter's number is a weak guide to next quarter's.
How should a company budget for token spend?
Stop forecasting a single total. Forecast from the bottom up, using the unit cost of each workflow.
- Pick one workflow. For example, invoice processing.
- Measure a pilot. Run 100 real items and record the token cost from your provider's usage report. Divide by 100 to get a cost per item.
- Multiply by expected volume. If you expect 5,000 items a month, you have a first estimate. Add a buffer for retries and growth.
- Compare with the work it replaces. The relevant question is whether the agent costs less than the time and delay it removes, not whether the bill is large.
- Set a cap and an alert. Most providers let you set spending limits. Use them per project, not just per company.
- Review monthly. Model prices and your own usage will both move.
Here is the kind of tracking sheet that makes this visible. Copy it and fill it in per workflow:
Workflow name:
Owner:
Items processed per month:
Average cost per item (from pilot):
Monthly estimate (items x cost):
Monthly cap:
Alert threshold (% of cap):
Which model is used:
Last reviewed:
What it replaces (hours or tools):
Other levers worth using: route simple tasks to smaller, cheaper models and keep the strongest model for hard ones; avoid feeding agents entire document libraries when a short excerpt will do; and limit how many steps an agent may take before it must stop and ask.
What does this mean for you?
If you manage a team or a budget: expect AI to show up as a variable line item, not a fixed one. Ask every proposed agent project for a cost per item and a monthly cap before it launches.
If you work in finance or procurement: you will be asked to approve spend that behaves more like a utility bill than a subscription. Push for usage reports by team and workflow so you can see where the growth comes from.
If you are an individual contributor: the people who can say "this task costs X per run and saves Y hours" will have an easier time getting agents approved. Learn to measure that.
If you are in IT or security: Levie's point about governance applies here too. Someone has to decide which data each agent may read, because every extra source also adds tokens and risk.
Can a small internal tool help you keep track?
Most teams start tracking AI spend in a spreadsheet and outgrow it quickly. With Remy, you describe what you want in plain English and it builds and deploys a full-stack app. A reader in this situation could ask for a token budget tracker: a request form where teams propose an agent workflow with an estimated cost per item, an approval step for the finance owner, and a dashboard showing monthly spend against each cap. You own the app and its code, and Remy bills AI tokens at cost, so the tool itself doesn't add a markup to the thing you're tracking.
Frequently Asked Questions
Did Aaron Levie say token usage will grow 1,000x?
Not in the portion of the interview available here. He said agents could outnumber people by 10 to 100 times (citing Jensen Huang) and that token spend can't be predicted in advance today. The 1,000x figure is a headline-style framing of how large the shift could be, not a number from a source.
Why can't companies predict token spend?
Spend depends on how many tasks agents run, how many steps each takes, which model is used and how much material is read. All of these change as teams adopt more workflows and models improve. Measuring cost per item in a pilot is the most reliable starting point.
Will tokens get cheaper over time?
Possibly per task, but total spend can still rise if you run far more tasks or use stronger models for harder work. Budget for the total, not just the unit price.
What is the first thing to do if my company is starting with agents?
Choose one workflow with a steady stream of items, run a pilot, and record the cost per item. Set a monthly cap and an alert before you scale it to more teams.
How we source, write, and correct our articles. Editorial standards
Related articles

Do You Need the Most Expensive AI Model? Task Saturation

How to Choose the Cheapest AI Model That Passes Your Evals

How Much Does Fable 5.2 Cost? Real Token Bills by Project
