The Gartner has put a number on a change that many companies already feel without knowing how to name it: up to $234 billion in corporate software spending is exposed to what it calls agentic arbitrage by 2030, about 20% of all software-as-a-service (SaaS) spending during this period. In plain language: one-fifth of the money companies pay today in system subscriptions may change direction, because AI has started delivering the result of the task without anyone needing to open the system.
This post translates what this means for those who sign up for software (i.e., almost every company), shows why the per-user charging model is failing, and tells a real case of Steply where this switch has already happened: a financial company that stopped paying people to operate screens and started receiving the result ready.
What is agentic arbitrage, without economese
First, the vocabulary. SaaS is software that you don't buy, you sign up for: you pay a monthly fee, usually per user, to access a system over the internet. CRM, billing system, HR tool, almost everything works like this today. The whole model is based on a premise: work is a person in front of a screen. More work, more people, more licenses, more revenue for the provider.
Agentic AI breaks this premise. An AI agent is a program that performs the task from start to finish: reads the data, crosses, decides what is routine and what needs people, and delivers the result. It doesn't need the nice screen, the menus, the dashboard. George Brocklehurst, vice president of Gartner, summed it up like this: agentic systems deliver results directly, bypassing user-centric applications. Translation: the agent doesn't open the report, it delivers the report.
And arbitrage is the name economists give to something simple: when there is a way to get the same result by paying less, the money migrates there. That's what Gartner is saying will happen with one-fifth of companies' software budget.
Why per-user charging is failing
Think of a company that pays for 40 licenses of a system. How many of those 40 people use the system to think, and how many use it to feed: enter data, check lines, copy information from one place to another? In most operations we attend, the second group is larger. And it's exactly their work that an AI agent does alone.
When this happens, the software provider's bill disconnects from the value delivered. The company no longer needs 40 licenses, it needs 10 for those who decide and an agent that does the rest. SaaS revenue growth has always depended on the growth of the number of users. Agentic arbitrage cuts this thread. It's not that the system got bad; it's that the customer discovered they were paying for access when what they always wanted was the result.
The corporate buyer, says Gartner, will stop asking "what tool or dashboard do I buy" and start asking "what result do I guarantee". This change of question seems subtle. It reorganizes the entire technology budget.
A real case: the financial company that stopped buying screens
This is not a forecast, it's something we've already delivered. An operation came to Steply with a classic financial bottleneck: beating what came in with what was expected. Every week, someone exported four sources (bank statement, card report, sales spreadsheet, and billing system) and crossed line by line. At the end of the month, this consumed two to three days of a person, and the owner only knew the real cash weeks after the month changed.
The traditional path for this problem would be to sign up for another SaaS: a reconciliation system, paid per user, with another screen for the team to learn to operate. In other words, solving the excess of manual work by buying more places to work manually.
What we built was the opposite: an agent that reads the four sources alone, understands that "Pix received R$ 1,240" in the bank is the same event as "order 8842 paid" in the sales system, automatically reconciles them, and delivers only the list of what doesn't match. No one operates any screen. The financial company receives the five or six discrepancies that really need human judgment, and the rest is already reconciled. When the month changes, the closing is 95% ready, because it was done gradually, every day, in the background. The difference that used to take an afternoon to find is pointed out on the same day.
Notice what the company bought. It wasn't a tool, it wasn't a license per head, it wasn't a dashboard. It was a result: closing that closes by itself, with exceptions pointed out for review. This is agentic arbitrage in practice, before it becomes a Gartner statistic. We detail this delivery in another blog post.
Metamorphosis, not apocalypse: SaaS doesn't die, it changes shape
Here, honesty is valuable, because Gartner itself makes this reservation. Brocklehurst describes the change as less apocalypse and more metamorphosis: SaaS will not be destroyed, it will emerge in a new form. We agree, and our practice shows where this new form is.
What remains valuable is what Gartner calls institutional memory and customer context: systems that store the operation's history, business rules, what went wrong and why. This accumulated knowledge is exactly what makes an AI agent good, because an agent without context makes mistakes. In the case of reconciliation, each discrepancy that a human resolves returns to the system as an example, and today's strange pattern is recognized alone tomorrow. The value migrated from the screen to memory.
What loses value is the generic screen charged per head: the system whose main function is to be the place where a human types what an agent would type alone. And there is a clear limit on the other side: not every task should become an agent. Where there is judgment, negotiation, and money decision, people in front of the screen continue to be the right answer. Our reconciliation agent doesn't approve payments or move a real: above a certain value or with low confidence, it stops and calls the human.
What this changes in your next contract renewal
If your company signs up for software (and it does), the practical reading of Gartner's report is three questions before renewing any contract. First: am I paying for a result or for access? If the value is in what comes out of the system, and not in how many people enter it, the per-user model is charging you wrong. Second: how many of these licenses exist just to feed the system with data? These are the first ones an agent eliminates. Third: if the mechanical part is automated, does this contract still make sense in this size?
The question for the technology budget of the next few years is not "which systems to sign up for". It's "which results to buy, and how much does each one cost". Gartner's $234 billion will come out of the pocket of those who continue to pay per-head licenses for work that can already be delivered ready, and will go to those who do this calculation first. Which side of the arbitrage will your company be on?
