On July 1, Gartner put a number on a shift a lot of buyers already sensed was coming. In a press release on agentic arbitrage, it projected that up to $234 billion of enterprise application spending is exposed to agents completing work across multiple systems, roughly 20 percent of application SaaS spend by 2030. The framing is that agents deliver outcomes directly and make the interface, the seat license, and the dashboard less relevant. Gartner calls the result a metamorphosis of the SaaS market rather than an apocalypse.
The number will get quoted in a lot of board decks this quarter. Before it shapes your roadmap, it is worth reading what the projection assumes, because the assumptions are where the work actually lives.
The arbitrage is an integration problem wearing a strategy costume
Strip the market language off and the mechanism is simple. Agentic arbitrage happens when an agent can reliably do a task that used to require a person clicking through two or three applications. The value does not come from the model being clever. It comes from the agent being able to reach each of those systems, read the right state, take an action, and confirm the result, without a human babysitting each hop.
That is an integration and access problem, and it is the part that does not fit on a slide. An agent that drafts a renewal in your CRM, checks entitlement in your billing platform, and updates the support record is only useful if it can authenticate to all three, if the data it reads is current, and if a failure in step two does not leave step one half-committed. None of that is a model capability. It is the same systems-integration discipline that has always separated software that works from software that demos.
Gartner's own language points at this. The release notes that the vendors already delivering agentic outcomes typically require heavy services engagement, and that better outcomes depend on systems that retain deep institutional memory and customer context over time. Translated: the agent has to know your business, and it has to be wired into your systems well enough to act on that knowledge. Neither of those ships in a box.
Retained context is the harder half
The institutional-memory point deserves more weight than it usually gets. An agent that forgets what happened last week is a chatbot with API access. The systems that create arbitrage are the ones that carry context forward: what this customer already asked for, which exceptions your team has approved before, what your actual approval chain is.
Holding that context is a real engineering constraint, not a prompt you write once. Anthropic's write-up on context engineering for agents is candid that a model's attention is a finite budget and that recall degrades as you stuff more into the window. Their answer for long-horizon work is not a bigger window. It is deliberate machinery: compaction, note-taking that persists outside the context window, and retrieval that pulls the right record in at the moment it is needed. The buyer-facing lesson is that retained context is architecture. If a vendor's pitch treats it as a feature that is already handled, that is the claim to probe first.
Legacy estates are the least ready, and the most exposed
Here is the uncomfortable part for the organizations Gartner is warning. The estates most exposed to arbitrage are often the least able to capture it, because arbitrage rewards clean, reachable systems and legacy estates are neither.
An agent cannot arbitrage across an application that has no real API, whose data model lives in one developer's head, or whose authentication story is a shared service account nobody wants to touch. A twenty-year-old ColdFusion or line-of-business app can absolutely be part of an agentic workflow, but only after someone gives it a reachable surface, a clear contract for reading and writing state, and access controls that let an agent act without becoming your biggest security hole. That work is remediation and modernization, and it is the precondition, not an optional follow-on.
This is why the honest version of the Gartner story is not "buy an agent platform." It is "the platform is the easy part." The competitive gap between the companies that capture this spend and the ones that watch it move will be readiness: whose systems can an agent actually operate, and whose institutional context can it actually hold.
What to do with the number
If the $234 billion figure prompts one useful action, make it an inventory rather than a purchase. Three questions separate a real opportunity from a slide:
First, which of your high-value workflows already cross two or more systems that a person clicks through today? Those are your arbitrage candidates. Ignore the ones that live entirely inside one application; there is no arbitrage there yet.
Second, for each candidate, can an agent reach every system in the chain through a real interface with scoped, auditable access? If the answer involves screen-scraping or a god-mode service account, you have integration and security work to do before any agent is safe to point at it.
Third, where does the context that makes the workflow correct actually live, and can it be retrieved reliably? If the answer is "in the heads of two senior people," the agent will be confidently wrong until that knowledge is captured somewhere a system can read.
Answering those candidly is the substance of an AI readiness assessment: it names the binding constraint before you spend real money finding it the hard way. The workflows that clear all three questions are where an automation effort will actually return something. The ones that do not are a modernization backlog with an AI business case finally attached to it, which is a better reason to fund the work than most estates have had in years.
The metamorphosis Gartner describes is coming. It will reward the organizations that treated it as an implementation problem with a strategy attached, and it will pass by the ones that bought the strategy and assumed the implementation would follow. The number is a forecast. Whether it lands in your favor is an engineering decision you get to make now.