Agentic AI enterprise applications will lose up to $234 billion of software spending to agents by 2030, according to a Gartner forecast published on July 1. The research firm calls the mechanism “agentic arbitrage.” An agent finishes a task across several systems, so the user never opens them. By 2030 that shift will touch about 20 percent of enterprise SaaS spending.
New here? I publish one agentic AI governance post every weekday. Subscribe to the blog and it lands in your inbox the moment it goes live.
The forecast lands on a market that has already started to move. For this post I tracked 25 of the largest companies behind agentic AI enterprise applications through their 2026 announcements. I left the names out on purpose, because the pattern matters more than the logos. Every one of them now ships agents, at least in a demo. The difference is where the agent lives, and who owns the record it writes to.
Agents inside the app, one system at a time
In 2026 the agent sits inside the application that sold it. The largest CRM vendor, for example, runs agents on its own customer data. The two ERP incumbents run agents on the ledger. A cloud HR vendor now markets a registry for agents the way it once marketed a registry for people. Each agent works one system at a time, so the user still opens five screens to close a quarter.
But pricing is the first thing to crack. A customer-service vendor now bills per resolved ticket instead of per seat. A code-hosting platform moved to token-based billing, and reporting on the Gartner forecast says the HR vendor changed its pricing the same year. One analyst put it in one line. The shift “breaks the link between user growth and revenue growth.”
The harness moves across systems
The second phase starts when the agent leaves the application. A harness is the software around the model: the loop, the tools, the permissions and the logs. In 2027 and 2028 the harness runs outside any one app and calls the CRM, the ERP and the HR system as tools. The user never sees those screens, because the harness did the work.
The data platforms want to be that harness. One of the two leading data platforms said in June that customers had built more than 100,000 agents on it since launch. The same announcement also puts models, tools, policies and agent traces in one catalog, with budgets per user and per group. Its argument is that governance should not sit “in a different vendor.” The other leading data platform makes the same claim with different product names.
The standards, for their part, settled early. A frontier lab published an open protocol for connecting agents to tools in November 2024. Within a year the data platforms, the app vendors and the harness companies had all adopted it. Once every system speaks the same tool protocol, the application stops being a destination. It becomes an entry in a catalog, and agentic AI enterprise applications start to look like plumbing.
More on agentic AI enterprise applications
- The Six Layers: My pillar page on agentic AI architecture names enforcement as the layer most teams skip.
- Govern the Harness: Why I govern the harness, not the model, across the Five-Stage Roadmap.
- The 2027 Plan: A quarter-by-quarter enterprise agentic AI roadmap for 2027 with a lender as the worked example.
- The Harness Itself: What an AI model harness is and why nobody drives an engine.
The record outlives the screen
By 2029 the question for every application vendor is the same. What do you own once the user stops looking at you? For the survivors the answer is the record. In short, the ledger of what happened, the memory of why, and the controls that decide what an agent may do next.
That is why the systems of record in regulated work hold their ground. The healthcare records vendor keeps the chart, because a clinical control gates every agent action. A loan-origination platform keeps the file, because an examiner will ask for it. So does the ERP with its ledger, because the auditor signs it. Gartner says the same thing in vendor language: “Better outcomes from AI require systems that can retain deep institutional memory and customer context over time.”
The dashboard vendors, on the other hand, do not hold. A vendor whose value was a screen full of charts loses the moment an agent answers the question in a chat window. Gartner calls this “an existential threat for vendors who are defending legacy dashboards and seat-based models.” I think that is the polite version.
Beyond 2030, the loop is the product
Past 2030, then, the enterprise stops buying applications and starts buying a governed loop. Intent goes in. Outcomes come out. Evidence stays behind for whoever has to check. The application layer, the data layer and the harness then collapse into two or three platforms per enterprise. The fight is over which two or three.
Here is the table I keep for agentic AI enterprise applications, phase by phase.
| Phase | Where the agent lives | Who gains | Who loses |
|---|---|---|---|
| 2026 | Inside each application | Incumbents with distribution | Point tools with thin data |
| 2027–2028 | In a harness across systems | Data platforms, frontier labs, harness vendors | Screen-heavy SaaS, seat licensing |
| 2029–2030 | On the record, under controls | Ledgers, charts, loan files, ontologies | Dashboard vendors |
| Beyond 2030 | In a governed loop | Whoever holds data, harness and enforcement together | Anyone selling a screen |
The frontier labs, of course, sell to every row of that table. The CRM vendor, the data platform and the harness startup all run the same three or four models. So the model is not the moat, and none of the 25 pretends it is any more. The moat is the data, the controls and the record.
What I do with this
In my work in regulated loan origination the record was never the screen. Instead it was the file an examiner opened a year later. So when I design for 2030, I put the harness on the data platform. Every application becomes a tool with a permission scope. The evidence stays where an auditor can read it without a vendor login.
My flat opinion: the real fight over agentic AI enterprise applications is not CRM versus IT workflow for the title of enterprise agent operating system. It is the data platform versus the application for ownership of the harness. The application vendors know it, which is why one ERP incumbent now ships its business data on a rival data platform’s engine. Instructions in, results out was IT. Intent in, outcomes out is agentic AI, and the screen was only ever where the instructions went in.
Which of your enterprise applications would survive a year in which nobody opened its screen?
Was this post useful? One click tells me what to write next.
Subscribe to my LinkedIn Newsletter
- Agentic AI Architect: control design for enterprise agents.
- Agentic AI Case Studies: deployment evidence, one teardown at a time.
- Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
- Agentic AI SDLC: the lifecycle that builds the agents.
- Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.