The next fight in enterprise software is no longer just about better chatbots. Microsoft, OpenAI, and Anthropic are pushing AI agents that can work across business apps, while the old guard – Salesforce, ServiceNow, Snowflake, and others – is scrambling to build the control layer that decides who gets to run them, monitor them, and profit from them.
That’s the real prize: not a single clever assistant, but a new layer of software sitting between employees and the systems they already use. If AI agents become the interface, then CRM, ERP, and collaboration tools could be reduced to something closer to back-end engines – useful, necessary, and far less visible than they are now.
Three layers are forming around AI agents
The market is splitting into three distinct layers. First are the tools for building agents. Then come the agents themselves, including browser-based systems that can navigate websites and desktop agents that handle local apps and files. Finally, there are management platforms – the emerging ”agent control panel” category – meant to coordinate fleets of agents from different vendors inside one company.
That is why names like Salesforce Agentforce and Google Gemini Enterprise matter, even if they sound like they were generated by a committee in a hurry. They are not just products; they are bids to become the operating layer for the new automation stack. The same logic explains Microsoft Agent 365 and OpenAI Frontier, which are both trying to manage agents rather than merely create them.
The promise is fewer apps, not more
Supporters of the shift imagine a workplace where people stop clicking around inside a dozen applications and instead direct a set of agents that talk to those systems on their behalf. In that version of the future, software licensing could even move away from per-user pricing and toward charging for agent usage. Vendors love that idea for obvious reasons; customers will probably love it less once the bill arrives.
This is also where the historical pattern shows up. Every major software wave tries to abstract the layer beneath it: GUIs replaced command lines, cloud platforms hid infrastructure, and now agents are trying to hide the applications themselves. The difference is that this transition is happening while the underlying systems are still deeply entrenched and heavily customized, which makes ”simple” deployment anything but simple.
Security and control are the first headaches
The technical pitch runs straight into an awkward reality: today’s browser and desktop agents are still brittle. They can expose credentials, create new attack surfaces, and misfire in ways that make security teams reach for the nearest stress ball. Corporate buyers have also complained that many of these systems are difficult to set up and harder to keep running.
There is also a split in how vendors describe maturity. OpenAI and Anthropic have leaned into research-preview language, while SAP and ServiceNow are presenting their offerings as ready for production use. That difference matters because enterprise IT does not buy ”maybe” – it buys guarantees, or at least something that can survive procurement.
The strange alliance between rivals
The most interesting twist is that the big corporate software vendors are both customers and competitors of the AI model makers. Snowflake, for example, is using models from OpenAI and Anthropic to build agents that work across warehouse data and connected apps such as Microsoft Teams, Salesforce, and SAP. That creates an odd dependency loop: the same model providers help power products that may eventually displace the interfaces they sit behind.
Expect more of that, not less. Large software companies rarely surrender control of a lucrative interface layer without a fight, and the AI agents era gives them a fresh reason to defend their territory. The unanswered question is whether companies will standardize on one orchestration layer or end up with a messy stack of incompatible agent systems that each promise to be the one true cockpit.