Microsoft, Salesforce, Oracle, OpenAI: Who Wins the AI Agent Wars?
Microsoft, Salesforce, Oracle and OpenAI are rivals and partners at once, all racing to control the enterprise AI agent layer. Here's the fight explained.
What is the fight between Microsoft, Salesforce, Oracle and OpenAI actually about?
It’s about who owns the “agentic control plane”, the software layer that decides which information an AI agent can see, which tools it can use, which company rules it has to follow, and where its finished work lands for a human to review. Microsoft, Salesforce, Oracle and OpenAI are all building toward that same position from different starting points (productivity software, CRM, business applications, and chat interfaces), even as they strike partnerships with each other to get there faster.
TL;DR
- Cloud infrastructure spending among the nine largest cloud companies is on pace to approach $957 billion this year and more than a trillion and a half next year, which gives a sense of the scale of money chasing this shift.
- Traditional B2B software vendors used to stay in their lanes, with Microsoft owning productivity tools, Salesforce owning customer relationship data, and Oracle owning finance and business applications, but agents are erasing those boundaries because real work (closing the books, selling to a customer) never respected them anyway.
- The real prize isn’t the smartest model, it’s the “agentic control plane”: the system that supplies context, enforces permissions, and coordinates agent work across a business, regardless of which AI model is doing the task.
- Rivals are actively helping each other reach customers. Salesforce built a bridge so Microsoft’s Copilot Studio agents can work inside Salesforce-coordinated processes without being rebuilt, and Oracle is building a workspace that brings its own applications into ChatGPT.
- Better models don’t end the competition. A more capable Microsoft agent can make a Salesforce-coordinated process more valuable, and Salesforce benefits without ever having to build that model itself, so both companies can gain from each other’s progress while still fighting over the next sale.
- Shared business context is the next battleground. Salesforce’s Agent Context Engine and Microsoft’s “site queue” style products are both attempts to give agents a common understanding of business data, definitions, and rules, even when those agents were built elsewhere.
- No one has built “Windows for agents” yet. The control plane that dominates enterprise AI the way Windows dominated the desktop in the 1990s doesn’t exist, and that open position is what’s driving billions in partnership and product announcements right now.
Without shadow IT.
How did these companies end up competing if they sell different things?
For most of the last two decades, Microsoft, Salesforce and Oracle sold adjacent but distinct products. Microsoft handled documents, email, and internal collaboration through Outlook, Excel, and Teams. Salesforce managed the sales pipeline and customer records. Oracle ran finance and back-office applications. Customers routinely bought from multiple vendors at once because each company had a legitimate, separate relationship with a different part of the business.
That arrangement worked because software followed departmental lines. But actual work never did. Closing the books touches spreadsheets, emails, and accounting systems all at once. Selling to a customer means reading an email, checking a number in a different system, and deciding what to do next, often without ever officially leaving the “sales” department. Humans have always bridged those gaps by moving between applications themselves.
Agents are now being asked to do that bridging work. And the moment an agent has to pull from multiple systems to finish a task, somebody has to supply the layer that governs which data it can touch, which instructions are current, and who has to approve what. That layer sits on top of all the old departmental boundaries, which is exactly why companies that used to coexist peacefully are suddenly circling the same territory.
Why would Salesforce make it easier to use a competitor’s product?
On October 27th, Salesforce rolled out a bridge that lets Microsoft’s Copilot Studio agents participate in Salesforce-coordinated work without the customer having to rebuild those agents from scratch. On its face, that looks like Salesforce doing Microsoft a favor. In practice, it’s a bet on where the money actually is.
Salesforce’s advantage has always been distribution: a massive base of sales teams and enterprise customers who have no appetite for ripping out their CRM just because a new model scored better on a benchmark. That distribution is still valuable, but it isn’t the whole game anymore. The more important question is who coordinates the agent’s work once it’s running, meaning who supplies the context, enforces the permissions, and decides where results go for review.
By making Microsoft’s agents work smoothly inside Salesforce-managed processes, Salesforce reinforces its claim to that coordinating role, even when a competitor supplied the underlying model. Microsoft, meanwhile, gets to put a more capable agent in front of the customer and has every incentive to sell more of the surrounding system around it. Both companies can win from the same deal while still competing hard for whatever the customer buys next.
What is Oracle doing with OpenAI, and why does it matter?
Oracle and OpenAI announced an expanded partnership the same day as the Salesforce-Microsoft bridge. Oracle is developing a workspace, still in development, that would bring Oracle’s business applications and data directly into ChatGPT. The example described is a finance employee using ChatGPT to investigate costs, review purchasing decisions, and check what’s outstanding before month-end close, pulling from finance, procurement, and planning applications, all subject to the company’s existing permissions.
Nothing gets lost.
The logic mirrors the Salesforce-Microsoft move. Oracle gets more value out of applications customers are already paying for, because now an employee doesn’t have to leave ChatGPT to use them. OpenAI gets a foothold as the place where that employee actually works through problems, and if the experience is good, the next request stays in ChatGPT too, with more departments and data sources added over time. A narrow starting point in finance becomes a wedge for a much larger role.
This builds on OpenAI’s earlier moves to connect business systems directly to its models, an effort it pushed forward with the launch of Frontier earlier in the year.
Is the agentic control plane actually worth fighting over?
Yes, because of what it controls, not because of which model happens to be attached to it. An agent can have access to perfectly accurate data and still get something wrong if nobody has defined what the numbers mean in context. A spreadsheet might hold the correct revenue figure, but an agent could apply the wrong definition of “revenue” when summarizing it. A customer record might be fully up to date and still miss a promise an employee made over email yesterday.
Connecting systems gets an agent access to data. It doesn’t automatically give that agent judgment about what the data means for a given decision. That judgment still has to come from somewhere, whether it’s business rules encoded by the vendor, context supplied by a system like Salesforce’s Agent Context Engine (currently in preview), or Microsoft’s related “site queue” style products that gather business data, definitions, and rules in one place.
Whoever supplies that judgment at the moment an employee needs it earns the right to keep getting paid for it. That’s the actual commercial logic behind the control plane fight: not owning every agent, but owning the layer that makes any agent trustworthy enough to act on inside a real business.
Frequently Asked Questions
What does “agentic control plane” mean?
It refers to the software layer that oversees how AI agents operate inside a business: which information they can access, which tools they’re allowed to use, what rules and approvals apply, and where their output goes for a human to review. It’s the governance and coordination layer sitting above individual AI models.
Why are Microsoft, Salesforce, Oracle and OpenAI partnering if they’re competing?
Because distribution and model quality aren’t the same prize. A company can benefit from a rival’s better model (Salesforce gains when a Microsoft Copilot agent performs well inside a Salesforce-coordinated process) while still competing for the bigger, more durable prize: being the system that coordinates and governs the agent’s work long-term.
Does a better AI model end this competition?
No. In an enterprise setting, model quality is only one input. A more capable agent can make a coordinated process more valuable without the company that owns the coordination layer needing to build that model itself. That’s why rivals can improve each other’s offerings while still fighting over the next contract.
What is Oracle actually building with OpenAI?
A workspace, still in development, that brings Oracle’s business applications and data into ChatGPT, starting with use cases like a finance employee investigating costs and purchasing decisions ahead of month-end close, within the company’s existing permission structure.
Is there already a dominant company in the agentic control plane space?
Nothing gets lost.
Not yet. The comparison the analysis draws is to Windows in the 1990s, which became the dominant layer shaping how people worked on computers. No single company has reached that position for AI agents in the enterprise, which is exactly why Microsoft, Salesforce, Oracle and OpenAI are all racing to define it on their own terms.
