Enterprise AI is entering a more interesting phase.
The first wave was dominated by copilots: place a conversational interface beside an enterprise application, retrieve some data, summarize it, and occasionally trigger an existing transaction. The current wave is materially different. SAP, Salesforce and Workday are all moving toward agents that can reason over business context, use governed tools, interact with other agents and execute work.
At first glance, their strategies look increasingly similar.
SAP has Joule Work, Joule Agents and Joule Studio. Joule Studio is positioned as an environment for building agents, applications and workflows, grounded in SAP business context. SAP is also exposing S/4HANA capabilities to agents through an Agent Gateway and MCP, and is adopting A2A for agent-to-agent interoperability. SAP Help Portal
Salesforce has Agentforce, Data 360, MuleSoft, Flow and an increasingly explicit combination of adaptive AI reasoning with deterministic automation. Agentforce can use Salesforce metadata and enterprise data for context, while MuleSoft connects agents to external systems and exposes actions through APIs, MCP and A2A. Salesforce
Workday is taking a somewhat different angle. In addition to application-specific agents, it is building an Agent System of Record to register, govern, monitor and audit agents—including third-party agents. Its newer Trusted Agent Platform and Data Cloud direction gives agents governed access to Workday data and business logic while retaining Workday permissions and approvals. workday.com
It is tempting to conclude that enterprise AI is becoming commoditized.
That would be the wrong conclusion.
The more useful question is: what does each of these platforms understand particularly well, and what happens when the work does not fit neatly inside that understanding?
The OEM advantage is not the language model
SAP, Salesforce and Workday do not have their strongest advantage because they can call a better LLM.
Their advantage is that they already sit on top of structured business meaning.
SAP knows a purchase order is not merely a JSON document. It exists inside purchasing organizations, plants, company codes, suppliers, goods receipts, invoices, authorizations and decades of process semantics.
Salesforce has an equally strong advantage around customers, leads, opportunities, cases, service interactions and the metadata describing how those objects behave.
Workday understands workers, organizational structures, positions, compensation, financial processes, approvals and security relationships.
This matters enormously for AI.
A model can reason impressively over text. Reliable enterprise action requires considerably more. It needs to know which data is authoritative, which action is permitted, which application owns the transaction, what the current state means and what constraints apply.
SAP is leaning heavily into precisely this advantage. SAP describes its Business AI stack as being grounded in its business data fabric, Knowledge Graph and domain models. Joule Studio’s current documentation describes agents that can access S/4HANA through MCP tools and make tool-selection decisions using business context. SAP
Salesforce is making essentially the same strategic move from a different starting point. Agentforce uses Data 360 and the Salesforce metadata layer to give agents context and combines this with Flow, MuleSoft and existing business logic for action. Salesforce
This is what makes OEM AI formidable.
The OEM already owns a large part of the semantic terrain.
But enterprises do not work one OEM at a time
This is where the problem becomes more difficult.
Consider a relatively ordinary activity: resolving an invoice exception.
The employee may need information from an ERP. The supplier correspondence may be in Outlook. An exception tracker may live in Excel. A logistics confirmation may come through Teams. A policy may exist as a PDF. One approval may happen inside SAP while another is informally obtained from a business manager.
The person’s job is not:
operate SAP.
Nor is it:
operate Microsoft 365.
The job is:
resolve the invoice exception.
That distinction becomes fundamental for enterprise AI.
OEM AI naturally tends to organize intelligence around the OEM’s existing application boundary.
Real work is usually organized around a responsibility that crosses those boundaries.
SAP itself is already moving beyond the traditional embedded-copilot model. Joule Work is described as a workspace spanning SAP and non-SAP systems, and SAP says it can consolidate information about a business process or transaction into a common experience. SAP
Salesforce is similarly extending outward through MuleSoft.
Workday is explicitly designing its Agent System of Record to govern third-party agents as well as Workday agents. blog.workday.com
All three therefore recognize the same problem: the enterprise boundary is larger than the application boundary.
But their centre of gravity remains different.
SAP: probably the deepest business-process context
Among the three, SAP arguably has the most ambitious enterprise-process AI direction.
The new Joule Studio is no longer simply an extension framework for a chatbot. SAP describes it as an AI-first environment for creating agents, applications and workflows from business intent. Its broader architecture increasingly combines Joule, SAP Business Data Cloud, SAP Knowledge Graph, Signavio, LeanIX, Integration Suite and an agent gateway. SAP News Center
That is powerful because SAP possesses something difficult to manufacture from scratch: decades of encoded enterprise process semantics.
For an SAP-centric enterprise, an AI system that already understands purchasing, finance, manufacturing and logistics—and inherits the security context of the logged-in user—has an enormous starting advantage. SAP explicitly documents that Joule operations are bounded by the user’s existing authorization. SAP Help Portal
The limitation is almost the mirror image of that strength.
A real enterprise is rarely only SAP.
The interesting question therefore becomes how effectively SAP can extend its semantic understanding into the messy territory that SAP does not own without forcing everything into SAP’s conceptual model.
That is not fundamentally a model-quality problem. It is an enterprise-context problem.
Salesforce: customer context plus an unusually strong action surface
Salesforce approaches the problem from a different direction.
Its native semantic advantage is customer-facing work. But Agentforce gains considerable reach from the rest of the Salesforce stack: Data 360, Flow and particularly MuleSoft.
That makes Agentforce interesting because the architecture is not merely:
agent talks to CRM.
It is increasingly:
agent reasons using Salesforce data and metadata and can invoke governed actions across a broader application landscape.
Salesforce explicitly describes Agentforce as combining deterministic business logic with adaptive reasoning and using MuleSoft as an automation layer across systems, APIs, workflows and agents. Salesforce
That hybrid model is important.
Enterprise AI will not become reliable by converting every decision into probabilistic agent reasoning. Stable work should remain deterministic. Agents should reason where reasoning adds value; established business rules and transactions should remain inspectable and predictable.
Agentforce appears to understand this increasingly well.
Its challenge is again one of scope. A customer-service agent living mainly in Salesforce is natural. An agent responsible for an end-to-end manufacturing exception involving SAP production orders, plant spreadsheets, engineering documents and supplier correspondence is much less naturally a Salesforce problem.
Workday: the governance problem may be larger than the agent problem
Workday’s strategy is particularly interesting because it has identified a problem that is likely to become acute very quickly.
Enterprises may soon have hundreds or thousands of agents.
Who owns them?
What are they allowed to do?
Which employee or organizational unit are they acting for?
What information may they access?
Are they still being used?
What did they do yesterday?
What did they cost?
Workday’s answer is the Agent System of Record.
Its metaphor is unsurprising but powerful: enterprises already manage human workers through systems of record; AI workers also require registration, ownership, permissions, lifecycle management and accountability.
Workday says ASOR can register, configure, activate and deactivate agents and maintain records of their activity, including third-party agents. workday.com
Whether the long-term winner of this category is Workday or something else remains open. But the category itself is significant.
Agents cannot simply proliferate as API keys plus prompts.
Enterprise AI eventually requires institutional memory and institutional control.
The harder problem: the context between systems
This is where I think much enterprise AI discussion remains too technology-centric.
The enterprise does not merely have a data problem.
It has a context fragmentation problem.
A significant amount of operational knowledge lives somewhere between formal systems and people’s heads:
a spreadsheet maintained because the ERP cannot represent a short-term planning need;
a Teams discussion explaining why a normal policy does not apply this week;
a supplier-specific practice known by an experienced buyer;
an SOP whose written process differs subtly from what people actually do;
an approval obtained informally because the configured workflow never anticipated the exception.
Calling all of this “unstructured data” understates the problem.
Much of it has structure in human work, but that structure has never been explicitly captured.
There are instructions, precedents, working assumptions, evidence, exceptions and local vocabulary. Humans reconstruct this context continuously while moving between applications.
AI can reduce that friction—but only if we give it sufficiently well-managed context.
This suggests that the next enterprise AI platform may need to think beyond both the application and the agent.
The natural unit may instead be the workspace around a responsibility.
A procurement workspace, for example, could bring together the relevant enterprise-system queries, working context, policies, evidence and AI assistance needed to perform a particular class of work.
The source systems remain the source systems.
Excel does not need to disappear tomorrow.
SAP remains authoritative for the purchasing transaction.
The AI layer does not need to reproduce the entire enterprise ontology.
It needs enough context to help a person perform this responsibility reliably.
Not everything should become a workflow
There is another trap emerging as AI and BPM converge.
Once we can combine agents, humans and deterministic activities inside increasingly sophisticated workflow platforms, there is a temptation to assume that the eventual destination of enterprise work is a formal executable process.
It isn’t.
Some work is extremely stable and should become deterministic.
Some has a stable sequence but contains judgment-heavy steps.
Some is largely contextual.
Some changes too frequently to justify formalization.
And some remains genuinely exploratory.
The better model is a continuum.
At one end is loosely structured human work assisted by AI. As repetition appears, guidance can become more explicit. Certain decisions can become rules. Recurring sequences can become workflows. Stable transactions can become conventional software.
But even mature processes continue to encounter exceptions, new information and contextual judgment.
The objective therefore should not be to turn the entire enterprise into BPMN.
It should be to apply the right amount of structure to each piece of work.
This is perhaps the most interesting distinction between enterprise AI and conventional process automation.
AI allows useful automation before everything has been formalized.
What should enterprises learn from the OEM platforms?
SAP, Salesforce and Workday should not primarily be viewed as three competing chatbot products.
They reveal three important truths.
SAP shows the value of deep business semantics and process context.
Salesforce shows the power of combining agent reasoning with an existing action and integration fabric.
Workday highlights the coming requirement to govern agents themselves as enterprise actors.
A serious enterprise AI architecture probably needs all three properties:
context, action and governance.
But there is a fourth problem that no single OEM automatically solves:
work itself often spans the boundaries between their systems.
That is where the next layer of enterprise AI is likely to become interesting.
The winning architecture may not replace SAP, Salesforce or Workday. In fact, attempting to reproduce the business semantics already encoded inside those products would be wasteful.
Instead, it may sit across them—bringing together the context needed for a responsibility, allowing humans and agents to work together, invoking the underlying systems under their existing controls, and gradually introducing stronger structure where repeated work warrants it.
The question for enterprise AI is therefore shifting.
It is no longer simply:
Which model should we use?
Nor even:
Which agent platform should we choose?
The more consequential question is:
How do we bring the right business context, authority, tools and degree of structure together at the point where work is actually being done?
The OEMs are each answering an important part of that question.
The space between their answers may ultimately be where some of the most valuable enterprise AI products are built.