Who Owns the Context Owns the Agent

On 6 October 2026, at SAP Connect in Las Vegas, SAP chief executive Christian Klein asked a rhetorical question: why build agents in isolation on a large language model platform? The question, reported by CustomerThink ↗, was not aimed at customers. It was aimed at competitors.

That same year, Salesforce built its strategy around the “agentic enterprise”; ServiceNow positioned itself as a place to govern everyone’s agents; OpenAI launched a platform to connect agents to enterprise systems; and Microsoft brought ontologies and business rules into its data platform. Each argues that agents should be able to work within its environment.

To understand what is at stake, it helps to look beyond the agents themselves and ask what each platform puts at their disposal.

The table summarises the companies’ stated positioning; it is not an independent performance comparison. Commercial figures are attributed to their sources: SAP Connect (SAP ↗), quarterly results (Salesforce ↗), Microsoft Fabric IQ ↗ and OpenAI Frontier ↗.

SAP introduced its “Autonomous Enterprise” at Sapphire in May 2026 and announced its expansion at SAP Connect. Since the Sapphire launch, the company says it has made more than 20 assistants and over 200 agents available, with a target of exceeding 400 by year-end. SAP has also described its Knowledge Graph as a representation of more than seven million data fields. These are capabilities and targets reported by SAP, not an independent measure of agent effectiveness.

According to SAP, its new Joule Work interface is used by its 110,000 employees. The company has reported productivity improvements of over 20% in HR, finance and procurement. That figure, too, should be read as a result reported by SAP, within its own context and measurement approach. Anthropic’s Claude is among the models available to Joule agents.

SAP’s argument is precision. In critical processes, “almost right” is not good enough: agents need consistent data and rules, and their actions must be traceable. It is a strong argument. But “autonomous enterprise” also needs to be read carefully: in practice, autonomy requires supervision, permissions and the ability to intervene. SAP itself presents the journey as gradual and governed.

Salesforce has taken an explicit route: connect the agent to the customer relationship and invest in the quality and management of the data that supports it. In fiscal Q3 2026, Salesforce said Agentforce and its Data 360 platform had reached nearly $1.4bn in annual recurring revenue, up 114%, with more than 9,500 paid Agentforce contracts. These are company-reported figures in its financial results ↗.

Another significant development was Salesforce’s acquisition of Informatica, completed in November 2025. Informatica’s portfolio includes data cataloguing, integration, governance, quality and master data management. The strategic interpretation offered here is that the acquisition may help Salesforce bridge the gap between data scattered across systems and context that agents can use. It does not prove that the problem has been solved. It does show how data readiness has become part of the contest over agents.

ServiceNow has adopted an ambitious position. At Knowledge in May 2026, it expanded AI Control Tower to discover, observe, govern and measure AI systems deployed not only on its own platform but also across environments such as AWS, Azure, Google Cloud, SAP, Oracle and Workday. The company announced 30 integrations to broaden this coverage. With Action Fabric, agents can invoke ServiceNow actions and workflows through MCP, with identity, permissions and audit controls.

ServiceNow’s thesis is that a process crossing finance, HR, legal and accounting is not fully visible to any single vertical system. It therefore wants to position itself as a control and action layer over the systems already in place. This could be useful in cross-functional processes; how much operational control is actually centralised, and with what dependencies, remains a case-by-case question.

Microsoft and OpenAI approach the problem from the data-platform and model side. In Fabric IQ, an ontology lets organisations define business concepts, entities, relationships, properties and rules, and connect them to data. The stated goal is to give people, applications and agents a shared representation of the business. Microsoft’s documentation notes that ontology features are in preview: the direction is significant, but that does not make the capabilities mature or universally available.

With Frontier, OpenAI describes a Business Context layer connecting data warehouses, CRM and internal applications, so agents can work with the same information and build institutional memory over time. Here, too, this is the vendor’s description of its platform. For a company, the challenge is not merely connecting systems. It is deciding which sources are authoritative, what the information means and who can authorise its use.

One feature makes this contest different from earlier platform wars: the competitors are contributing to shared protocols and connections.

In December 2025, Anthropic donated the Model Context Protocol (MCP) to the Agentic AI Foundation (AAIF), hosted by the Linux Foundation. AAIF was founded by Anthropic, Block and OpenAI, with support from organisations including AWS, Google, Microsoft and Cloudflare. At the time of the announcement, MCP had more than 10,000 public servers, according to the project’s official blog ↗. The protocol standardises how AI applications connect to tools and data sources.

Google introduced A2A, a protocol for agent-to-agent communication. In August 2026, A2A joined AAIF as a hosted project, placing it alongside MCP within the same foundation. SAP documents MCP for connecting agents to systems and A2A for communication between agents in Joule Studio.

If connections become more open and interoperable, competitive advantage may shift to what travels through them: quality, meaning, relationships, rules and permissions. SAP is investing in its Knowledge Graph and Business Data Cloud; Salesforce acquired Informatica; Microsoft offers Fabric IQ; OpenAI includes Business Context in Frontier; and ServiceNow is building control and action tools. The offers are not equivalent, but they converge on one point: an agent needs more than a model and a technical connection.

The contest is not only about who builds the best agent. It is also about who holds and governs the meaning of a company’s data.

Away from the stage, adoption data paints a more cautious picture. In McKinsey’s global 2025 State of AI survey, 62% of respondents said their organisation was at least experimenting with AI agents. Twenty-three per cent said they were scaling their use in at least one area; in no single function did the share exceed 10%. These are results from a survey with a specific sample and definitions, not a universal estimate of every company.

Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls. The prediction was published in June 2025. Gartner also estimated that only about 130 of the thousands of vendors calling themselves agentic offer genuinely agentic products, warning about “agent washing”. This is a forecast attributed to Gartner, not an outcome already observed.

In the SAP world, the starting point is even more concrete. The 2026 investment report from DSAG, the German-speaking SAP user group, found that 54% of responding organisations used SAP ECC or the earlier Business Suite, down from 68% in 2024; responses about ERP solutions in use may include multiple environments within a company. Standard maintenance for SAP Business Suite 7 ends at the end of 2027 for most customers. This does not mean that all these companies are standing still or unprepared. It is a reminder that many are still managing complex systems and migrations while evaluating agents.

Data quality, undocumented customisations and processes held together by spreadsheets and email matter in this context. Announcing an agent does not remove the work needed to make the ground beneath it reliable.

This is where the debate about agents returns to earth. SAP has a substantial legacy of processes and transactions; Salesforce, customer relationships; ServiceNow, operational workflows. But owning data does not mean it has been made usable by an agent.

An agent working with inconsistent master data, misaligned hierarchies or forgotten customisations does not automatically create intelligence. It can automate confusion, and do so faster than a human team can detect it. If “active customer” means one thing in CRM and another in accounting, the system will not necessarily choose the right definition. It will use the one made available through its context and tools.

That is why the most important work is not only building agents, but preparing the conditions in which they operate: data quality, shared definitions, explicit rules, clear permissions and people accountable for the business. It is not the most spectacular work. It determines whether an agent deserves trust.

If value concentrates in context, the provider hosting it may become difficult to replace. A company can change its model relatively easily; transferring definitions, business rules, relationships and organisational memory is harder if they live only inside one provider’s platform.

Tensions are already visible. In 2026, SAP published a harmonised API policy governing the use of its cloud interfaces. SAP describes it as a way to protect stability, security and innovation at scale. DSAG has raised concerns about the effect of the new provisions on customer use cases and third-party applications. Both positions matter: the policy alone does not prove an intent to prevent portability, but it makes clear that data-access terms are a strategic and contractual issue.

Open standards such as MCP and A2A can make connections easier. They do not automatically make the meaning of data portable. Portability of definitions, rules and relationships therefore needs to be designed deliberately.

One concrete example of a bounded experiment is J4C, an SAP assistant with controlled access to four verifiable sources. It is a hypothesis to be measured, not a demonstrated result.

Before entrusting hundreds of agents with a complex SAP environment, one could start with a bounded assistant, J4C, with controlled access to four verifiable sources. The idea is that each answer should be traceable to a source consulted.

The connection could use MCP, which SAP documents in Integration Suite ↗, together with SAP and community tools. The model does not automatically “know” how a company’s environment works: it consults defined sources and tools. The question then becomes how good those sources are, what permissions are granted and how answers are checked.

The value of a system such as J4C would depend above all on its operating rules:

What has not been demonstrated? That four sources will be enough. A system such as J4C may be useful for technical and development questions whose answers can be found in documentation or code. It will not by itself resolve inconsistencies in business data accumulated over decades; that takes data models, quality rules and clear accountability. It is a possible starting point for making the problem visible and measurable, not proof of a solution that can be generalised.

For business leaders, the platform contest suggests four decisions.

**Map your context.** Identify critical processes and data, and check which are reliable before granting an agent autonomy.

**Decide where it lives.** Definitions, rules and relationships are strategic assets. Deciding where to represent them involves the business, IT, governance and procurement.

**Design for portability.** Prefer documented interfaces and, where possible, maintain your own representation of data meaning, independent of a single provider. Examine data-access terms before committing.

**Start small and measure.** Run bounded experiments with real questions, explicit quality criteria and human oversight. An assistant such as J4C is one example of this approach, to be validated in practice.

Is the autonomous, agentic or governed enterprise promised by every provider a concrete direction, or another layer of complexity on top of systems companies are still modernising?

The answer will not depend on the number of agents, but on two factors: how trustworthy the data they use is, and who holds its meaning. Owning context can give a platform an advantage. For companies, the priority is to know what that context contains, who can change it and whether it can be taken elsewhere.

This article provides general information and reflects the author’s opinion where indicated. Announcements and market data are attributed to their cited sources; J4C is presented as an experimental example, not a validated outcome. Sources checked on 17 October 2026.