Agentic AI in EPM: What Do OneStream, IBM, and Anaplan Have in Common?

The first wave of AI in business applications was largely built around generative chatbots layered on top of existing systems. They could summarize documents, rephrase text, or answer simple questions, but they had no understanding of the data the organization actually relied on.

The second wave introduced specialized point solutions. While effective within specific use cases, they still operated alongside core business processes rather than as an integral part of them.

Both approaches shared the same fundamental limitation. The language model had no understanding of the corporate structure, which reporting periods had been closed, or what data the person asking the question was authorized to access.

Two major developments changed that.

1. Interoperability standards matured, most notably MCP (Model Context Protocol). They allow applications to expose their capabilities as tools for any AI agent while providing the appropriate business context and enforcing user permissions.
2. EPM vendors recognized that their greatest competitive advantage in the AI era is not another chatbot. It is the structured, multidimensional financial data model they have spent years building—together with its business rules, hierarchies, and security framework.

An agent connected to such a model can respond within the context of the organizational structure, intercompany relationships, the fiscal calendar, and the organization’s security model. An agent without that context—even if powered by the latest large language model—may generate a compelling summary while misinterpreting the company structure or exposing information beyond a user’s access rights.

During the Splash 2026 keynote, this distinction was described as the difference between an answer grounded in data and an eloquent guess.

And that is precisely why, today, the architecture providing context to the language model matters more than the language model itself.

The fact that these platforms are converging on a similar architecture does not mean their offerings have become the same. Each vendor has taken a different path to Agentic AI, and that journey continues to shape the way its platform evolves.

Among the three vendors, OneStream has the longest track record of developing AI as an integral part of its platform. Its journey began with early machine learning initiatives in 2017, continued with the general availability of SensibleAI Forecast in 2022, and expanded further with the launch of SensibleAI Studio in 2025, featuring more than 60 prebuilt AI components.

At Splash 2026, this strategy reached another milestone.

• Four AI agents—Search, Finance Analyst, Deep Analysis, and the new Forecast Agent—became generally available. They operate not only within the OneStream platform but also across Microsoft 365 applications, including Excel and Teams.
• OneStream also introduced Finance Agentic Layer, which exposes platform capabilities through the Model Context Protocol (MCP) to external AI assistants such as Copilot, Claude, ChatGPT, and Gemini. Importantly, all interactions respect the customer’s existing security model.

During the keynote demonstration, an external AI assistant retrieved an income statement, analyzed performance variances, and answered follow-up questions—all while enforcing exactly the same user permissions defined within the OneStream platform.

At the same time, OneStream strengthened the platform’s technical foundation. A new cube engine removes previous data scalability limitations, while Developer Studio introduces a new approach to building applications and platform extensions.

This is a significant distinction. Rather than adding AI on top of an existing platform, OneStream has redesigned its architecture to support AI agents as a core part of the platform.

IBM has taken a very different approach. Rather than building a proprietary set of AI agents within Planning Analytics, it has focused on integrating the platform with the broader watsonx ecosystem.

In November 2025, Planning Analytics was integrated with watsonx Orchestrate. As a result, the planning platform evolved from a system used primarily for data storage and calculations into an active participant in business processes.

If a sales forecast falls below a predefined threshold, an AI agent can automatically trigger the appropriate workflow. If inventory drops below a critical level, an alert is generated. Synchronization between finance and operations can take place automatically, without requiring users to manually initiate tasks.

The user defines the business logic. watsonx Orchestrate builds the agent. The data and calculations remain within Planning Analytics.

Today, IBM supports a “bring your own agent” approach, as well as the use of agents available through watsonx Orchestrate. At the same time, the company has announced dedicated AI agents for forecasting, scenario analysis, and the financial close process.

The ecosystem is further strengthened by IBM’s evolving MCP server, which enables external AI agents to access Planning Analytics capabilities, as well as ongoing enhancements to PA Assistant and AI-powered multidimensional forecasting.

Of the three vendors, Anaplan has been the most vocal about its AI strategy. This is no coincidence. The company has announced a $500 million investment in the development of Agentic AI.

Its pace of innovation has been equally ambitious. In December 2025, Anaplan introduced its first role-based AI agents, including Finance Analyst, designed to support performance monitoring and automate monthly finance processes.

In March 2026, CoModeler became generally available—an AI agent that helps model builders create and enhance planning models using natural language. At the same time, Anaplan launched Agent Studio, enabling organizations to build their own AI agents while maintaining governance and control.

The strategy culminated in the announcement of Agentic Enterprise. Anaplan unveiled a comprehensive portfolio of AI agents for the Office of the CFO, covering FP&A, treasury, tax, and investor relations, with general availability planned for October 2026. By the end of the year, similar capabilities are expected to extend to supply chain, HR, and sales. The entire framework is being built on Amazon Bedrock, while remaining open to multiple large language models and AI frameworks.

At the same time, Anaplan strongly emphasizes the economic dimension of its approach. According to the company, large language models should not serve as the calculation engine, as this would be not only less predictable but also significantly more expensive. Instead, the platform’s deterministic calculation engine delivers faster, more cost-effective, and—most importantly—consistent results.

If we set aside product names and marketing messages, it becomes clear that OneStream, IBM, and Anaplan are converging on virtually the same architectural model. At its core are three layers and one guiding principle.

The foundation remains the EPM platform itself. It stores the data, dimensions, hierarchies, exchange rates, business rules, and consolidation logic—and it is responsible for performing all calculations.

Regardless of the vendor, this role is fulfilled by OneStream’s cube engine, TM1 in IBM Planning Analytics, and Anaplan’s calculation engine. Their defining characteristic is determinism: the same input data will always produce the same result.

Large language models operate differently. Their outputs are probabilistic, meaning the same prompt can generate slightly different responses each time. In many use cases this is perfectly acceptable. In finance, however, consistency is not a preference—it is a requirement.

Anaplan’s CEO illustrated this with a logistics analogy. A company that delivers shipments to the wrong warehouse in five percent of cases is not operating “almost correctly.” It is simply operating incorrectly.

The same principle applies to financial reporting. A report that appears credible but contains incorrect figures has no value. That is why, across all three platforms, the large language model does not replace the calculation engine. It is never responsible for performing the calculations.

This does not mean, however, that the role of the large language model is limited. Quite the opposite. The LLM is responsible for exactly the tasks it performs best:

• Understanding questions asked in natural language,
• Explaining results,
• Comparing scenarios,
• Summarizing changes,
• Helping users interpret variances,
• Initiating actions such as running a forecast, starting a workflow, or preparing commentary on financial results.

The key point, however, is the source of the information. The language model does not generate its own calculations. All figures come directly from the EPM platform.

The most important layer of the entire architecture, however, is governance. It is what ensures that an AI agent operates within the platform’s rules, rather than alongside them.

The agent inherits the user’s permissions. If a controller has access only to specific legal entities or cost centers, the agent sees exactly the same data—regardless of whether the question is asked within the EPM platform, in Microsoft Teams, Excel, or through an external AI assistant.

As a result, every response remains grounded in the organization’s data, aligned with its existing security model, and fully auditable. This is what distinguishes enterprise-grade AI from a conventional chatbot connected to a database.

That level of trust cannot be achieved through prompt engineering alone.

The similarities do not end there. All three vendors have embraced one additional principle: openness.

The large language model landscape is evolving rapidly. New models, new protocols, and new AI environments emerge every few months. It is difficult to assume that any single AI platform will remain the industry standard for years to come.

That is why none of the platforms attempts to lock customers into a proprietary AI ecosystem. OneStream exposes its capabilities through Finance Agentic Layer, IBM is expanding its integration with watsonx Orchestrate and its MCP server, while Anaplan emphasizes model and hyperscaler neutrality. Although their technical implementations differ, the underlying message is the same.

The customer chooses the AI agent. The EPM platform ensures that, regardless of which agent is used, it always operates on the same trusted data and within the organization’s existing security and governance framework.

This shift is about much more than adding another AI assistant. Its real significance becomes clear only when the underlying architecture is translated into the day-to-day work of finance teams.

Take variance analysis as an example. Until recently, answering a simple question such as “Why did margin decline?” meant opening multiple reports, comparing different views, filtering data, and manually tracing the underlying drivers. Gathering the necessary information alone could take anywhere from several minutes to several hours.

An AI agent can perform the same analysis in seconds. It can identify where the variance occurred, determine its primary drivers, quantify the impact of each one, and prepare a data-driven narrative based on information retrieved directly from the EPM platform.

The same transformation is taking place in planning model development. Traditionally, building sophisticated calculations or optimizing business logic has been the responsibility of a small group of model architects and administrators. Tools such as CoModeler are beginning to act as intelligent collaborators, helping design models, suggesting improvements, and accelerating development.

This does not replace specialists. It enables them to complete work that previously required significant manual effort much faster.

Forecasting is evolving as well. Rather than being a quarterly exercise carried out under tight deadlines, forecasting is increasingly becoming a continuous process supported by AutoML models that automatically update forecasts while explaining which business drivers caused the projected outcome to change.

This distinction is important. AI is no longer treated as a black box that simply produces a number. Increasingly, it can also explain why the forecast looks the way it does.

There is another change that may seem less dramatic but is often even more important from the user’s perspective. Modern EPM platforms no longer require users to switch applications simply to ask a question. They can interact with AI directly from the tools they already use every day—whether in Excel, Microsoft Teams, or their organization’s preferred AI assistant.

This may ultimately become one of the most important factors driving the successful adoption of Agentic AI across enterprises.

Perhaps most importantly, these capabilities are no longer theoretical concepts or pilot projects. All three vendors are now demonstrating production deployments in large enterprises, ranging from global manufacturing companies to multinational organizations in the media and consumer goods sectors.

Every technology announcement focuses primarily on new capabilities. From an implementation perspective, however, the limitations and prerequisites are just as important. These determine whether Agentic AI delivers real business value—or simply becomes another impressive demonstration.

Today, the biggest limitation is not the large language model itself. More often, it is the quality of the underlying data model.

Disorganized hierarchies, inconsistent dimensions, outdated model structures, or access permissions that have accumulated over years of shortcuts will not disappear simply because an AI agent has been introduced.

On the contrary, the agent will rely on exactly the same data and business rules—it will simply do so much faster and present the results in a far more convincing way. That is why improving the quality of the data model is no longer a task that can be postponed. It has become one of the key prerequisites for a successful Agentic AI implementation.

New capabilities also create new pathways to enterprise data. If users can interact with the EPM platform not only through its native interface, but also via Excel, Microsoft Teams, or external AI assistants, any weaknesses in the underlying permission model can have far broader consequences than before.

For that reason, reviewing and strengthening the organization’s security model should precede the deployment of AI agents—not follow it.

In finance, an answer that is almost correct is often simply incorrect. That is why explainability is so important. An AI agent should be able to show the source of the data, the calculation logic, and the assumptions or parameters used to generate its response.

The final decision—particularly when it has material business implications—must remain with the human decision-maker. This is the position all three vendors advocate today.

The real question, however, is not whether this principle will be upheld, but how it will be implemented in practice during a real-world deployment.

It is equally important to distinguish between capabilities that are available today and those that remain on the product roadmap. OneStream’s SensibleAI agents and Finance Agentic Layer are already available for production use, as are Anaplan’s CoModeler and Agent Studio.

By contrast, Anaplan’s full portfolio of AI agents for the Office of the CFO is scheduled for release in October 2026, with additional functional packages expected by the end of the year. Similarly, several of IBM’s planned AI agents have been announced but are not yet generally available.

This distinction matters. Technology decisions should be based primarily on capabilities that organizations can deploy today—not solely on the long-term direction of a vendor’s roadmap.

There is another aspect that receives far less attention in product announcements: licensing. Increasingly, AI capabilities are licensed separately from the core platform and are often priced based on actual usage.

For that reason, when building a business case for Agentic AI, organizations should consider not only the potential savings generated through automation, but also the total cost of AI adoption—including licensing, consumption-based pricing, and the mechanisms needed to monitor and control AI usage over time.

The good news is that adopting Agentic AI does not require an overnight transformation. What it does require is the right sequence of steps.

Start by strengthening the foundation.

Reviewing your data model, hierarchies, and access permissions will often deliver greater value than immediately deploying another AI assistant. It is also an ideal opportunity to address the technical debt that has accumulated within the EPM platform over time.

Then choose a single, high-value use case.

Automated variance commentary, planning model development, or forecast analysis are all strong starting points. The benefits are easy to measure, while the implementation risk remains relatively low.

Design the architecture deliberately.

If your organization is adopting a corporate AI assistant such as Microsoft Copilot, it is important to understand how your EPM platform exposes its capabilities to that assistant—and, just as importantly, where user permissions and security controls are actually enforced.

Compare architectures, not marketing messages.

Although OneStream, IBM, and Anaplan are moving in a remarkably similar strategic direction, they differ in the maturity of individual capabilities, licensing models, and the way AI agents are embedded into business processes. These differences are likely to have a greater impact on long-term success than the AI branding itself.

The most interesting aspect of the current shift is not that every major vendor is talking about Agentic AI. It is that they have all arrived at remarkably similar conclusions.

A large language model does not replace the EPM platform. It does not replace business rules, the underlying data model, or governance. Instead, it becomes a new way of interacting with them.

This is likely the most significant evolution in Enterprise Performance Management in many years—not because AI will replace planning or financial consolidation, but because it will enable organizations to work with existing EPM models more quickly, more intuitively, and with far less effort.

The organizations that gain the greatest competitive advantage, however, will not necessarily be those that deploy the next AI agent first. They will be the ones with well-governed data, consistent business processes, and a robust operating model. That is the foundation on which Agentic AI delivers its greatest value.