Financial Reporting AI Agent for IBM Planning Analytics

We have created a Financial Reporting Agent for IBM Planning Analytics. The agent automates periodic financial reporting using data in IBM Planning Analytics and related infrastructure. It can be deployed on AWS or Azure following the Microsoft Foundry or AWS Bedrock framework, and managed via Watsonx Orchestrate.

Our business objective and opportunity is to automate financial reporting. When actuals update at the end of each month, the agent generates a financial report in PDF format that outlines the financial results including drivers of performance, variances, etc. 

Here is a sample report output from the agent: 

The agent is unique for the following reasons:

• It is customized for every individual Planning Analytics implementation

• It understands the structural relationships between cubes, rules, and processes

• It operates autonomously based on simple parameters (Company, Period, etc.)

• It returns insight based on summarization of data, and leverages its toolset along the way to drill down into the drivers of performance

Client Use Case – Example  

We deployed the agent on a Business Profitability Model with one of our customers. The initial setup contains specific tooling to interrogate and document the TM1 model to build up to the knowledge base layer. Once the custom knowledge base is in place, the agent can start creating reports. Information in the report includes— but is not limited to— the following: 

• Financial summary with a fully allocated P&L at the business level  
• Analysis of variances to prior periods
• Comparison of Standard and Fully Allocated P&L, including indirect expenses
• Drill down to key expense drivers and allocation 
• Analysis of revenue by project and client (leveraging level of detail stored in separate cubes) 
• Drill down to employee level detail across cubes to shows allocation of employees to projects and the downstream contributions to revenue 
• Drill down to expenses at the line-item level 

How Does it Work

Here is a conceptual picture of the solution. The description of each layer is below. 

Solution Description: 

1. Existing Finance Infrastructure: This node represents a reasonably complex financial system that includes multiple cubes and data stores. Typically, there are rules and processes that manage the data flow in an application, and these need to be parsed so that the lineage of specific data can be understood at the top level in IBM PA. Data also enters the Planning Analytics application from a wide variety of relational and multi-dimensional data sources that must be reflected by the documentation layer. In our case, we assume that the multi-dimensional data store is IBM PA, and all our current use cases are on IBM PAoC (V11) or IBM PA SaaS (V12). 

2. The "ACG Knowledge Base" layer is a set of custom tools and logic for building and accessing metadata about the TM1 model. They scan the finance application and document the relationship between different objects on the TM1 server (summary information is pushed to detail with "Process X"). We can then serve this documentation to the agent so that it knows where to look for arbitrary data in the TM1 model.
3. The reporting agent is a custom agent invoked with simple parameters. Once the agent receives a request from the user, it will begin to compile the required data from the finance infrastructure. If at any point the agent encounters a data quality issue or a significant variance, the agent will leverage its "Knowledge Base" and retrieval tooling to drill into the root cause or relevant drivers.
4. The top layer represents the User Interface as well as the access and governance controls. We have tested the recommended approach using Watsonx Orchestrate, Microsoft Foundry, and AWS Bedrock. While they all function similarly, the specific configuration and approach differs between providers. Any of the above frameworks can be used depending on existing client infrastructure. 

Business Value

We like this approach for a number of reasons: 

Low entry point for AI 

It provides a low entry point option for leveraging AI and for providing immediate business benefit. The agent can be deployed in 2-3 weeks and provides real tangible results early on. Clients can then continue to explore options to expand AI use as appropriate. 

Low Cost 

It is a low-cost step to start the AI journey. Instead of rolling out agents to hundreds of users on Day 1 and risking data quality and performance issues, this approach offers a way to manage the cost and risk early on. The cost of the LLM in directed workflows like this is minimal, when compared to the cost of many users asking arbitrary questions in a chat interaction. 

Customizable 

The agent is configured on the specific client TM1 model so that it can understand the structure. The output can be customized to each client’s specific reporting requirements. 

Low risk of errors and “hallucinations” 

Since we are accessing a well-defined and documented model in a controlled fashion, the risk of “hallucinations” or incorrect responses remains minimal. This can be further mitigated by fine-tuning the agent's instructions and if needed, using the agent to validate responses before distributing the final report. 

In case of questions, or if you would like to discuss in more detail, contact us at solutions@acgi.com.