Automating KPI commentary is one of the most practical applications of Claude AI in management accounting. This article explains how the comments are generated, how they are imported directly into IBM TM1 – and why that makes a real difference.
The Problem with manual KPI Annotation in IBM Planning Analytics
The figure is in the report. The context is missing.
12.4 million EUR in revenue for Q3. Is that good? Is that bad? Is it due to project business? To growth in the DACH region? To the major contract lost in the north?
Controlling teams know the answer. But they have to formulate it manually every time, enter it, and make sure the comment ends up in the right place.
That’s time-consuming. And at the system level, it happens often enough that comments end up in a note instead of directly in the cell. Or worse: they aren’t recorded at all because there isn’t enough time.
How does automatic KPI annotation work with Claude AI?
Claude AI provides commentary on KPIs in three steps – all via the MCP server:
- Claude reads the variance – actual vs. plan and vs. prior year – directly from the TM1 cube.
- Claude analyzes the context: Which region? Which cost element? Which time period? Which driver stands out?
- Claude formulates the comment and writes it back directly into the configured TM1 system cell.
Here’s the result:
Write-Back in IBM TM1: Why comments and numbers go hand in hand
The comment must be placed where the number is. Not in a separate document, not in a note, not in an email.
The write-back mechanism ensures that:
- Comments and numbers are always linked – no media discontinuity
- Versioning: Who wrote the comment and when?
- Auditability: Every comment is traceable
- No manual transfer – no source of error
This is particularly relevant for audits and internal reviews: The rating is versioned and traceable right where the number appears.
What this looks like with your own TM1 cubes
Together, we’ll look at how write-back fits into your existing Planning Analytics environment.
KPI Commentary on Quality: What AI does well – and where humans still play a role
That’s a valid question. Here are some insights from real-world experience:
- Quantitative Deviations: Claude is almost always accurate. He reads the numbers directly from the cube.
- Root Cause Analysis: Claude interprets the data based on the structured information. The cleaner the cube structure, the more precise the commentary.
- Tone and Language: Configurable. Factual, suitable for management, and – upon request – in the company’s corporate style.
Recommendation: Give the comments a quick once-over for the first 2–3 months. Most teams report that after that, a quick glance is enough. Fundamental errors are rare.
Automating KPI Commentary: Prerequisites and limitations
Suitable for:
- Controlling teams using IBM Planning Analytics / TM1
- Regular management reports with numerous KPIs
- Environments with a clean cube structure and defined plan values
- IT infrastructure that supports MCP server operations
Not suitable: Highly customized commentary that requires in-depth industry expertise and cannot be derived from the Cube data. In such cases, humans remain the better tool.
- Typical time for manual KPI commentary: 30–90 minutes per reporting period for 20–50 KPIs
- With Claude AI: seconds per KPI, write-back directly into the cube
- MCP server: role-based, P&L / sales / balance sheet kept separate, audit log enabled
- Anthropic Claude: an AI model by Anthropic, specialized in precise, structured text generation
FAQ: Automating KPI annotation with IBM TM1 and Claude AI
Can Claude also write industry-specific comments?
Yes, provided the relevant data is available in the cube. The more structured the cube data is, the more context-specific the comment will be. Industry-specific knowledge can also be configured as context. For more information on how structured data and AI agents work together in Controlling, see our blog post on the topic.
Can we review the automatic KPI comments before the write-back?
Yes. There is an optional review step in which a controller reviews and approves the comment before it is written back to the cube. For an overview of the basics of AI in IBM Planning Analytics and how such workflows are structured, see our overview of the Planning Analytics Assistant.
What happens if Claude misinterprets a deviation?
In that case, the controller edits the comment manually. This is rare, but it does happen, and it’s easier than writing everything by hand. If you’re wondering how ad hoc analysis and automated commenting complement each other effectively, there’s a separate article on that topic as well.
We’d be happy to discuss whether this is a good fit for your commenting processes.
What would this look like with your KPIs?
In our showcase, we’ll show you how Claude AI reads your KPIs live from the TM1 cube, comments on them, and writes them back directly – using your own data.
Book a meeting →
