Why LLMs cannot do math, and how to fix the data layer to solve this problem

AI models process financial data as language tokens, not mathematical objects, leading to confident errors that bypass traditional finance controls. Finance leaders must solve the data normalization problem, mapping disparate ERP text strings into a single data layer, before deploying AI for financial analysis.

The Anatomy of a Confident Error

When a junior analyst makes a mistake in a spreadsheet, it leaves a trail. You see a broken formula or a wildly inaccurate figure that triggers a review.

AI errors do not look like errors. They arrive wrapped in grammatically perfect executive summaries. The AI does not flag its own uncertainty. If it misinterprets a date field and includes a partial month of data, it will confidently present a revenue figure that is off by millions but still looks entirely plausible. Because the output is professional, reviewer alertness drops and the error survives.

This token-processing limitation leads to specific, recurring failures in financial AI.

• Aggregation failures occur when the AI adds numbers that should not be added, or averages figures that should be summed, arithmetic errors a basic calculator would never make.
• Context confusion happens when the model mixes thousands with millions, or blends EUR and USD figures, because both appear in the source text and the model loses track of which currency applies to which token.
• Temporal errors arise when the AI misinterprets fiscal year boundaries, producing a number that is systematically wrong across every period and therefore invisible to trend analysis.

The Multi-ERP Normalization Crisis

The problem compounds exponentially when a group runs multiple entities across different ERPs. Every ERP has its own data model. Tripletex handles revenue recognition differently than Business Central. Fortnox structures its chart of accounts differently than Microsoft Finance & Operations. None of them were designed to communicate with each other, and none were designed with AI reasoning in mind.

When an AI agent tries to answer a simple question it is not looking at just one dataset. It is looking at four structurally incompatible representations of financial reality. It is attempting to translate four different languages without a dictionary. The outputs will reference real numbers, sound authoritative, and be wrong in ways that are extremely difficult to detect.

The Solution: Build the Data Layer First

Before Agentic AI can do anything useful with financial data, finance leaders must solve the normalization problem. Map the data model from every ERP into a single structure. Every account code from every system must map to a consistent group-level equivalent. Every intercompany flow must be identified and tagged so the agent does not double-count it as external revenue. Once this layer exists, account 4010 in one system and account 3000 in another become the exact same mathematical object. The AI is finally reasoning on one version of financial reality.

Implement AI-Specific Financial Controls

Even with a single data layer, traditional finance controls are insufficient. Build input-output reconciliation, an automated check that compares key totals between the raw input data and the AI's final output. If the input sums to $100 million and the AI reports $102 million, the workflow halts. Implement provenance tracking so that every number in an AI-generated report traces back to a specific source record. If you cannot trace it, treat it as a hallucination.

Pro tips

• Never trust inline arithmetic from any LLM in production. Route all computation to a verified executor (Python interpreter, calculator API). This is non-negotiable.
• GSM8K is necessary but not sufficient. It's become a benchmark that models overfit to. Also evaluate on MATH, AIME problems, and especially out-of-distribution multi-step word problems your domain actually contains.
• Verify your fine-tuning data, not just your eval data. A single incorrect reasoning trace in a fine-tuning set can introduce systematic errors. Use symbolic verifiers to check every training example.
• Process reward models (PRMs) outperform outcome reward models (ORMs) for math training. Reward each correct intermediate step, not just the final answer. This is the key insight from DeepMind and OpenAI's process supervision research.
• Mixture of synthetic + formal + human-solved data beats any single source. Aim for roughly 60% synthetic traces / 20% formally verified / 20% human-curated for math-intensive fine-tuning runs.

Stop buying AI tools that promise to magically understand your messy ERP data. Fix your chart of accounts, build a single source of truth data layer, and implement strict input-output reconciliation. That is how you turn AI from a liability into a strategic advantage.