In the Age of AI, What Makes an AI Right for FP&A?
Generative AI sounds convincing — but it can get the numbers wrong. Is that really a fit for finance? Meet Numen's rule-based AI, purpose-built for FP&A.
Have you tried generative AI yet? With the recent buzz around turning photos into Ghibli-style images, it feels closer to everyday life than ever. And it's not just personal use — AI is being adopted aggressively across a wide range of industries and functions. Automating customer support, generating images and video, drafting documents — more than any tool before it, AI feels like an inflection point for workplace productivity.
We covered the benefits of bringing AI into an FP&A team in an earlier piece.
[Reclaiming Your FP&A Team's Wasted Time with AI]
We're as convinced as ever that AI will lift productivity in finance too — but given the nature of finance, you can't just reach for any AI. So what conditions does the AI an FP&A team uses need to meet?
The blind spots of LLM-based generative AI
1. "It sounds right, but the numbers can be wrong"
Ask generative AI a question and it'll understand anything you throw at it and give you an answer — even when that answer is flat-out wrong. According to a 2024 study, three to nine out of every ten pieces of information AI generates can be incorrect.
Hallucination rates for LLM-based generative AI (*)
GPT-3.5: 39.6%
GPT-4: 28.6%
Google Bard: 91.4%
In fact, when Microsoft demoed its ChatGPT-based AI features in Bing, it summarized a financial report and got figures like gross margin and operating margin wrong (**).
(**) Source: [PC MAG - Demo of Microsoft's AI-Powered Bing Included Several Small Mistakes]
Even if the report itself looks polished, a single wrong unit or misplaced comma is enough to destroy the credibility of an accounting or financial report — which makes a hallucination rate like this a serious risk.
2. Leakage in the name of "training"
Generative AI "trains" on data — that's how it improves itself. The problem is, you have no idea how the data it trains on gets used or stored.
In 2023, Samsung Electronics used ChatGPT for source-code review, only to have sensitive internal code leak out — and ended up banning ChatGPT company-wide.
A finance team has even more reason to be on guard about leakage. Financial data isn't just numbers — it's sensitive information that encodes a company's strategy and intent, which sets it apart from other data entirely.
What an "AI for FP&A" has to deliver
For the reasons above, AI built for FP&A has to meet a different set of prerequisites than generative AI.
First, it must never state a wrong number. Second, it must rigorously protect data security.
What finance needs is an AI that eliminates both numerical risk and security risk — one built on integrity and trust.
FP&A needs "rule-based" AI, not an LLM
As we've seen, working with finance and accounting data — where accurate inputs and the strict application of rules are non-negotiable — is inherently limited for LLM-based AI, which is swayed by the user's wording and context. Numen's rule-based AI, by contrast, works from accurate data and the strict application of rules.
1. Analysis grounded in ERP ledger data, with integrity checks
Numen runs on automation logic grounded in accounting standards to deliver accurate figures. It performs financial analysis on the journal entry (ledger) data recorded in your ERP. Every figure is verified through AI-driven reconciliation logic, and when an error surfaces, it traces back to the source data to re-verify accuracy.
2. Evidence-based automation = rule-based AI logic
Numen's automation operates on verified accounting principles and financial standards — not uncertain guesses. It's rule-based AI logic that runs on predefined formulas and conditions: reproducible, explainable automation that CPAs and CFOs can trust.
e.g., automated KPI calculations, automatically applied account-classification rules, and more.
3. How it differs from LLM-based AI
Unlike an LLM (large language model) that reasons probabilistically, Numen is accuracy-first AI optimized for ERP-based numerical logic. When useful, an LLM plays a supporting role — explaining, summarizing, or offering recommendations — but financial figures are always produced with quantitative, rule-based logic alone.
4. Data security and a user-driven access model
Numen never transmits customer data externally, and ledger data cannot be accessed without the customer's explicit consent. All data is stored encrypted, and rigorous internal security policies apply — including an AI auto-deletion feature for sensitive data like counterparty names and memo descriptions.
[How to get started with Numen] ｜ [Numen's security policy]
In the age of AI, finance — of all fields — needs a finance-specific AI. Right now, countless finance teams are already using AI accurately and securely with Numen. If you're weighing AI adoption, choose the Numen FP&A solution.
Looking to bring AI into your company's FP&A function?
Get FP&A insights validated on the floor — delivered every Tuesday morning.
