In 2024, every accounting software company added "AI" to its marketing. Some of them meant it. Most of them did not. The distinction matters — not because AI is inherently better, but because the wrong kind of AI in a finance context creates audit risk, not efficiency.
The three things vendors call "AI"
A rules engine says: if the vendor name contains 'AWS', code it to cloud infrastructure. This is not AI. It is a lookup table. It works well for known vendors and predictable transactions. It breaks immediately when a new vendor appears or a transaction description changes. Most 'AI-powered' accounting tools are rules engines with a machine learning wrapper that helps you build the rules faster.
True machine learning in accounting means: given the last 18 months of transactions, predict how this new transaction should be coded. This is more powerful than a rules engine because it handles novel situations. But it has a critical weakness: it learns from your historical data, which means it also learns from your historical errors. If your team miscoded a category for six months, the model will learn to miscode it too.
LLMs can read a contract and extract the performance obligations. They can read a vendor invoice and identify the service period. They can draft variance commentary from a set of numbers. These are genuinely new capabilities. But they require clean, structured data as input — and they produce probabilistic outputs that require human review before they touch your books.
Why AI fails in accounting — and it is not the model's fault
The most common reason AI implementations fail in finance is not the quality of the model. It is the quality of the data. A machine learning model trained on inconsistent, incomplete, or incorrectly coded historical data will produce inconsistent, incomplete, and incorrectly coded predictions. Garbage in, garbage out — but with a confidence score attached.
This is why we built the data layer before we built the AI. Lunari's Rules Engine is not a standalone AI product bolted onto an existing accounting workflow. It is a component of a unified data model that captures every transaction at the source, in a consistent format, with a complete audit trail. The AI works because the data is clean. The data is clean because it was never fragmented in the first place.
"AI fails in accounting because the data is broken, not because the models are bad. Fix the data architecture first. The AI will follow."
The audit question every CFO should ask
Before adopting any AI-powered accounting tool, ask this question: if the AI makes a mistake, can I find it? In a well-designed system, every AI-generated coding decision is logged with a confidence score, a rationale, and a human review status. The auditor can see exactly which transactions were coded by the AI, which were reviewed by a human, and which were overridden. The audit trail is complete.
In a poorly designed system, the AI makes decisions silently. The journal entries look correct. The reconciliations balance. But there is no record of how the AI reached its conclusions, and no way to verify that it reached them correctly. This is not a hypothetical risk. It is the reason several high-profile AI accounting implementations have resulted in restatements.
How Lunari's Rules Engine works
Lunari's Rules Engine is a hybrid system. It combines deterministic rules (which your team defines and which are always auditable) with machine learning predictions (which are always flagged as AI-generated and always require a confidence threshold before they post). Every AI decision is logged in the immutable audit trail. Every override is logged. Every exception is queued for human review.
After 30–60 days, the Rules Engine typically handles 70–85% of routine transactions automatically. After 12 months, it knows your specific vendor-to-GL mapping better than any new hire could. But it never posts to the GL without a human having the ability to review and override. The AI accelerates the work. The accountant remains responsible for the result.