"AI" in finance software covers three different things. Sorting every proposal into one of them is the first skill a finance leader needs.
1. Predictive AI: scores and forecasts
Statistical and machine-learning models trained on your history estimate what will probably happen: which invoice is a duplicate, which customer will pay late, what cash will look like in six weeks, which journal entry looks unusual. Finance has used this for years in fraud detection, credit scoring and forecasting. It is only as good as the data it learns from.
2. Generative AI: drafts and answers
Large language models (LLMs), the technology behind ChatGPT, Claude, Gemini and Copilot, predict the next words of a text. That makes them good at drafting (variance commentary, board-pack sections, policy summaries), summarizing (contracts, long email threads, accounting standards) and answering questions over documents. They can also be fluent and wrong at the same time.
3. Agentic AI: taking actions
An agent is an LLM connected to tools: it can read an invoice, look up the purchase order, post a draft entry, send a reminder or prepare a payment run, with few or no human steps in between. Agents multiply both value and risk, because a mistake can repeat at machine speed.
How common is it?
Gartner found 58% of finance functions were using AI in 2024, up 21 points in a year. The most common uses were process automation and anomaly detection, not chatbots. And budgets are moving: Gartner's 2026 budget research found nearly 60% of CFOs plan to raise finance AI investment by 10% or more.
A simple test for any AI proposal
- Which kind is it: predict, draft or act?
- Which number moves, by how much, against what baseline?
- What does a human check before it reaches the ledger, a payment or someone outside the company?
If a proposal can't answer the second question, it is a demo, not an investment.