What AI can and can't do in finance today

A lot gets promised about AI in finance and little gets delivered. Half the stories say your controller will be redundant next year, the other half that it's all toys. Neither is true. AI does a few things surprisingly well today, and a few things it really doesn't. Knowing the difference saves you a lot of wasted time and money.

What AI does well today

The strongest use cases combine a lot of text and context with a repeatable pattern. Think of:

• →Receivables: drafting reminders, summarising payment behaviour, prioritising follow-up
• →Month-end close: working through checklists, flagging deviations, writing a first draft of the notes
• →Reporting and board prep: a readable management summary from the figures, with the questions a board asks
• →VAT and ledger checks: reconciling and flagging outliers
• →Cash flow: scenarios and notes around a forecast that you feed

The common thread: AI speeds up drafting, summarising and checking. It is a strong assistant for the first 80 percent of the work, not the last 20.

What AI doesn't do (yet)

AI takes no responsibility. It can present a number with confidence that is wrong, and it doesn't know when it doesn't know. For finance, where a mistake costs money or trust directly, that means a human checks and signs off. This is called human-in-the-loop, and it's not a temporary measure but a design principle.

Judgement also stays human work. Whether a deviation is an error or a signal, whether an investment is wise, whether a customer deserves credit: that requires context the data doesn't hold. AI delivers the material, you make the decision.

Where to start: quick wins versus strategic projects

Start small and concrete. Pick one recurring task that takes a lot of time and carries little risk, for example drafting reminders or a first draft of the month-end notes. A quick win proves the value, builds trust in the team and teaches you where the limits are. Only then come the larger, structural projects where AI gets truly embedded in your processes.

The order matters. Teams that start with one big, ambitious AI project get stuck on data, governance and adoption all at once. Teams that start with a quick win have something working within weeks.

Conditions: governance, the EU AI Act and access to your data

Three things you want in place from the start. One: governance, who may use which AI output and who checks it. Two: the EU AI Act, which sets requirements for AI use that are relevant for finance applications. Three: secure, controlled access to your administration, because AI is only useful once it can reach your real figures, and dangerous if that happens uncontrolled.

And then?

AI in finance is not a switch you flip, but a craft in its own right: a different methodology, different tooling, building instead of advising. That has its own proposition and its own site, with deeper articles per topic, from the month-end close to agents and governance. See the resources on start2scale.ai for the depth.