Every finance team has the same list: twenty or more places where AI could help. Finding ideas is not the problem. Choosing the first one is. So I wrote down 21 and scored each one the way I would if it were my roadmap.
go first. Two I would not build yet. The other 14 are good ideas that can wait.
Each use case gets a score from 1 to 5 on three questions. The scores are my judgement, not measurements. That is the point of the model below: change the weights and see if the answer survives.
Thousands of invoices a month score high. A decision made four times a year scores low, however clever the AI is.
A matched bank line is right or wrong, and anyone can check it. A forecast is an opinion until the quarter ends.
A wrong draft that a person edits costs minutes. A wrong entry in the ledger costs a restatement.
Set how much each question matters to you. The list reorders as you move. Then try "Volume only": that is the ranking a vendor demo gives you.
A ranking is a guess about where the value is. One month-end close is enough to find out, and it costs far less than a platform decision.
Invoice capture for one business unit, or bank reconciliation for one set of accounts. Not the whole company.
For one close the team works as usual and the system works in parallel. Nobody's month depends on it.
How many items went through with no touch. How many it sent to a person. How many it got wrong that a person would have got right.