Your finance team is ready for AI when at least one of its processes passes a short test. The data is reachable. The work repeats. The rules fit on a page. One person owns it. And a mistake gets caught before it costs money.
Below are the 10 questions we use to run that test, how to score them, and what to do with the result. It takes about 10 minutes, and you don't need to give anyone your email to use it.
"Is my team ready for AI?" is the wrong question
Most CFOs we talk to stall at the same point. The question feels too big, so the answer becomes "we need a data strategy first" or "let's wait until the ERP upgrade." Months go by.
A team is never ready for AI in general. A process is. Accounts payable can be ready while month-end close is not. So before you assess anything, pick one process to assess.
Good candidates to start with (for a wider view, see our map of where AI helps a non-tech company):
- AP invoice matching against purchase orders and receipts
- Bank reconciliations
- Month-end accruals and recurring journal entries
- Collections follow-up on overdue invoices
- Expense report review against policy
Pick the one your team complains about most, write its name at the top of a page, and answer the 10 questions below for that process only.
The 10 questions
Answer each one yes or no. If you have to think hard about it, the answer is no. Two of them are knockouts; more on those in the scoring.
Data: can the AI get to what it needs?
1. Could you pull the last three months of this process's inputs in one afternoon? Invoices, statements, emails, exports. It doesn't need to be clean. It needs to be reachable without a two-week IT request.
2. Do you have examples of correct outputs? Approved journal entries, finished reconciliations, invoices that were matched and paid. These are what you check the AI's work against.
3. Do you know whether this data is allowed to leave your servers? Some finance data can go to a cloud AI provider. Some can't, because of contracts, regulation or your own policy. If you don't know which yours is, find out first. The answer decides how the AI gets set up.
Repetition: is there enough volume to matter?
4. Does the process run at least weekly, or dozens of times a month? A task done four times a year rarely pays back the setup.
5. Do most cases look alike? If most invoices, matches or entries follow the same pattern and exceptions are the minority, AI can take the routine part and leave the rest to people.
Rules: could someone new follow it?
6. Could you write the rules on one page so a new hire could follow them? If the process lives only in one person's head, document it first. That alone is often worth doing.
7. When an exception comes up, is it clear who decides? AI should hand exceptions to a person. That only works if everyone knows which person.
Owner: is someone accountable?
8. Knockout: is there one named person who owns this process end to end, and can they give it two to three hours a week during a pilot? Someone in finance who knows the work and will say whether the output is right.
Cost of error: what happens when it gets one wrong?
9. Knockout: is there a review step where someone would catch a wrong output before it leaves the building? A person approving the entry, the payment run or the email before it goes out.
10. Is a mistake reversible? A wrong journal entry can be reversed. A wrong wire transfer or a wrong regulatory filing is much harder to undo. Start where mistakes are cheap.
How to score it
Check the knockouts first. If question 8 or 9 is a no, stop: the process isn't ready, whatever the total. A process with no owner, or one where nobody would catch a wrong output, is how AI pilots fail quietly.
If both knockouts are a yes, count your yeses:
| Score | What it means | What to do |
|---|---|---|
| 8–10 | This process is ready | Start here. Make it your first AI process and measure it against how it runs today. |
| 5–7 | Close, with gaps | Fix the no's first. They're usually about rules or ownership, and you can fix them without buying anything. |
| 0–4 | Not this one | Pick a different process and run the check again. |
If every process you try scores low, that is useful too. It tells you the first project is documenting and owning your processes, before any AI.
An example: accounts payable at a mid-size distributor
Say a distributor with three warehouses processes a few hundred supplier invoices a month. Each one is matched by hand against a purchase order and a receiving report in an ERP that is more than ten years old.
| # | Question | Answer | Why |
|---|---|---|---|
| 1 | Inputs reachable | Yes | Invoices arrive as PDFs in a shared inbox; POs export from the ERP |
| 2 | Correct outputs | Yes | Years of matched and paid invoices |
| 3 | Knows where data can go | Yes | Supplier contracts keep pricing data in-house |
| 4 | Weekly volume | Yes | Daily |
| 5 | Cases alike | Yes | Most invoices match cleanly; the rest have price or quantity differences |
| 6 | Rules on one page | No | Tolerance rules live in the AP lead's head |
| 7 | Who decides exceptions | Yes | The AP lead |
| 8 | Owner (knockout) | Yes | The AP lead, with the controller's backing |
| 9 | Review step (knockout) | Yes | Payment runs are approved weekly |
| 10 | Reversible | Yes | A wrong match can be undone before the payment run |
Score: 9 of 10, no knockouts. Ready, with two things to settle first. The tolerance rules get written down, which takes an afternoon with the AP lead.
And because pricing data stays in-house, the setup has to keep it there. Neither one is a reason to wait.
What teams usually get wrong
- They start with the most painful process instead of the most ready one. The painful one usually has no clear rules or owner, so it fails this check. Win on a ready process first, then take on the hard one.
- They assume the data has to be clean first. Modern AI reads PDFs, emails and messy spreadsheets well. What matters is whether you can reach the data.
- They buy a tool before naming an owner. Software doesn't own a process. If nobody in finance is accountable for the output, the pilot ends when the novelty does.
What to do next
If you scored a process at 8 or above, you have a first AI project. If you landed in the middle, you have a short list of things to fix, and you can fix them in-house.
If you'd like a second opinion on which process to start with, we're happy to talk it through. It's 30 minutes on one process, and you'll leave knowing whether you need outside help at all.
If you're weighing what a first engagement should deliver, we wrote about that here: What a first AI consulting engagement should actually deliver.
Frequently asked questions
How long does an AI readiness assessment take?
This self-check takes about 10 minutes per process. A proper assessment, where someone maps the process, looks at the real data and tests a few cases, takes about two weeks.
Do we need to replace our ERP before using AI?
Not necessarily. AI can often work on top of an older ERP, reading its exports and writing back through the same steps a person would. Whether that's the right call depends on the ERP and the process.
What is a good first finance process for AI?
One that scores 8 or more on this check. In practice that's often invoice matching, bank reconciliations or collections follow-up: high volume, clear rules, and a person already reviewing the output.
Does AI replace people on the finance team?
In a well-run first project, no. It takes the repetitive part of a process so the team spends its time on exceptions, analysis and judgment calls.
