The question underneath "should I use AI for this?"
That's usually the wrong question, because it's framed at the wrong level. AI isn't generally good or bad for a business; a specific step either genuinely needs it or doesn't. The useful question is narrower: does this particular step have an input a fixed rule can't handle, and does getting it wrong occasionally cost little enough to risk?
Three questions that decide it in practice
- Is the input actually messy? A scanned invoice, a free-text enquiry, a handwritten note, these don't fit a clean rule. An order form with the same three fields every time does, and doesn't need AI to process it.
- Would the decision need explaining? If you'd have to justify exactly why a particular outcome happened, that's a sign for a rule, which can always show its logic, or a person, who can explain their reasoning, rather than a judgment call handed entirely to AI.
- What happens if it's wrong, and would you catch it? A low-stakes step that's easy to check tolerates more risk than a high-stakes one that silently gets something wrong. The combination of high cost and low visibility is where a person needs to stay in the loop, whatever else is true.
Running a real example through the three questions
A supplier invoice arrives as a scanned PDF attached to an email. Is the input messy? Yes, a scanned document has no structure a fixed rule can reliably read. Would the decision need explaining? No, extracting an amount and a due date isn't a judgment call, there's nothing afterwards to justify. What happens if it's wrong, and would you catch it? A misread amount would most likely be caught at the approval step before payment went out, so an occasional mistake is low-cost and already checked anyway.
That combination, genuinely messy input, no real judgment involved, and an existing check further down the process, is exactly the shape of task AI suits. Compare that with a different step in the same business: deciding whether to extend a struggling client's payment terms. The input (an account history) is perfectly structured. The decision absolutely needs explaining if it's ever questioned, and getting it wrong, in either direction, is costly and not always obvious straight away. That's the shape of task that stays with a person, regardless of how capable AI becomes in general.
Where AI actually helps
- Reading something unstructured. Pulling the amount, date, or reference number out of a scanned invoice or a messy form.
- Classifying or routing. Sorting an incoming message into the right category before it reaches a person, rather than someone triaging every item by hand.
- Drafting a first pass. A reply, a summary, or a description, written for a person to check and send, not sent automatically.
- Flagging what's unusual. In a batch of otherwise routine items, surfacing the ones that don't match the normal pattern, so a person's attention goes where it's actually needed instead of across everything equally.
Where it doesn't
- A step that's already a clean rule. If the logic can be written as "if this, then that," a rule is more predictable and easier to check than an AI judgment call doing the same job.
- A genuine, high-stakes decision. Anything where getting it wrong is expensive and the mistake wouldn't be obvious until later stays with a person.
- Anything that needs to be explainable after the fact. If the actual reasoning behind a decision matters, not just the outcome, that reasoning needs to come from a rule or a person, not an AI step nobody can fully account for.
- Output nobody's actually reviewing. AI drafting something for review only works if the review is real. Skipping that turns a safeguard into a formality.