Using AI in accounts

Claude prompts for chasing overdue invoices: what to ask, what to avoid

Ask Claude "who should I chase first" and it will happily invent a ranked list from whatever you paste in — with total confidence, and no way for you to check its arithmetic. That's the one thing on this page you shouldn't do. Everything else below is genuinely useful.

Where it genuinely helps

1. "I've called this customer twice and they keep saying 'next week.' Draft what I should say on the third call to actually get a commitment, not another promise." — good use: it's about the conversation, not the numbers.
2. "This customer disputes the amount on invoice #4471 — they say they were only delivered 80 units, we billed for 100. Draft a message asking for their delivery acknowledgment without accusing them of anything." — good use: de-escalating a real dispute is exactly the kind of tone work a model is decent at.
3. "I have three customers who've promised payment and not delivered before. How should my tone differ for each on the next call?" — good use, if you tell it what happened last time with each one; it can't infer history you haven't given it.
4. "Explain the difference between sending a legal notice and just calling again, for a customer 60 days overdue on ₹1.2L, in terms of what it actually signals to them." — good use: a plain-language explanation of a real decision.

Where it quietly goes wrong

• "Here's my customer list with amounts and dates [pasted from Excel], tell me who to chase first." This is the trap. A model reading a pasted table can misalign a row, miscount days between two dates, or just average things incorrectly — and it will present a wrong answer with the same confidence as a right one. You have no way to spot the error without redoing the calculation yourself, which defeats the point of asking.
• Trusting a "risk score" it invents. If you ask "which of these customers is most likely to default," it will give you an answer, because that's what a language model does — it doesn't mean the answer reflects anything real about that customer's actual payment behaviour. It has no memory of how they've paid you before unless you paste that history in every single time.
• Asking it to track cooldowns or escalation state across conversations. "Did I already chase Ramesh this week?" is not something a chat tool remembers reliably between sessions — it needs actual stored state, not a model's recollection of an earlier message in the same thread.

Why this distinction actually matters

The failure mode isn't "the AI is bad" — it's that ranking who to chase is a job for consistent arithmetic over real data, and drafting a message is a job for judgement about tone. A language model is built for the second, not the first. Ask it to do the first and it will still answer, fluently, and you'll have no signal that anything's wrong until a customer you should have called first quietly goes another month unpaid.

This is the exact reason Collection Plan's ranking, DSO, and dispute detection are plain deterministic math over your Tally export — the same ten lines of arithmetic every time, not a model's best guess at a pasted table. AI shows up in the product exactly once, as an advisory fallback when the parser can't identify a column on its own — and even then, a human clicks "analyse" before anything reaches a number you'd act on.

Next: The full Claude collections playbook · Claude for payment reminders · 25 Claude prompts for accounts teams · Spotting a disputed invoice