How to automate collections with AI: aging, prioritisation and drafted dunning
Most collections teams don't lack effort. They lack a good answer to one question each morning: which ten invoices, and what do I say? AI is good at both halves of that question.
Why the aging report alone doesn't collect cash
The aging report tells you what is late. It doesn't tell you what to do about it. Sorting by the 90+ bucket sends collectors after the oldest balances, which are often disputed, small, or already written off in everyone's head. Meanwhile a large invoice at 35 days from a customer who always pays at 30 slips without anyone noticing.
The useful questions are different:
- Which invoices are behaving differently from this customer's normal payment pattern?
- Which ones are blocked by something we control: a missing PO number, a pricing dispute, a proof of delivery nobody sent?
- Where would one email this week move the most cash before month end?
The daily loop
- Read. Pull open invoices, the last 12 months of payment history per customer, credit limits, and the collections inbox or CRM notes. Read-only access is enough.
- Explain. For each overdue invoice, find the likely reason: dispute in an email thread, short payment against a remittance, promise-to-pay date missed, or no contact at all.
- Rank. Score each invoice on cash impact and likelihood that action this week changes the outcome (see the table below).
- Draft. Write the next touch for the top of the list: email, call script, or an internal note asking sales or operations to clear a blocker.
- Hand over. The collector gets a short worklist with drafts attached. They edit, send, and log the outcome, which feeds tomorrow's run.
How to prioritise: signals that matter
| Signal | Why it matters | What the tool does |
|---|---|---|
| Days late vs this customer's normal | A habitual 45-day payer at 40 days is fine; a 25-day payer at 40 is not | Compares against each customer's median days-to-pay |
| Invoice size | Effort should follow dollars | Weights by balance, capped so one huge invoice doesn't hide the rest |
| Open dispute or short pay | Chasing payment won't work until the dispute is resolved | Routes to the internal owner instead of the customer |
| Broken promise to pay | The strongest predictor of further delay | Moves the invoice up and escalates the tone one step |
| Credit exposure | Total open balance near or over the credit limit | Flags for the credit owner before new orders ship |
Drafted dunning: what good looks like
Good dunning is specific and boring. It names the invoice, the amount, the due date and the one thing needed to get paid. A tone ladder keeps it consistent:
| Stage | Typical timing | Tone and content | From |
|---|---|---|---|
| Reminder | 3 days before due | Friendly, attaches the invoice, confirms remit details | AR |
| First follow-up | 7 days late | Asks for a payment date or the reason it's held | Collector |
| Second follow-up | 21 days late | Lists all open items, asks for a call | Collector, sales cc'd |
| Escalation | 45+ days or broken promise | States consequences under the contract terms | Controller or CFO |
The draft should reference the actual thread. "Following up on your note of 14 October that the invoice was with your approver" gets answered. "This is a reminder that your account is past due" gets ignored.
The controls that make it safe
Draft, never send. The tool writes the email; a collector reads it and presses send. Nothing goes to a customer without a person who owns the relationship seeing it first.
- Every number opens its source. Each amount, date and invoice number in a draft links to the ERP record it was read from, with the time it was read.
- A reviewer pass checks figures. A separate check confirms every invoice cited is still open and every amount matches before the draft reaches the worklist. Nothing is more damaging than chasing an invoice the customer paid yesterday.
- Define DSO once. Pick one DSO formula, write it down, and have every report use it, so collections progress isn't argued over. (See the number book guide.)
- Escalations stay human. Credit holds, payment plans and write-offs are decisions. The tool can recommend; a named person decides.
Common mistakes
- Automating the send. Auto-sent dunning saves a few minutes and costs relationships. The gain is in the drafting and prioritising, not the click.
- Ignoring the internal blockers. A large share of late invoices are late because of something on your side: a wrong PO, a missing document, a price mismatch. Route those internally first.
- Ranking by age only. You'll spend the week on the oldest balances instead of the ones you can still collect this month.
- Keeping a worklist nobody uses. If collectors are still working from their own spreadsheet after a month, find out why and fix it, or retire the tool.
FAQ
Should AI send collection emails automatically?
No. Let AI draft every email and call note, and keep the send with the collector who owns the customer relationship. The time saving is in preparing and prioritising the work, and a person reading each draft prevents chasing an invoice that was already paid or disputed.
What data does an AI collections tool need?
Open invoices and payment history from the ERP or billing system, credit limits, and the collections inbox or CRM notes. Read-only access is enough. The payment history lets it compare each invoice against that customer's normal days-to-pay.
How quickly does AI-assisted collections reduce DSO?
Most teams see movement within one or two month-end cycles, because the first gains come from catching large invoices that have drifted past a customer's normal pattern and from clearing internal blockers. The size of the change depends on how much of the current backlog is genuinely disputed.
Get new playbooks by email
Practical notes on AI in finance from a sitting CFO. Occasional, no selling.
Subscribe