Use AI to Chase a Supplier for a Late Delivery

To use AI to chase a supplier for a late delivery, write for a different goal than an invoice chase. When a customer owes you money, pressure works, because the relationship is already strained and the outcome you want is a payment. When a supplier owes you goods, pressure is counterproductive: you need the stock, you need it from them, and you will need the next order too. The thing to escalate is specificity, not force.

What follows is a three-stage ladder, a prompt that produces each stage, and the one instruction that stops the model inventing contract terms you never agreed.

Why the invoice playbook fails here

Chasing a late invoice works through consequence. Each message raises the cost of continued silence until paying is easier than ignoring you, and the ladder in using AI to chase late invoices runs on exactly that logic.

Reverse the roles and the logic inverts. You have no consequence to apply that does not hurt you more: cancelling the order leaves you without stock, threatening to leave costs you a supplier relationship built over years, and a legal letter buys a dispute instead of a delivery. Meanwhile the person reading your email usually cannot make the goods arrive any faster. They can only tell you the truth about when they will.

So the goal of each message is information, in this order: a committed date, then the reason, then your options if the date is wrong. Information is also the thing a supplier can give you immediately, which is why this ladder gets replies when a pressure ladder gets silence.

The three-stage ladder

Stage one assumes nothing is wrong. Most late deliveries are an administrative slip, and a short, specific email solves them without anybody being managed. Stage two arrives only after a missed promise, and its job is to find out whether this is a delay or a failure, because those need different plans from you. Stage three is where you stop asking about their problem and start solving yours.

Do not skip to stage three because you are frustrated. Offering a substitution before you know the cause is how you end up accepting a worse product for a delay that was going to resolve itself on Thursday.

Stage one: the date request

Keep it to five lines. The email has one job, which is to produce a date and a tracking number.

Write a short email to a supplier about a late delivery.

Facts, use only these:
- Order: [number], placed [date]
- Promised: [date or lead time as quoted]
- Outstanding: [items and quantities]
- Today: [date]

Rules:
- Ask for a specific shipping date and a tracking reference.
- Neutral and friendly. Assume an administrative slip, not bad faith.
- Do not mention penalties, interest, contract clauses or legal terms.
- Do not state what our contract says. I have not given you the contract.
- Under 120 words. No subject line pleasantries.

That fourth rule carries most of the value, and it is worth understanding why it is phrased as an instruction rather than a prohibition. Anthropic's prompting guidance recommends telling the model what to do instead of what not to do, with the example of replacing "Do not use markdown in your response" with "Your response should be composed of smoothly flowing prose paragraphs." Applied here, "state only the facts I gave you about the order" works better than a list of forbidden topics, so include both: the positive instruction binds, the prohibition catches the rest.

Stage two: after a missed promise

The second message is the one people get wrong, usually by writing a longer version of the first. It needs a new question, not more emphasis.

Write a follow-up email. The supplier promised [date] and it passed
with no shipment and no contact.

Facts, use only these:
- Previous promise: [date], from [who, which channel]
- Still outstanding: [items]
- What it affects: [the job, order or deadline at our end]

Rules:
- Ask two questions: what caused the delay, and what the new
  date is, with how confident they are in it.
- State our deadline as a fact, not a threat.
- Keep the relationship intact. We intend to keep buying from them.
- No legal language, no penalties, no invented contract terms.
- Under 150 words.

Asking how confident they are is the trick worth stealing. A supplier who answers "the 22nd, confirmed with the warehouse" has told you something. One who answers "we are hoping for the 22nd" has told you more, and you can act on it a week earlier than you otherwise would.

The instruction that matters most

Never let the model talk about your contract. It does not have your contract, it has a statistical sense of what supply agreements usually say, and it will produce confident sentences about notice periods, acceptable delays, late-delivery remedies and your right to cancel. Any of those can be wrong, and a wrong assertion about your own terms is worse than silence: you have now put a claim in writing that your supplier can check and you cannot support.

Three habits keep this safe. Give the model only the facts you can see in your own records. Tell it explicitly that it has not been given the contract. And read the draft once for any sentence that asserts an obligation rather than a fact, because that sentence is the one to cut. The same discipline applies whenever you let AI write to a counterparty, which is why negotiating supplier rates with AI leans on your own numbers rather than the model's sense of a fair price.

Stage three: give them a choice you can accept

By now you need a decision more than an explanation. The strongest version of this email offers two or three options that are all acceptable to you, which converts an open-ended wait into a reply you can plan around.

1. Partial shipment now of whatever is in stock, remainder to follow.

2. A named substitute product at the same price, if you have checked one is genuinely usable.

3. Cancellation of the outstanding lines with a credit, if you have an alternative source lined up.

Write those yourself, or at least check each one, because this is the stage where an invented option causes real damage. A model asked to "suggest alternatives" will cheerfully propose a substitution you cannot sell or a credit arrangement the supplier never offers.

What to keep out of AI entirely

Two things. Anything that is going to become a legal dispute, which needs advice rather than drafting. And the decision itself: whether to wait, substitute or cancel is a judgement about your business that depends on things the model cannot see, including what else you can sell that week.

Used for drafting rather than deciding, this saves a genuine twenty minutes per chase, and it makes the messages better than the ones you would write while annoyed. That is the general pattern for AI for small business: it is good at the writing, and it should never hold the pen on the decision. The same split shows up in handling a supplier price increase and in catching supplier billing mistakes, where the arithmetic is yours and the phrasing is the model's.

FAQ

Should I tell the supplier the email was written with AI?

No more than you would disclose using a template. The facts are yours and the judgement is yours. What matters is that every claim in it is true.

How long should I wait between stages?

Tie it to their promise rather than the calendar. Stage two goes out the day after a promised date passes, not a polite week later, because the information is most useful while the warehouse still remembers what happened.

What if they stop replying altogether?

Change channel and change person. Telephone the account manager, and if that fails, write to whoever signs off orders rather than sending a fourth email into the same inbox. Silence across two channels is a sourcing decision, not a chasing problem.

Can I automate the whole ladder?

Automate the drafting and the reminder, not the sending. Every one of these emails contains a fact that changes between stages, and an automatically sent stage two that arrives after the goods did costs you credibility you will want later.