Papers, Please Is a Lesson in Writing Support Bot Rules

I helped a small online shop set up a returns chatbot last year. It sold kitchenware, mostly to home cooks, and around 70 percent of its customer messages were some version of "can I return this?" The owner wrote a list of return rules, we loaded them into the bot, and within a week she was getting angry emails from customers who had been told yes and then no, or no when the answer should have been yes.

The rules were all correct. The problem was the order they were checked in, and what the bot did when two rules disagreed. I recognised the problem because I had spent a few evenings that month playing Papers, Please, a game that is, underneath the bleak setting, a long training course in exactly this. Here is what it teaches about writing rules for a support bot, and what we changed in that shop.

What the game asks you to do

Papers, Please was made by Lucas Pope and released in 2013. You play a border inspector in the fictional country of Arstotzka. Each day a queue of people arrives at your booth with documents, and you stamp each one approved or denied. You earn a small amount for every person you process correctly, and you need that money to pay rent and keep your family fed and warm.

The twist is that the rules change. On the first day, you only let in citizens. The next day, foreigners can enter with a valid passport. Then they need an entry ticket. Then an entry permit, then an ID card, then work passes for workers, and so on. Every few days a new bulletin adds a check. Make a mistake and you get a citation. After a couple of free warnings each day, every further mistake costs you money.

By the second week, players have built a mental checklist, and the order of that checklist decides whether they survive. Good players check the cheap, disqualifying things first. Is the passport expired? Is the issuing city a real one? Does the photo match? Only then do they bother with the detailed documents.

Why order matters for a support bot

A support bot that handles returns, refunds or bookings is doing the same job. It takes a request, checks it against rules, and gives an answer. The shop's rules looked like this in the owner's notes, in the order she wrote them down:

• Items can be returned within 30 days of delivery.
• Items must be unused and in original packaging.
• Knives and sharpening stones cannot be returned once opened, for safety.
• Sale items are final sale.
• Gift orders can be returned for store credit only.
• Orders placed through the wholesale portal follow separate terms.

The bot checked them roughly top to bottom. So a customer who had bought an opened knife, on sale, through the wholesale portal, 12 days ago, would first hear "yes, within 30 days," then get asked about packaging, then be told knives cannot be returned. Three messages in, they felt misled. That is a bad experience even when the final answer is correct.

The inspector who checks the passport photo last has already wasted the customer's time.

The order we settled on

Borrowing from how good Papers, Please players work, we reordered the rules so the bot checks things that end the conversation outright first, and things that need detail later. We also made it collect everything it needed in one message instead of drip feeding questions. The new sequence:

1. Which channel was the order from? Wholesale orders go straight to a human, with no further questions. That is the equivalent of the wrong country's passport. Nothing else matters until it is sorted.
2. Is it a final sale or restricted item? Sale items and opened knives get a clear, polite no, with the reason, before anything else.
3. Is it within the 30 day window? The bot reads the delivery date from the order number, rather than asking the customer to work it out.
4. Was it a gift? If so, it explains store credit up front.
5. Condition and packaging. Asked last, because it only matters once everything else says yes.

That ordering alone cut the "you said yes and then no" complaints almost to zero. Over the first month after the change, the share of chats that needed a human dropped from about 18 percent to about 11 percent, mostly because customers stopped arguing with answers that had seemed to change.

Discrepancies should go to a person

The other thing Papers, Please does well is how it handles conflicting information. In inspection mode, you click on two pieces of evidence that do not match, a name spelled differently on two documents, a weight that does not match the scale, and the game highlights the discrepancy. Then you decide what to do: ask the traveller about it, fingerprint them, or deny.

Bots need the same habit, and often lack it. If a customer says the item arrived three days ago but the tracking shows delivery five weeks ago, the bot should not quietly pick one. It should flag the mismatch and hand over to a person. We added a simple rule: whenever the customer's claim and the order data disagree on dates, item or quantity, the bot says it needs a colleague to check and routes it to the inbox with both values shown.

This is where I disagree with a common piece of advice, that a good bot should "resolve as many tickets as possible without escalation." That metric encourages a bot to paper over discrepancies. I would rather a bot escalate more often and be right when it does answer. Messages that bundle several questions together are especially risky here, a pattern covered in why chatbots struggle with multi part questions.

New rules need a bulletin, not a patch

In the game, every new rule arrives as an official bulletin, and you can see the full current rule book in a drawer at any time. Shops tend to add policies by editing a bot prompt in a hurry. Three months later, nobody knows which rules are live.

We set up a one page rule document with a date at the top of each rule, and the bot's configuration is copied from that document, never edited directly. When the owner added a holiday extension to the return window in December, she added it to the document with an end date, and it came off on time in January. It sounds bureaucratic. It saved her two awkward conversations.

If you are just starting with bots in a small business, how small businesses can use bots sensibly covers which jobs to hand over first. Returns, with clear rules, is usually a decent one.

Reorder your own rules this week

Write down every rule your support bot or canned reply system follows, on one page. Then mark each rule with one of two labels: "ends the conversation" or "needs details." Move every "ends the conversation" rule to the top, so the bot checks them before asking the customer anything else. Finally, add one rule that sends any mismatch between what the customer says and what your records show to a person. It takes about an hour. Then read through ten recent conversations and count how many would have gone differently.