In an ordinary three days, a homeware shop I will invent for this example gets 600 chats, and 30 of them ask about the terms of an offer. In a three-day sale the same shop gets 1,500 chats, and 600 of them ask about terms. Every figure in this post is hypothetical, built to show a method, not taken from a client or from any market. Total chats rose 2.5 times. Questions about terms rose 20 times.
The shop answers on WhatsApp and web chat with an AI first reply and two people behind it. Its owner plans the sale by the total: two and a half times the chats, so more hours for the team. That plan is sensible and incomplete. The load that grew twenty times is one particular kind of question, and nobody has written down what the correct answers are.
Total volume hides the shift
I call them terms questions, and I mean five kinds: does the offer apply to this item, can I combine it with a code or points, when does it end, will it arrive by the date I need, and can I return it if I buy on the offer. A customer asks these with the basket open and money about to move. An unclear answer either loses the sale or starts a dispute a week later.
In the example, the ordinary days run at 200 chats a day, so 200 x 3 = 600 chats, and terms questions are 5 percent of them: 0.05 x 600 = 30. The sale days run at 500 a day, so 500 x 3 = 1,500 chats, and terms questions are 40 percent of them: 0.40 x 1,500 = 600. Total chats rise 1,500 / 600 = 2.5 times. Terms questions rise 600 / 30 = 20 times. Everything else, stock, sizes, order status, goes from 570 to 900, which is 600 - 30 and 1,500 - 600.
The proportions are invented and yours will differ. The pattern is what I want you to keep. A sale does not only add conversations, it changes what they contain. Planning by chat count staffs for the 2.5. Planning by question type shows you where the 20 lands.
Where the terms live
In most firms the terms of an offer live in the banner, the launch email and the marketing team's brief. The AI assistant was set up on standing policy: normal returns, normal delivery times, normal prices. So were any temporary staff hired for the season, who were shown the catalogue and the tone of voice but not the fine print. While standing policy and sale terms agree, nothing goes wrong. They disagree most in three places: what is excluded, what combines with what, and what happens to a sale item after it is bought.
This is nobody's failing. Marketing wrote the offer to attract buyers, and a banner does that job well. Customer service was set up to answer, and it answers from what it was given. What is missing is one page that both sides treat as the source. Separately, check the rules that apply to your market on advertised prices and terms; that is a different question from this one.
The Offer Sheet: five lines
The Offer Sheet is my working method, not research. It is one page of five lines, filled in by the person who owns the offer, in words a customer would use.
- In. Exactly which products, services or dates the offer covers.
- Out. What is excluded, named one by one, including anything customers commonly assume is included.
- Combine. Whether it stacks with codes, loyalty points, bundle prices or other offers, yes or no for each.
- Ends. The date and time in the customer's own clock, and what happens if stock runs out first.
- After. The delivery cut-off for the offer, and the return or exchange terms for items bought on it if they differ from standing policy.
Two steps surround the sheet. Before you write it, forecast the question mix: use the chats from your last sale if you have them, or a first guess by line. After you write it, run what I call the Twenty-Question Rehearsal, described next.
A worked example: 600 terms questions, 20 rehearsal questions
Every number in this section is hypothetical and illustrative. The five kinds of terms question become five buckets for the rehearsal. In and Out share one bucket, "does it apply to this item", and the After line feeds two, delivery cut-off and returns on sale items. So five lines give five buckets, and I write four questions for each: 4 x 5 = 20 questions. Each is answered from standing policy alone, first without a sheet, and every wrong or unsupported answer is marked.
| Bucket (hypothetical) | Terms questions in the sale | Wrong in rehearsal, no sheet (of 4) | Expected wrong, no sheet | Expected wrong, with sheet |
|---|---|---|---|---|
| Applies to this item (In, Out) | 180 | 2 | 90 | 0 |
| Combine | 120 | 1 | 30 | 30 |
| Ends | 120 | 1 | 30 | 0 |
| Delivery cut-off (After) | 80 | 1 | 20 | 0 |
| Returns on sale items (After) | 100 | 2 | 50 | 0 |
| Total | 600 | 7 of 20 | 220 | 30 |
The buckets add up: 180 + 120 + 120 + 80 + 100 = 600. Seven answers out of 20 failed, since 2 + 1 + 1 + 1 + 2 = 7, and 7 / 20 = 35 percent. Applied flat to the sale, 0.35 x 600 = 210 bad answers. But the buckets are not the same size, so I weight each bucket's failure rate by how many customers will really ask it: 0.5 x 180 = 90, 0.25 x 120 = 30, 0.25 x 120 = 30, 0.25 x 80 = 20, 0.5 x 100 = 50. The sum is 90 + 30 + 30 + 20 + 50 = 220, which is 220 / 600, about 36.7 percent. The flat figure is 10 short, since 220 - 210 = 10. The reason is that the two buckets that failed twice hold 180 + 100 = 280 of the 600 questions, about 47 percent of the volume, while they supplied only 4 + 4 = 8 of the 20 rehearsal questions, which is 40 percent.
Then I wrote the sheet, filled the gaps it exposed, and ran the same twenty questions again from the sheet alone. One answer in 20 was still wrong, in Combine. The sheet said codes do not combine and said nothing about loyalty points. The hole was in the sheet, not in the assistant. Weighted, that one miss is 1/4 x 120 = 30 expected wrong answers, so the total fell from 220 to 30, a reduction of 220 - 30 = 190. I added the loyalty-points line and retested Combine: all four questions came out right.
A rough staff effect, again invented. Suppose one wrong answer in five becomes a dispute that a person spends 10 minutes on. Without the sheet that is 220 / 5 = 44 disputes, 44 x 10 = 440 minutes, or 7 hours 20 minutes. With the first sheet it is 30 / 5 = 6 disputes, 6 x 10 = 60 minutes, or 1 hour. The difference is 440 - 60 = 380 minutes, 6 hours 20 minutes of staff time across the sale.
I want to be plain about what this is. Twenty questions are a sample used as a planning estimate, not a measurement. Real mixes differ, and four questions per bucket will miss things. Use last year's sale chats if they exist. The point is the method: forecast by type, weight by volume, rehearse before launch. Nothing here is a statistic about any market.
What to do before the banner goes up
- Pick your next offer and ask whoever owns it to fill the five lines on one page.
- Write twenty questions, four per bucket, at least two working days before launch, and answer them from the sheet alone. Put them to the AI and to someone who was not in the briefing.
- Fix the sheet first, then the assistant, then rerun the same twenty.
- Give the sheet an end date and time in the customer's clock, and retire it the moment the offer ends, so a finished offer is not quoted back to a customer.
- Rerun the rehearsal whenever the offer changes, including a late extension or a new exclusion.
The late-November sale season is close, and this is a good week to do it. Before your next offer goes live, write its five lines on one page, put twenty questions to the sheet, and fix the sheet before you fix anything else. If you want a second pair of eyes on your Offer Sheet, write to me.
