At three in the afternoon on October 2, a row of cars sat parked on the hard shoulder of a highway, people eating oranges by the roadside.
At the same moment, a guesthouse owner's phone buzzed a third time in his pocket. He did not pick up — the queue had already reached the ticket barrier. Two hours later, when he finally pulled the phone out, seven unanswered enquiries were sitting in his WeChat list: two asking about early check-in, two about parking, three asking "can we still cancel for tonight".
He answered five of the seven. The other two had already booked somewhere else.
1. A counter-intuitive truth: holiday enquiry volume actually falls
I assumed the holidays were customer service's Spring Festival travel rush, too. Then I saw an e-commerce operator dig out his own numbers from last National Day, and the conclusion ran against intuition: during the holiday, enquiry volume runs 30 to 40 percent below normal.
But he had a second half to that sentence, and it matters more — the distribution shifted: morning enquiries fell off noticeably (everyone sleeps in), while afternoons and evenings became more concentrated, and late-night volume was higher than usual, because everyone is up late scrolling on their phones.
So the pressure of holiday service was never about total volume. It lands when staffing is thinnest, at exactly the moment customers most need one definite answer.
Ask "when will this ship" before the holiday and he can wait. Ask at ten at night on October 2, standing at a hotel front desk, "can my booking really be moved to the 4th" — he cannot wait, he checks out tomorrow. Questions that could be parked on an ordinary day all turn into "I need the answer now" during a holiday.
2. What actually breaks service is not being busy — it is having nobody to ask
The second half of that operator's post is the most valuable part of it. He made a one-page "holiday policy card" for the on-duty agents, spelling out five things:
- Holiday shipping: which days ship normally, and the daily cut-off time
- Estimated delivery by region: re-estimated on holiday logistics pace, not the usual numbers
- After-sales turnaround: how long refunds take to review, how long replacements take to arrange
- Service hours: which hours have humans, and how to leave a message outside them
- Escalation rules: what must be reported, and to whom
Why go to the trouble of writing that card? In his own words:
Holiday shifts are usually thin, and there is no supervisor nearby. When an agent hits a question and has nobody to ask, they estimate an answer themselves.
That phrase — "estimate an answer themselves" — is the single source of holiday service failures. Normally, when an agent is unsure, they turn their head and ask a supervisor; three seconds, done. During a holiday the supervisor is at his parents' place, at a scenic spot, at the dinner table, and a WeChat reply takes two hours. In those two hours the agent can only guess — guess "refunds take about three days", guess "it should be possible to change the date", guess "we cover the shipping".
The cost of a wrong guess is not settled that day. It accumulates, and on October 8 it all turns into bad reviews, refund disputes and complaints.
3. Why holidays are structurally understaffed: triple pay is an arithmetic problem
The Ministry of Human Resources and Social Security is explicit: when an employee works on a statutory holiday, the employer must pay, on top of normal wages, no less than 300% of the employee's daily or hourly wage. This year the National Day statutory holidays are October 1 to 3.
In other words, putting one agent on shift during the holiday costs close to four times the normal rate. For a small shop with two or three people, there is basically one answer: schedule the fewest people, cover the most hours.
And the enquiries a holiday brings are extremely repetitive. An article on cultural and tourism scenarios in the Alibaba Cloud developer community gave a number: during holiday peaks, more than 60% to 80% of enquiries are repeat questions — opening hours, parking, whether pets are allowed, change and cancellation rules, how to get an invoice.
That creates an awkward structure: the most expensive labour is spent on the most repetitive questions, while the 20% that genuinely needs human judgement gets no time to think.
4. What it costs: a set of complaint numbers
In June this year, 21st Century Business Herald exclusively obtained an industry report from the consumer protection platform Xiaofeibao: in the first five months, online travel platforms nationwide received 19,029 complaints involving more than 30.67 million yuan, and only 47.01% of valid complaints were resolved.
Two cases in the report are especially typical: one consumer cancelled a hotel booking five months in advance and was charged a penalty of half the amount by the platform; another, at the airport, watched the free rebooking window pass because the platform was slow to process the change, losing more than 20,000 yuan.
Zoom out. The China Consumers Association's analysis of complaints received by consumer associations nationwide in the third quarter of 2025 shows 536,761 complaints received, up 7.9% year on year, and among the eight hotspot categories, "a hot outbound travel market with service shortfalls at related travel platforms" is explicitly listed.
Almost all of these disputes start with one small thing: an inaccurate answer, or an answer that came too late.
5. Turn the policy card into decidable Q&A (code included)
The "policy card" above can go one step further and become decision logic that runs inside a system: if an enquiry hits a policy, answer from the policy; if it is not in the policy, hand it to a human. The code below runs as-is.
import re
POLICY = {
"shipping": "Oct 1-3 ships normally, daily cut-off 16:00; normal schedule resumes Oct 4.",
"transit": "Holiday logistics run 1-2 days slower; East China 2 days, South China 3, Northwest 4 - all estimated at holiday pace.",
"refund": "Refund requests are reviewed the same day; arrival depends on the payment channel, usually 1-3 business days.",
"exchange": "Exchanges are logged during the holiday and processed from Oct 8 in order of registration, within 3 business days.",
"hours": "Human agents are on duty 10:00 to 01:00 the next day.",
"escalate": "Anything above 500 yuan, or any explicit request to complain, goes to a human supervisor.",
"invoice": "E-invoices can be issued during the holiday; paper invoices are mailed after Oct 8.",
}
ROUTES = [
("complain|your manager|too expensive|compensation|refund me", "escalate"),
("refund|money back|when will I get", "refund"),
("exchange|swap|different color", "exchange"),
("invoice|receipt", "invoice"),
("anyone there|what time|are you open|opening hours", "hours"),
("when will it ship|how many days|shipping|dispatch", "shipping"),
("how long|transit|when will it arrive|when do I get it", "transit"),
]
def answer(q):
for pat, key in ROUTES:
if re.search(pat, q, re.I):
return "answer", POLICY[key]
return "human", NoneRunning it over the ten most common holiday questions (including three the card does not cover):
Hit policy 7/10 = 70%; escalated to a human 3/10 = 30%
Answered outside the policy (made up) = 0
The three escalated questions are "what is your boss's phone number", "can you give me a discount" and "can you deliver it to a different address" — none of them on the card, and the system improvised on none of them.
There is one pitfall worth spelling out. In the first version I had the routes as an unordered dict, and "I want a refund, how long will it take" got answered with the transit time — the answer itself was correct, but it answered a different question. That is the code-level cause of answering the wrong question. Once the intent words (refund / exchange / escalate) were ordered before the time words, wrong-question answers went from 1 to 0.
6. Three things you can do before the holiday ends
No waiting for a system, no waiting for a budget. Over the last few days of the holiday, three things are enough:
First, write a one-page policy card. One page: how shipping runs over the holiday, how many days it takes, how long refunds take to review, which hours have humans, and what gets escalated to whom. The standard is simple — the person on duty must not have to guess.
Second, list the ten questions you get asked every single day. Once you write them down you will find that most enquiries fall inside those ten. Those ten are your first knowledge base.
Third, set one escalation rule. State clearly which questions the AI or the on-duty agent is not allowed to answer themselves and must pass to a human. Handing off something you cannot answer beats answering it wrong a hundred times over.
Do those three things and you will notice they are really one thing: moving the policy out of people's heads and onto paper and into systems.
In closing
We do enterprise knowledge bases and AI customer service deployment, and the scenario we run into most often in a year is not "the customer asked something impossibly tricky" but one question, three people in the company, three different answers. The holiday just magnifies that normally invisible problem sevenfold.
What makes AI customer service genuinely valuable is not that it can answer, but that it will not guess — it answers only from the policy you have written down, and hands anything outside that policy to a human.
The most expensive thing during a holiday is not triple pay. It is the sentence "I figure it's probably".
Who we are: we deliver enterprise knowledge bases and AI customer service for small and micro businesses; on method we promise only a verifiable process — whether the policy is written down, whether the fabrication rate can be driven to zero, can all be run and seen on the spot.