TikTok Customer Service Automation Script Tutorial: 24/7 Auto-Replies for Comments and DMs

1. Customer Service Is a Hidden Cost of Cross-Border E-Commerce

Anyone doing cross-border e-commerce knows a harsh truth: buyers live on the other side of the world, and customer service cannot stay online 24 hours a day. The “customer service” problem of domestic e-commerce is magnified three times over in cross-border scenarios.

Source of hidden cost	What it looks like	Impact on the business
Time zone differences	The active hours of U.S./European buyers are the small hours in China, leaving comments and DMs unanswered	The window of purchase intent passes and users buy from someone else
Multiple stores and platforms	Comments and DMs are scattered across multiple accounts and platform entrances	Messages are missed, and support staffing must double
Response speed	Slow replies make users leave before they commit	Conversion rate drops quickly as response time grows

The third point deserves special attention: response speed is directly tied to conversion. TikTok Vietnam once published data showing that if a brand replies to a customer message within 3 minutes, conversion can increase by up to 2.2x (source) — an official platform figure that makes it clear “replying fast” is itself a form of productivity.

But 24-hour human staffing is more than most small and medium sellers can afford. The answer is therefore straightforward: hand the repetitive work of “seeing messages, judging intent, and crafting replies” to a program, and let humans handle only what the program cannot. That is where TikTok customer service automation starts.

2. Real Cases Rewritten: How Others Answer Around the Clock

Let us first look at two real cases circulating online, then rewrite them into a scenario closer to small and medium sellers.

Case 1: Auto-reply tools let sellers take orders 24 hours a day. The Mixdesk official blog systematically explored how to choose an auto-reply tool, with the headline “helping cross-border e-commerce sellers take orders 24/7” (source), and later broke down how auto-reply tools improve efficiency in the “private-domain traffic loop” (source). The common logic of these tools is simple: a program watches comments and DMs in place of a human, auto-replies when rules match, and the seller only maintains the scripts and rules.

Case 2: The platform itself rolls out an AI customer service bot. In November 2025, TikTok Vietnam launched the AI chatbot “TikTok Lead Genie” for business sellers. It can interact with users directly in DMs and automatically screen potential buyers, providing round-the-clock responses; according to platform data, if a brand replies to a customer message within 3 minutes, conversion can increase by up to 2.2x. Vietnamese beauty group Saigon Star’s brand DrHoat achieved round-the-clock customer acquisition after deployment, cutting lead costs by 80%, saving about 300 hours of manual work per day, and generating over $23,000 in average daily revenue (source).

Case rewrite (a scenario adapted from the public cases above; all figures are illustrative and do not represent measured results of any product): a 5-person cross-border team running 3 TikTok stores focused on the North American market. Time zones were brutal — Beijing midnight is the active period for U.S. users, and DMs plus order comments piled up, only to be seen the next morning. They tried two shifts, but doubled staffing costs and still missed messages. The final solution: a batch of idle Android phones (real devices, no root), one store account per phone, running an automation script, plus an AI API. The script polls the DM and comment lists every 10 seconds, OCR-recognizes new message text, matches intent with keyword rules, calls the AI API to generate an English reply for matched inquiries, and automatically types and sends it. One month after launch (case data): average response time during night hours (Beijing time 00:00–08:00) dropped from about 8 hours to under 1 minute, missed overnight inquiries fell by about 40%, no night staffing was needed for the 3 stores, and only about 1 hour a day was spent handling sensitive messages the AI was unsure about (case data).

The technical pipeline in this rewritten case is exactly what the rest of this article breaks down: poll → OCR recognition → keyword judgment → AI-generated reply → auto-send. Each step is illustrated with the real APIs of EasyClick Android phone automation (no root) (official docs).

3. How a Script Monitors and Replies: A “See → Judge → Reply” Pipeline

An automation script essentially simulates human operations, replacing the human “eyes, brain, and hands” with a program. The whole pipeline has four steps:

Step	Capability used	Corresponding API (EasyClick example)
Poll for new messages	Timed loop / background thread	sleep(), setInterval, thread.execAsync
Read message text	Screenshot + OCR recognition	image.captureFullScreenEx(), ocr.newOcr() / initOcr() / ocrImage()
Judge reply intent	Keyword rule matching	Compare the label field of recognition results against a keyword library
Locate the input box and send button	Node-based UI locating	text() / desc() / id() selectors, waitExistNode(), getOneNodeInfo()
Type and send	Node operations	node.inputText() to type, node.click() to click

Step one, polling. The script refreshes the page on a timed loop, e.g., checking the DM/comment list every 10 seconds; you can also use setInterval or put the polling in a separate thread with thread.execAsync (global module docs).

Step two, OCR recognition. Comment sections and DM lists are ordinary UIs; when a standard node tree is unavailable, “look at the screen” instead: take a screenshot and hand it to an OCR engine to recognize the text. EasyClick ships with the full PPOCR-V4 / V5 / V6 model family, all free and running offline on the device, independent of cloud APIs; each recognition result carries label (text), confidence, and coordinates (OCR docs). V6 requires Android 8.0 or above; on lower versions it falls back to V5.

Step three, keyword judgment. Compare the recognized text against a keyword library: if it hits purchase-intent words like “多少钱”, “怎么买”, “price”, “how to order”, or “shipping”, enter the reply flow; otherwise skip it.

Step four, locate the node and send automatically. Find the input box, type the reply, and click the send button. This uses node-based UI locating: text() / desc() / id() selectors find the control by name, waitExistNode() waits for the element to appear instead of blind sleep, node.inputText() types text, and node.click() clicks (selector & node docs).

Putting the four steps together gives the complete illustrative script below (OCR recognition + keyword judgment + auto-reply):

let ocrIns = null;

// Script startup: initialize the automation environment and OCR (all offline on device)
function init() {
    startEnv();
    image.requestScreenCapture(10000, 0);   // Request screenshot permission
    sleep(1000);
    ocrIns = ocr.newOcr();                   // Create an OCR instance
    // PPOCR-V4/V5/V6 are all free and offline; example uses V5, Android 8.0+ can use V6
    ocrIns.initOcr({"type": "paddleOcrOnnxV5", "numThread": 2, "padding": 50, "maxSideLen": 960});
}

// Recognize screen text: screenshot -> OCR -> return result array (label/confidence/coordinates)
function readScreenText() {
    let img = image.captureFullScreenEx();
    if (!img) return null;
    let results = ocrIns.ocrImage(img, 20 * 1000, {});
    image.recycle(img);
    return results;
}

// Keyword rules: hit "purchase intent" and enter the reply flow
function matchIntent(text) {
    let keywords = ["多少钱", "怎么买", "price", "how to order", "shipping"];
    for (let i = 0; i < keywords.length; i++) {
        if (text.indexOf(keywords[i]) >= 0) return true;
    }
    return false;
}

// Auto reply: find the input box, type the reply, then click send
function autoReply(question) {
    let input = text("输入消息").waitExistNode(5000);  // Input box node; label depends on the actual UI
    if (!input) return;
    let reply = "Hi! Thanks for your message. Please DM us for details~";  // Example reply
    input.inputText(reply);
    sleep(500);
    let sendBtn = text("发送").waitExistNode(3000);    // Send button node
    if (sendBtn) sendBtn.click();
}

// Main loop: poll once every 10 seconds
function main() {
    init();
    while (true) {
        let results = readScreenText();
        if (results) {
            for (let i = 0; i < results.length; i++) {
                let label = results[i].label;
                if (label && matchIntent(label)) {
                    autoReply(label);
                    break;   // Handle only one message per round to avoid spamming
                }
            }
        }
        sleep(10 * 1000);
    }
}
main();

Two details in this example deserve attention. One is deduplication — in real use you must record processed messages and compare before handling on each poll, or you will reply to the same user repeatedly. The other is frequency control — replying to only one message per round structurally prevents spamming. Both will come up again in the compliance section.

4. AI-Generated Replies: Connecting LLM APIs

Keyword rules plus fixed templates handle high-frequency questions like “how much” and “how to ship”, but the moment a user asks something slightly more complex, fixed replies feel stiff. The advanced approach: send the recognized user question to an LLM or translation service over an HTTP API and let it generate a reply in your brand’s voice, then have the script send it automatically.

EasyClick has a built-in network module you can call directly: http.httpGetDefault() for GET requests, http.postJSON() for JSON POSTs, and http.httpPost() for form POSTs, all supporting custom timeouts and headers (HTTP API docs). Example code (API URL and parameters depend on your actual server docs):

// Call an AI API to generate a reply (illustrative: URL and parameters follow your actual server docs)
function generateReply(userMessage) {
    let url = "https://your-ai-server.example.com/v1/chat";  // Replace with your AI API URL
    let params = {
        "message": userMessage,   // The user question (text recognized by OCR)
        "language": "en",         // Target language
        "tone": "friendly"        // Tone style
    };
    // http.postJSON(url, jsonData, timeoutMs, headers) — signature per the http-api docs
    let resp = http.postJSON(url, params, 20 * 1000, {"Content-Type": "application/json"});
    if (resp) {
        // Parse the reply text according to the actual API response structure
        return JSON.parse(resp).reply;
    }
    return "Hi! Thanks for your message. We will get back to you soon~";  // Fallback reply
}

Plugging generateReply() into the autoReply() flow from the previous section completes the “recognize → judge → AI generate → auto-send” loop. The AI API is entirely your choice: LLMs, translation APIs, or self-hosted services, as long as it speaks HTTP. Also worth noting: EasyClick scripts are written in JavaScript and can call all Java libraries, so any SDK you need can be wrapped yourself — plenty of flexibility (product docs).

5. Unified Multi-Account Support: Cloud Control Collects the Data

One account per real device means 3 stores equal 3 phones, and 30 accounts equal 30 phones. Once device counts grow, a new problem appears: how is each device doing? How many replies today? Any stuck devices? Operators cannot watch every screen one by one.

This is where a batch management platform comes in. EasyClick’s cloud control system revolves around four core concepts: device, script, task, and data (cloud control docs):

• Device: the phone executing tasks, each with a unique number;
• Script: the code that automates a task;
• Task: what is to be executed, including the target device group, execution time, cycle, and status; its parameters are pushed to the script together;
• Data: the data produced when scripts run, saved in cloud control and manageable via API for add, delete, update, and query.

The flow is: devices connect to the cloud after numbering → the cloud pushes tasks to devices → the EC client on each device loads and runs the script → the script fetches task parameters via ecloud.getTaskInfo() → runtime data (daily reply counts, keyword hit statistics, AI call counts, error logs, etc.) is reported back and aggregated in real time. Operators view everything in one web dashboard instead of staring at individual screens.

As for scale, the two management tiers can be chosen as needed (central control docs):

Management tier	Deployment location	Management radius	Scale reference
Standalone script	On the device	One device	1 device
Central control	LAN computer	Local batch management with screen monitoring	About 100 devices per machine
Cloud control	Cloud web platform	Any network, remote devices	About 500 devices per machine

A note: once device counts reach double digits, cloud control is worth prioritizing. It is not just about managing more devices — more importantly, it aggregates data scattered across phones into a single view. Which reply scripts convert well, which time slots get the most inquiries, and which keywords hit most often are all visible in the dashboard, giving customer service strategy something to iterate on.

6. Reply Quality and Compliance: Machines Screen, Humans Backstop

Automation boosts efficiency, but auto-reply does not mean letting a bot say whatever it wants. TikTok has its own platform rules around automation and bulk operations; before deploying, it is wise to read through compliance material (see, e.g., the TikTok Compliance FAQ 2026). Combined with industry practice, compliant operations should hold at least three red lines:

1. No non-compliant promises. Reply scripts must not hype efficacy, promise returns, or guarantee absolute results — phrases like “guaranteed viral sales” or “conversion will double” violate platform rules and invite after-sales disputes. Every reply template should pass human review before going live.
2. No spamming or harassment. Control reply frequency: handle one message per round, never disturb the same user repeatedly, and never blast bulk messages.
3. Human review of sensitive content. High-risk messages — price negotiation, complaints, after-sales disputes, medical or financial inquiries — must never be auto-replied; route them to humans instead.

Translated into architecture, it is one sentence: machines do the initial screening, humans provide the backstop. Low-risk, high-frequency inquiries (price, shipping, sizing) get auto-replies; high-risk, low-frequency messages (complaints, disputes, complex questions) get flagged and routed to a human queue. In the script this is easy to design: split the keyword rules into an “auto-reply word list” and a “route-to-human word list”. Hitting the former replies automatically; hitting the latter only logs the message without replying and pins it to the top of the dashboard for attention. The “1 hour a day spent on messages the AI was unsure about” (case data) in the rewrite is exactly this backstop — automation saves repetitive labor, not the human responsibility to judge.

7. FAQ

Q1: Do TikTok comment/DM auto-replies require a rooted phone? A: No. Taking EasyClick as an example, it relies on official system capabilities such as the accessibility service and can take screenshots, run OCR recognition, locate nodes, and type input without root, supporting Android 5.0 up to the latest systems.

Q2: Will auto-replies get my account flagged as non-compliant by the platform? A: It depends on how you use it: control reply frequency so you never spam, avoid non-compliant promises in your scripts, and route sensitive content to human review, all while following platform rules. Automation tools are neutral by nature; compliant usage is the operator responsibility.

Q3: Does OCR recognition cost money? Does it need the internet? A: No cost and no internet needed. The full PPOCR-V4/V5/V6 model family is provided free and runs offline on the device, so recognition does not depend on cloud APIs and the text never leaves the device.

Q4: Do the AI replies write themselves? Do I have to connect a large language model? A: Not necessarily. For a start, keyword rules plus fixed reply templates are enough; for more advanced use, send the user question to an LLM or translation service over HTTP to generate personalized replies. The API is your choice.

Q5: How does the script tell new messages apart from already-replied ones? A: You need to do deduplication yourself: record the text or position of processed messages (in a local file or reported back to cloud control data), compare before processing on each poll, and avoid replying to the same user twice.

Q6: How do you manage multiple accounts and multiple stores in a unified way? A: One real device runs one account. Once device counts grow, use central control (LAN, about 100 devices per machine) or the cloud control platform (any network, about 500 devices per machine) to push scripts, check running status, and aggregate reported data in one place.

Q7: Can it handle English or other non-Chinese users? A: Yes. OCR can recognize English and other languages, replies can be generated in the target language through a translation API or LLM, and the keyword library can maintain both Chinese and English entries.

Q8: Do I need to reinstall the APK when the script is updated? A: No. Scripts support hot updates, so updating the script code takes effect directly, without the hassle of repackaging and reinstalling the APK.

Q9: How many phones can one computer or server manage? A: LAN central control handles about 100 devices per machine; the cloud control platform handles about 500 per machine. The actual ceiling depends on computer performance and network conditions, so leave headroom when planning.

Q10: What is the compliance bottom line for auto-replies? A: Three rules: no non-compliant promises (no hype about efficacy or returns), no spamming or harassment, and human review of sensitive content. The core principle is machines do the initial screening and humans provide the backstop, with high-risk messages always routed to humans.

---

About EasyClick: a mobile automation AI agent platform covering Android (no root), iOS (no jailbreak), and HarmonyOS Next, offering script development, Apple cluster control, local central control, and a cloud control system. → Explore all products