Judgment vs. Rule: How Shopee and TikTok Shop Automate Refunds

Shopee and TikTok Shop both automate refund decisions, but one trusts AI to judge each case, the other trusts a rule the seller set in advance.

Jason K Hanani

9/14/20266 min read

Every online marketplace claims its refund process "runs on AI" now. That claim alone says almost nothing, because it can mean anything from a simple bot that matches keywords to a canned response, all the way to a system that weighs evidence and decides whether a customer's damaged-item claim is believable. The more useful question isn't whether Shopee and TikTok Shop use AI when a buyer asks for a refund (they both clearly do), it's what each company actually trusts that AI to decide, versus what it only trusts the AI to carry out.

Compare how the two companies actually process a refund and a real pattern emerges. Shopee has built something closer to a judge: its AI looks at an individual case and makes a call. TikTok Shop has built something closer to a rulebook: its AI carries out a decision the seller already made in advance. Both would get described as "AI-powered refund automation" in a press release. They are not the same thing, and understanding the difference reveals a lot more about how each company thinks about risk than either chatbot's tone of voice does.

Two very different companies, one shared problem

Shopee and TikTok Shop are both major forces in Southeast Asian e-commerce, but they didn't get there the same way. Shopee built its business as a traditional marketplace: buyers browse, sellers list products, and transactions happen at massive scale across the region. TikTok Shop grew out of short videos and livestreams, turning entertainment into shopping almost by accident, and has grown fast enough to become a real challenger to Shopee in just a few years.

One more difference matters here, and it shapes everything else in this comparison. Shopee's parent company, Sea Limited, is publicly listed on the New York Stock Exchange, which means it has to report business results to investors every few months, including specific numbers about AI's role in refund processing. TikTok Shop's parent, ByteDance, is privately held and has no such obligation. That difference in what each company has to say publicly turns out to matter almost as much as what their systems actually do.

Shopee: an AI that makes the call

On its most recent earnings call, Sea said it used large language model capabilities specifically to enhance its buyer return-and-refund process, and in the same quarter, resolution times for customer cases in Asia markets improved by more than 40% year-over-year, with almost six out of ten cases resolved within a single day.

That's a different kind of claim than a chatbot answering routine questions faster. Sea credits the LLMs with improving how a case gets resolved, not just with responding to it faster, which points to the AI doing something closer to weighing a case than following a script. What Sea hasn't disclosed is how much of that process still runs through a human reviewing or approving the outcome before a refund is finalized, or what happens to the roughly four in ten cases that don't resolve same-day.

TikTok Shop: an AI that follows the rules.

TikTok Shop's documentation for its sellers is very detailed here, and it paints a genuinely different picture. Sellers can set up automatic rules in advance for returnless refunds. TikTok's own Seller Center guide for Singapore, for instance, walks a seller through configuring exactly which return reasons, product categories, refund amounts, and per-customer limits should trigger an automatic payout, and specifically recommends capping the automatic amount at S$15 to limit financial exposure. Once a rule is active, TikTok's documentation is explicit that the system "will automatically refund the buyer" with no seller action required, and just as explicit that any refund a rule auto-approves "cannot be appealed."

Nothing in TikTok Shop's public documentation for this feature describes a model weighing the specifics of an individual case: a condition is met, and the system simply carries out what was already agreed, permanently. The finality is the guardrail. There's no AI judgment to second-guess and no appeal step, because the seller already made the call when they set the rule, not when the refund actually happened. A request that doesn't match any configured rule (a return reason outside what the seller listed, an amount over the threshold, no rule set up at all) falls outside this automated path entirely and goes through TikTok Shop's standard, human-handled returns process instead.

That's not automation done halfway; it reflects a different philosophy about where risk should sit. The seller sets the policy once, and after that the system's only job is to apply it instantly and correctly, with no guesswork and no chance of a wrong judgment call, because no judgment is being made.

Why the difference makes sense for each company

Neither approach is simply better. They're two reasonable solutions to the same problem, shaped by two very different businesses.

Shopee runs an enormous, high-volume marketplace where refund disputes genuinely vary from case to case: was an item really damaged, is a complaint credible, does the evidence support a refund. Centralizing that judgment in one system the company controls directly is a sensible way to keep outcomes consistent, rather than hoping that millions of individual sellers each handle ambiguous cases fairly and consistently on their own. It's also the kind of investment that only makes sense with serious, ongoing engineering effort behind it, which lines up with everything Shopee has said publicly about where it's investing.

TikTok Shop, by contrast, deals with demand that spikes hard and fast around livestream events and sales campaigns, and it's a platform where far more of the actual customer relationship belongs to the seller, not to TikTok itself. Letting each seller set their own simple refund rule (anything under a certain amount, no return required, for instance) lets a business owner encode their own comfort level immediately, without waiting for a model to learn what "reasonable" means for their specific store, and it does so instantly and predictably every single time.

The bigger lesson

The specific finding here (one company currently trusts judgment for refunds, the other currently trusts rules) only matters directly if you're choosing between these two platforms. It's worth remembering that "currently" is doing real work in that sentence: this is a snapshot of what each company has shipped, not a permanent difference in what either one is capable of building. But the underlying question is useful anywhere automation gets introduced into a business: is this system being trusted to decide something, or only to carry out something that's already been decided? Those are genuinely different problems. A rule that's set up wrong fails the same way every single time, which at least makes it easy to spot and easy to fix. A judgment that's wrong fails unpredictably and is much harder to catch, because understanding why it went wrong means understanding how the system was reasoning in the first place. Before automating any decision (a refund, an escalation, who gets contacted next), it's worth asking not just whether AI is capable of it, but whether the situation actually requires judgment at all, or whether it's really just a rule nobody has bothered to write down yet. Most businesses have more of the second kind than they realize, and the costly mistakes tend to go both directions: building an expensive, judgment-based system for something that was always just a simple rule, or trying to force a genuinely complicated situation through a rigid rule that can't account for its exceptions.

Sources

Jason Kester Hanani

Operations & Product Professional

contact@jasonkhanani.com