How TikTok Shop cut ticket resolution time by 150%

TikTok Shop may sound like just another online marketplace – it is anything but. 
Unlike platforms like Amazon, TikTok Shop is far more complex with many more variables. It doesn't have full control of its inventory or logistics, and commerce activity is driven by multiple stakeholders, including merchants, customers, creators and affiliates. 

For sellers, the reach is unrivalled; the discovery-based ecommerce ecosystem promises to put brands in front of millions of potential customers. The buying experience is unprecedented, with impulse, rather than intent-driven, buyers browsing listings every day. 

As a result, the stats are eyewatering: the platform has more than one million sellers working alongside creators, affiliates, and logistics partners. But there are up to 50,000 support tickets a week and at this scale, complex edge cases are incredibly difficult to handle with efficiency.  

This article looks at the key insights shared by Sujit Mohanty, senior CX program manager for TikTok during All Access: Future Contact Centers 2026. Speaking to senior event producer, Chloe Chappell, Mohanty explains how he created and implemented an AI-powered workflow and put it through rigorous experimentation to find the right solution for each unique support request and reduce resolution times by 150 percent in some categories. The full interview can be viewed on CX+

Don't miss any news, updates or insider tips from CX Network by getting them delivered to your inbox. Sign up to our newsletter and join our community of experts. 

Learn More

How big is TikTok Shop?

TikTok Shop launched as a pilot in Indonesia in 2020. It was then introduced to the UK market in 2021 with a US launch in 2023 and subsequent moves into the EMEA and LATAM markets. Growth was rapid. At launch, revenue was less than US$2 billion and in 2026 is trending toward $25 billion, representing a year-on-year increase of more than 500 percent. 

But runaway growth does not automatically lead to sustainable success. During its growth phase, TikTok Shop was publishing and updating 10 to 15 new policies a month on top of constant product and feature releases. Agent training couldn't keep pace, and the knowledge scattered across multiple documents made it hard for agents to find the right answer quickly. Compounding it all, normal contact-center turnover was eroding institutional knowledge.

"All of those things created a lot of confusion also for merchants and creators on how to deal with all of this information," says Mohanty. "We had to create good knowledge bases for our contact centers to support merchants and creators who were dealing with issues. We had to tackle how to keep them structured, how to do version control, which information is the most relevant, and the ownership of who deals with the contacts."

Despite receiving 40,000-50,000 support tickets a week, volume wasn't the primary issue. It was the long tail: the many complex edge cases where a support agent couldn't resolve an issue alone and had to hand off to a seller or merchant for a decision. Merchants, busy running their own businesses, often took time to respond. The ticket would then bounce back to the agent to act on that response.

This multi-step relay dragged the average resolution time down and saw some tickets taking up to eight days to resolve.

The fix: Agent co-pilot and selective automation 

Rather than trying to automate the edge cases away, Mohanty's team built a new, AI-powered workflow based on "a lot of agent co-pilots", which exponentially reduced decision times. The technology pulls the relevant, up-to-date policy information and combines it with the specific context of the ticket, so agents aren't searching a document repository by keyword, while under pressure to work at speed.

Mohanty says there are two factors that make the system work:

A living knowledge base: The system is built to retrieve only current information due to strict version control.

Multi-modal context: The co-pilot doesn't just read ticket text. It can assess images and video submitted by merchants and affiliates, extracting context that even well-trained agents can miss. "Even with proper training, you cannot cover all the edge cases with agents, especially when it's media files, videos, and images," Mohanty says. 

Decision quality was also inconsistent from agent to agent; feeding everyone the same extracted context closed that gap.

For low-stakes decisions, the co-pilots give agents everything they need to resolve a case immediately. For higher-risk decisions, it still speeds things up, but the agent retains the call on whether to escalate or hand off.

The framework: One-way doors vs. two-way doors

While it sounds like the only solution when dealing with huge ticket volumes, full automation would not have solved TikTok's problems. Instead, Mohanty built a simple two-route decision tree for determining when automation should and should not be applied. "It's all about classifying what is a risky decision, what is a non-risky decision, and then making a call based on that," Mohanty says. 

Two-way doors: These tickets are prime candidates for full automation, as they resolve fast and reduce contact volume. They are low-risk, reversible decisions that do not need human review every time. They're the small refunds, replacement orders and anything where a mistake is cheap to correct and doesn't meaningfully damage trust or the business. "The benefits outweigh the cost," as Mohanty puts it. 

One-way doors: These are the decisions with real financial or trust consequences, where getting it wrong is expensive to undo, for example account suspensions. If a model misreads a submitted document or image and freezes a merchant's funds incorrectly, that seller can be locked out of their business for weeks. Fully automating that decision almost guarantees repeat contacts as the merchant keeps escalating until a human intervenes. For these cases, the answer is agent co-pilot combined with human judgment. That means AI surfaces the context, but a trained person makes the call and can override the machine when needed.

The unglamorous foundation: Knowledge base quality

Triaging the support tickets is only part of the job. Without accurate and up-to-date knowledge to draw on, no level of automation will work.

Although Mohanty describes this as "the least glamorous part of the project", it is essential to success and preventing the hallucinations that can erode trust in the entire system. 

"AI is all good, it speeds up everything," Mohanty says. "But eventually it relies on the foundational data. We know that having a quality knowledge base is a major part of making AI-assisted support effective."

In Mohanty's view, to be ready for AI-assisted support, a good knowledge based must be:  

• Authoritative: This demands a single, current source of truth for each policy, not competing versions across teams.
• Version-controlled: This prevents models retrieving outdated guidance.
• Structured and consistent: LLMs work best with uniform, markdown-friendly documents, not freeform text scattered across formats.

For TikTok Shop, getting there was a significant cross-team effort, aligning different departments on a shared format for policies, educational documents, and solutions, and reformatting any documents that didn't fit the model. 

It also required real iteration on prompt engineering: testing different instructions and example cases until outputs were deterministic close to 99 percent of the time. Hallucination isn't tolerable in decisions that touch merchant funds or account status.

"It was a big organizational effort to align different departments and make sure that knowledge base was formatted and consistent across different departments," Mohanty says.

The final piece of the puzzle is a consistent feedback model. Mohanty and his team monitor all contact center tickets and edge cases, particularly the one percent exceptions, then continuously adjust the prompts to ensure the model is performing in line with current requirements. 

The bottom line: 150% reduction in resolution time and improved institutional knowledge

By deploying the AI-powered agent co-pilots and using multi-modal AI models to assess images and video submitted by sellers, resolution times for TikTok Shop service tickets were reduced by 150 percent in some categories.

Additionally, the issue of institutional knowledge erosion was solved and the experience of human agents working in the service suite improved dramatically. 
"When you see such a drastic improvement, it helps align everyone to make changes much faster," Mohanty says.

Watch the full interview with Sujit Mohanty on CX+

Quick links

• AI to augment human roles and elevate experience: Practitioner insights from All Access Future Contact Centers 2026
  
• 5 Lessons on contact center AI from the companies developing it
• What happened at All Access: The AI Revolution in CX

CCW EU 2026

CCW Europe will explore how service is transforming from execution to a strategic driver of trust, insight, and value.

The agenda features over 250 world-class speakers, delivering actionable strategies, global best practices, and inspiring insights on AI, innovation, and customer-centric transformation. From the Failures & Fixes stage to international benchmarks, each session equips leaders to navigate this fast-paced environment and drive meaningful change.

Learn More