Interview with Seel Co-Founder and CEO Zack Peng: AI-Powered E-Commerce After-Sales Service Moves from Answering Questions to Taking Responsibility for Outcomes
Seel Co-Founder and CEO Zack Peng advocates enabling AI agents to handle e-commerce after-sales service across merchant, logistics, and payment systems, while managing risk through clearly defined permissions, human escalation, and accountability for outcomes. AI can accelerate the handling of 80% of support tickets, Seel Agencies has a resolution rate above 90%, and the overall support case resolution rate exceeds 99%.
Body
Zack Peng is Co-Founder and CEO of Seel, where he focuses on the problems shoppers and merchants face after checkout. His work began with predicting returns and refunds and has expanded into AI systems that resolve customer requests across commerce workflows. He advocates pairing agents that can take action with clear operating limits, human escalation, and accountability for the outcome customers receive.
Seel provides post-purchase services and AI agencies for ecommerce, including purchase and delivery protection, customer support, and resale. Its platform connects merchant systems, order information, and service workflows to help resolve shopper issues, while its LLM Gateway supports access to models and task-based routing. In this interview, Peng discusses the shift from answering questions to owning outcomes, the controls needed for agents to act across systems, and how merchants can judge whether AI services are dependable.
What inspired you to co-found Seel, and how did solving the problems shoppers face after checkout lead you toward building an AI workforce for ecommerce?
It started with one marketplace. We met a founder running a marketplace, and saw how much effort the industry put into predicting chargebacks, which cost about 0.5% of GMV. Returns and refunds ran 10 to 20%, and nobody was predicting those. So we built a model that could, and used it to let shoppers pay a small fee to make final-sale items refundable. I expected 3 to 5% to opt in. It was 10 to 20%. That told us shoppers will pay for confidence after checkout, the same reason people buy AppleCare or a car rental waiver.
The AI workforce came from our own support inbox. More than 70% of our support tickets had nothing to do with returns. Shoppers wanted to know where their order was or whether they could exchange something. Instead of sending them back to the merchant, we started resolving those tickets ourselves. We learned that AI’s real value in commerce isn’t saving money. It’s in doing things instantly at scale even during peak volume. This includes having AI systems that can assess and approve eligible refunds before receiving return shipments, as well as having 24/7 AI agents that can take action to resolve customer requests.
You describe a progression from chatbots that answer questions, to agents that take action, to AI agencies that own an entire outcome. What changes technically and operationally between those stages, and can you illustrate that with a real customer issue?
The difference is who owns the outcome. A chatbot answers ‘where’s my order?’ An agent completes a single task, like filing a claim. An agency owns the result: a customer who got what they paid for.
Technically, an agency works across systems, including the merchant’s store, the carrier and payments, rather than inside one. It tracks a case over days, not one conversation, and it has to know when the outcome is actually done. Operationally, you stop staffing for ticket volume and start staffing for exceptions.
Take a package marked delivered that never arrived. A chatbot tells the shopper to wait. An agent opens a claim. Seel Agencies checks the carrier scan and order history, decides eligibility, pays out or ships a replacement, updates the merchant’s records and tells the shopper it’s done. As a partner that takes on end-to-end post-purchase tickets for merchants, we tested this AI infrastructure on our own tickets first. This resolution-driven approach helped us reach a resolution rate north of 90% while being able to flag and refund eligible requests immediately.
Consider a shopping agent that purchases the wrong size on someone’s behalf. What would need to happen between that agent, the merchant, and Seel to resolve the mistake, and who should be accountable for the outcome?
The shopper shouldn’t have to referee it. Regardless of who made the mistake, the fix is the same: a fast exchange or refund.
Three things make that possible. The order needs a record of who placed it and on what instruction, so the agent’s decision can be traced. Second, the merchant’s return policy has to be machine-readable, so an agent can act on it. Third, someone has to stand behind the outcome before fault is sorted out.
That’s where Seel sits. We take on the liability. If the order was protected, the shopper is made whole first and attribution comes after. If an AI shopping agent orders the wrong size on someone’s behalf, the shopper should have the same path to resolution as if they’d ordered it themselves—as long as the right protection was added at checkout and the issue falls within that coverage. We stand financially behind the outcomes we cover. The shopper shouldn’t have to referee whether they or their agent got it wrong.
Your team says AI accelerates 80% of Seel’s support ticket volume. What does “accelerates” mean in practice, how much is resolved without human intervention, and which measures show whether customers actually receive better service?
It means AI does the work that used to slow a person down: reading the case, pulling order and carrier data, checking the merchant’s policy and, in most cases, completing the resolution. AI does that on 80% of our ticket volume. Overall, we resolve 99%-plus of support cases, where the industry usually talks about 60 to 70%.
But closure rate alone can hide bad service, because a bot that deflects shoppers looks great on a dashboard. So we track metrics like ticket reopen rate, escalation rate to a human agent, and customer satisfaction. Speed matters in the moments that frustrate shoppers. Changing a shipping address usually means an email chain across separate order and fulfillment systems. Our agents do it in minutes.
The measure I care about most is whether the shopper buys again. Across our network, 69% of repeat buyers protected their first order.
You’ve described yourself as “pro-agent, just not pro-autopilot.” Which decisions do you deliberately reserve for people, and how does your system recognize when an apparently routine case requires human judgment?
Autopilot means nobody’s watching. We want agents doing most of the work and people making the calls where judgment matters.
We keep three kinds of decisions with people: anything that sets policy, cases where a wrong call is expensive to unwind, and cases unusual enough that the rule’s answer may not be the fair one. The AI handles routine cases within those boundaries and doesn’t get to expand its own authority. If an action hasn’t been authorized, or a case requires additional approval or special handling, it escalates straight to a person.
The system looks beyond the type of request to the context around it: risk, urgency, customer frustration, and policy sensitivity. High-risk cases are routed to people, while policy-sensitive cases can require additional verification before action. Those thresholds are configured for the merchant and workflow rather than expressed as one universal dollar amount or confidence score. An expedited or elevated-payout request, for example, can fall outside the AI’s autonomous authority.
Your shopper research suggests many Gen Z consumers still want a person involved in resolving their issues. What have you learned about designing AI support that earns trust, particularly when customers want to challenge a decision or reach a human?
Gen Z is the generation most comfortable with AI, and also the quickest to notice when a company uses it to avoid them. In our 2026 Gen Z shopping report, 49% of Gen Z said they’d rather resolve an issue with a person. They still want it fast: 87% of shoppers expect an issue resolved within two to three days.
What we’ve learned is that trust comes from the exit, not the bot. When shoppers know they can reach a person, they’re happy to let AI handle things. When AI becomes a wall, trust collapses, even if the answer was right.
So we design for three things:
Speed, so issues are resolved before frustration sets in
A clear reason in plain language whenever the AI says no
And one-step escalation to a person who immediately sees the full history, so the shopper never repeats themselves.
When agents can issue refunds, change orders, or act across merchant systems, mistakes have financial consequences. How do you enforce permissions, prevent duplicate or unauthorized actions, and reconstruct what happened when something goes wrong?
Every agent gets the access its job needs and nothing more. We scope that access to the specific workflows the merchant has agreed to. We use dedicated accounts or API tokens, separate from employee identities, so the merchant can adjust, suspend or revoke access at any time.
For higher-risk actions, like large refunds, we can configure human review or amount- and rule-based thresholds around the merchant’s requirements. Giving an agent access to a system shouldn’t mean giving it an open-ended mandate.
Every system action is logged with when it happened, what was done and what triggered it. If something goes wrong, that gives us a record to investigate with the merchant, assist with rollback and tighten the rules afterwards.
With 1-Click Resale, agents generate listings, set prices, and coordinate sales across multiple marketplaces. How do you validate product descriptions and pricing, and prevent an item from selling twice when marketplace updates are delayed?
Our Resale agency does the work a seller usually wants to avoid. It reviews the shopper’s photo to check the item’s condition, prices it against the market, creates the listing, and delists it once it sells.
Descriptions start from data we already have. Because the item comes from a merchant order, the agency works from the original product record. Pricing starts with what comparable items have actually sold for, adjusted for condition. The shopper confirms the quote before we proceed, and our team handles the listing and any missing details.
The goal is to make reselling easier, so shoppers can recover value from items they no longer want, even after the return window closes
What operational challenges led you to open Seel’s LLM Gateway, and how do you choose models for different tasks based on reliability, latency, and cost? Do you evaluate the economics per model call or per successfully completed customer outcome?
We built it for ourselves first. Developing Seel Agencies to run our own workflows taught us that AI at scale forces you to get disciplined about how you source models. We needed scale, reliability, access to the best models and predictable unit economics, so we built the infrastructure for it. Merchants were feeling the same cost creep, model by model and request by request, so we opened it up.
We route by job. Pulling an order number from an email doesn’t need the same model as working through a complicated, multi-step problem. LLM Router assesses the request’s complexity and routes it within the models the customer has allowed. Customers choose performance, savings or a balance of the two. That lets them decide how aggressively to save on simpler work.
Reliability and latency still have to hold up on the actual workload. We recommend testing quality, response time and cost before making routing more aggressive. For complex workflows or tasks that need strict consistency, customers can keep a specific model rather than use automatic routing.
LLM Router charges for model usage, not per completed customer outcome. With smart routing, you’re billed at the rate of the model that actually runs. Seel covers the cost of assessing the request’s complexity to make that routing decision.
For founders building AI businesses, what has Seel taught you about the difference between an impressive demonstration and a service customers can depend on? What evidence should a merchant demand before handing an entire business function to an AI agency?
A demo shows the system working once. A service shows it working on the ten-thousandth case at 2 a.m., when the carrier’s API is down and the shopper is angry. The hard part was never the model. It’s bad data, systems that disagree, and knowing when the AI shouldn’t act.
Merchants should ask for evidence, not demos. What share of cases does the agency resolve end to end? What are its resolution times and repeat-purchase rates, beyond ticket deflection? When it’s wrong, who pays, how fast, and can you see exactly what it did? And how many people does it take to run?
Thank you for the great interview, readers who wish to learn more about AI agencies, post-purchase services, and automation for ecommerce should visit Seel.
AI Trends