Startups & Funding

Seel resolves 99% of ecommerce support cases with AI agents that act

Seel's AI agencies resolve over 99% of support cases by completing resolutions, not deflecting them. CEO Zack Peng explains the controls, liability model and metrics behind it.

Zack Peng, Co-Founder and CEO of Seel – Interview Series
Zack Peng, Co-Founder and CEO of Seel – Interview SeriesMediamodifier / pixabay
By Sophie Lindqvist7 min read

Why it matters

  • Seel resolves 99%-plus of support cases end to end, versus an industry norm of 60-70% deflection.
  • AI accelerates 80% of Seel's ticket volume; 69% of repeat buyers protected their first order.
  • Returns and refunds run 10-20% of GMV while chargebacks, at ~0.5%, get all the prediction effort.
  • Seel's 2026 Gen Z shopping report found 49% would rather resolve issues with a person, and 87% expect resolution within two to three days.
  • LLM Router bills per model usage, not per outcome; Seel covers the cost of complexity assessment.

By Sophie Lindqvist

Seel resolves 99%-plus of its support cases by having AI complete the resolution — pay the refund, ship the replacement, update the merchant's records — rather than deflect a ticket to a cheaper channel. The industry benchmark companies usually cite sits between 60 and 70%, and it often measures deflection, not resolution. That gap is the core of Zack Peng's argument about where AI in commerce is actually heading.

Peng is Co-Founder and CEO of Seel, a company that started with post-purchase protection and now runs what it calls AI agencies for ecommerce: systems that own entire outcomes across merchant stores, carriers and payment systems. In an interview, he laid out the mechanics, the guardrails and the numbers behind that shift.

The stakes are straightforward. Returns and refunds run 10 to 20% of gross merchandise value for many merchants, while chargebacks — which absorb enormous predictive effort — cost about 0.5% of GMV. Nobody was predicting the bigger number. Seel's answer to that mismatch became a business.

How did a returns-prediction model become an AI workforce?

Seel's origin story starts with one marketplace founder. Peng's team noticed the industry poured effort into predicting chargebacks while ignoring returns. So they built a model to predict returns and refunds, then 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%," Peng said. "That told us shoppers will pay for confidence after checkout, the same reason people buy AppleCare or a car rental waiver."

The AI side came from Seel's own support inbox. More than 70% of its support tickets had nothing to do with returns — shoppers wanted tracking updates or exchanges. Instead of sending them back to the merchant, Seel started resolving them directly, testing its AI infrastructure on its own tickets first. That approach produced a resolution rate north of 90%, with eligible requests flagged and refunded immediately.

"AI's real value in commerce isn't saving money," Peng said. "It's in doing things instantly at scale even during peak volume."

What's the difference between a chatbot, an agent and an agency?

Peng frames it as a question of ownership. "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 rather than inside one. It tracks a case over days, not one conversation, and it has to know when the outcome is actually complete. Operationally, Peng said, merchants stop staffing for ticket volume and start staffing for exceptions.

His example: a package marked delivered that never arrived. A chatbot tells the shopper to wait. An agent opens a claim. Seel's agency 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.

The takeaway is that the value sits in the last steps, not the first one.

Who pays when an AI shopping agent orders the wrong size?

As AI shopping agents buy things on people's behalf, mistakes become a liability question. Peng's position is blunt: "The shopper shouldn't have to referee it."

He lists three prerequisites. The order needs a record of who placed it and on what instruction, so an agent's decision can be traced. The merchant's return policy must be machine-readable, so another agent can act on it. And someone has to stand behind the outcome before fault is sorted out.

That's where Seel positions itself. "We take on the liability," Peng said. If the order was protected, the shopper is made whole first and attribution comes after — provided the right protection was added at checkout and the issue falls within that coverage.

What does 'AI accelerates 80% of tickets' actually mean?

Seel says AI accelerates 80% of its support ticket volume. In practice, that means AI does the work that used to slow a human down: reading the case, pulling order and carrier data, checking the merchant's policy and, in most cases, completing the resolution.

Peng is careful about metrics, because a high closure rate can hide bad service. "A bot that deflects shoppers looks great on a dashboard," he said. Seel tracks ticket reopen rate, escalation rate to human agents, and customer satisfaction. Speed matters too: changing a shipping address, normally an email chain across separate systems, takes Seel's agents minutes.

The number Peng cares about most is repeat purchase. Across Seel's network, 69% of repeat buyers protected their first order — evidence that resolution quality drives revenue, not just cost savings.

Which decisions stay with humans?

Peng describes himself as "pro-agent, just not pro-autopilot." Autopilot means nobody's watching; his model keeps people making the calls where judgment matters.

Three categories stay human: 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 can't expand its own authority — unauthorized actions or cases needing special handling escalate straight to a person.

The system reads context, not just request type: risk, urgency, customer frustration and policy sensitivity. Thresholds are configured per merchant and workflow rather than as one universal dollar amount. An expedited or elevated-payout request, for instance, can fall outside the AI's autonomous authority.

Why does trust depend on an exit rather than accuracy?

Seel's 2026 Gen Z shopping report found that 49% of Gen Z would rather resolve an issue with a person — the generation most comfortable with AI is also the quickest to notice when a company uses it to avoid them. Meanwhile, 87% of shoppers expect an issue resolved within two to three days.

"Trust comes from the exit, not the bot," Peng said. "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."

Seel designs for speed, plain-language reasons whenever the AI says no, and one-step escalation to a person who sees the full history immediately.

How does Seel control agents that touch money?

Every agent gets the access its job needs and nothing more, scoped to the workflows the merchant has agreed to. Seel uses dedicated accounts or API tokens, separate from employee identities, so merchants can adjust, suspend or revoke access at any time.

Higher-risk actions like large refunds can require human review, with amount- and rule-based thresholds configured to the merchant. Every system action is logged — when it happened, what was done, what triggered it — giving Seel and the merchant a record to investigate failures, assist with rollback and tighten rules afterwards.

The company's 1-Click Resale product extends the same pattern. Its agency reviews the shopper's photo to check condition, prices against actual comparable sales, creates listings, and delists items once they sell — with the shopper confirming the quote before anything goes live.

Why did Seel open its LLM Gateway?

Seel built its LLM Gateway for itself first, then opened it after merchants reported the same cost creep, model by model and request by request. The system's LLM Router assesses a request's complexity and routes it among the models the customer has allowed; customers choose performance, savings, or a balance. Seel recommends testing quality, response time and cost before making routing more aggressive, and customers can pin a specific model for tasks that need strict consistency.

Billing is per model usage, not per completed outcome — you pay the rate of the model that actually runs, and Seel covers the cost of the complexity assessment.

What should merchants demand before handing over a business function?

Peng's closing distinction is the sharpest in the interview. "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, he said, was never the model — it's bad data, systems that disagree, and knowing when the AI shouldn't act. His checklist for merchants: end-to-end resolution share, resolution times, repeat-purchase rates, who pays when the system is wrong and how fast, full visibility into what it did, and how many people it takes to run.

For an industry still sold on polished demos, that evidence-first standard may be the most durable idea Peng offers.

Source: unite.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

777 articles

Related articles

  1. Salesforce forecasts $75.8B AI-shopping wave as checkout lags
  2. Mirakl Bets on AI Agents and ChatGPT Enterprise to Rebuild Commerce
  3. Meta Claims Millions Shop With Muse. Executives See 'Hundreds' of Orders
  4. Bloomreach acquires AthenaHQ to connect AI search and personalization
  5. AI Could Soon Set the Price of Your Big Mac and Groceries

« Previous articleNext article »