AI Support for Marketplaces: Handling Buyer–Seller Disputes

Two customers per ticket

Marketplace support has a shape single-merchant support doesn't: there are two customers on most tickets — the buyer and the seller — and their accounts both live on the marketplace. "Item not as described," "never arrived," "the seller won't respond," "the buyer is abusing returns" — each of these has two sides. An AI support agent for a marketplace has to work within that.

The agent doesn't adjudicate

Deciding who's right in "the item was damaged" versus "the buyer damaged it" is a judgement call with money and reputations attached. That's not the agent's job. Its job is to gather, check, and route — resolve the clear cases and hand the genuine disputes to a human with everything assembled.

What the agent can do

• Collect structured evidence from both sides — photos, tracking, the message history, the return reason.
• Check the claim against marketplace policy and the transaction record.
• Resolve the clear cases — tracking shows delivered to the right address and the buyer confirms receipt; a return within the stated window; a refund the seller has already agreed to.
• Escalate the real disputes to a human with both statements, the evidence, the policy check, and the transaction timeline attached — so a person picks up a complete case, not a fresh investigation.

Identity and permissions

The agent acts for one party at a time. A buyer-side conversation and a seller-side conversation on the same order run separately, each scoped to that party's view. Neither sees the other's private notes or account details.

Grounding

The agent reads the actual order, the actual messages, and the actual policy. It never characterises the other party's position from memory — if the seller's response matters, the agent quotes it from the message record. On our support agent project the pattern is the same one used for single-merchant returns: function calling over the real records, grounding in the real policy content, confidence-based escalation.

The evidence-gathering flow

A buyer opens "item not as described." The agent asks for specifics and photos, structured — what's wrong, which item, when it arrived. It checks the order, the listing text, the return window, and the policy. If it's a clear, in-policy return, it processes it and closes the ticket. If it's contested — the description is arguably accurate, or the seller has grounds — the agent opens a case, notifies the seller with a structured request for their side, and collects their response and evidence. Then it assembles the case and escalates. The human who picks it up reads a complete file, not a starting point.

Why the two-sided model is more work

A single-merchant agent has one customer and one view. A marketplace agent needs per-party identity, per-party views of the same order, per-party notifications, and a case object that a human reviewer can see in full while each party sees only the slice meant for them. That's real integration — it's the main reason a marketplace support build is bigger than a merchant one, and it's worth scoping before you start.

Where this stops being right

• A heavy trust-and-safety operation may want humans on all disputes — the agent then does intake and evidence-gathering only.
• The two-sided identity model is more integration than a single-merchant agent — budget for it.
• Vague dispute policy ("resolved fairly") produces a vague agent. The clearer the written policy, the more the agent can resolve.

FAQ

Can the agent decide a dispute? No. It gathers evidence, checks policy, resolves the clear cases, and escalates the real disputes with a complete package.

Does it talk to both the buyer and the seller? Yes, but as separate conversations, each scoped to that party's view of the order.

What does it need to work well? A clear, written dispute policy and access to the order, the messages, and the evidence. Vague policy limits it to intake, because there's no rule for the agent to apply.

Can it act against a seller — hold a payout, issue a warning? Those are enforcement actions with real consequences for the seller's business. Keep them behind a human decision; the agent gathers the evidence that a person acts on.

ISTRALLEN builds marketplace support agents that assemble disputes for human review and resolve the clear cases directly; see AI for E-commerce.