Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- A digital-goods marketplace's AI customer service results on Intercom at a glance
- What does the marketplace do?
- Which helpdesk does the marketplace use?
- How did the marketplace train their AI customer service agent?
- When did the marketplace decide to turn on 'direct replies' to customers?
- What was the biggest thing the marketplace did to improve their AI agent's resolution?
- How does the marketplace customize their AI agent setup to work for their business?
- Shaping how the agent talks and when it asks
- Handing over the moment a customer is upset
- What impact is the marketplace's AI customer service agent having now?
- Where does the marketplace go from here?
Created time
Sep 10, 2026 08:10 PM
Title length (<60)
Author
Last optimised
Ecomm?
Image
Publish date
Sep 10, 2026
Video
Slug
digital-goods-marketplace-ai-customer-support-case-study
Featured
Featured
Type
Case study
Ready to Publish
Ready to Publish
A digital-goods marketplace resolves 58% of 36,413 Intercom conversations with AI over the last 30 days, holding 92% AI CSAT and saving 1,768 hours of agent time.
A digital-goods marketplace runs its customer support on Intercom, and our AI agent answers there directly, resolving 58% of conversations without a human and saving 1,768 hours of agent time. Almost every ticket is somebody's money or somebody's item in transit, so the answer turns on which trade and which payout the customer means, and the agent's first job is to find out.
A payout has not landed, or a trade canceled after the items were already sent. The customer usually arrives convinced something has gone wrong, and often they are right.
I have read plenty of AI resolution rates that were earned on password resets and where-is-my-order. Tickets like that barely exist in this inbox.
What follows is what they connected, what they wrote, and where a human still steps in.
A digital-goods marketplace's AI customer service results on Intercom at a glance
Fact | A digital-goods marketplace |
Industry | Digital-goods marketplace, consumer, global |
Helpdesk | Intercom (direct replies) |
Support volume | 36,413 conversations (30 days) |
AI resolution rate | 58% |
AI CSAT | 92% |
Time saved | 1,768 hours |
Escalation rate | 10% |
Key features used | Tasks, Guidance, Custom Answers, human handover |
Knowledge sources | Intercom Help Center, uploaded internal documents, Custom Answers |
Go-live mode | Direct replies inside Intercom |
What does the marketplace do?
Players buy and sell digital goods for real money, mostly in-game cosmetic items known as skins. Every completed sale moves two things at once, an item to a buyer and a payout to a seller.
That gives the team two support jobs inside one inbox, and our agent has to tell them apart on the first message.
Buyers ask where their item is and why a payment failed. Sellers ask where their payout is, what the fee was, and why verification blocked them.
The support team sits in the 20,000+ conversations a month band.
Which helpdesk does the marketplace use?
Intercom. Our AI agent runs inside that same inbox, on the same conversations their human agents see.
They keep their Intercom setup, their agents, their tags and their routing, so nothing about the way they work had to move.
Every conversation the agent handles comes through Intercom. If this team ever moves helpdesk, the trained agent moves with them (the Custom Answers, the Guidance rules and the Tasks all travel too).
Pricing models pull apart at this kind of volume. Paying for every conversation and paying only for the ones the AI resolves are two different bills at 20,000 a month, and both sit on top of the Intercom seats this marketplace keeps paying for either way.
How did the marketplace train their AI customer service agent?
Three kinds of knowledge, added in this order.
First, the Intercom Help Center connector. Their existing help center syncs straight in, and it covers the questions a help center is good at, such as how selling works and what the fees are.
Second, their internal operational documents, uploaded directly as files. These are the procedures that never make it into a public help center, and in an inbox like this one they carry more weight than the public articles do.
Third, Custom Answers, which are exact question-and-answer pairs our AI returns word for word. On a marketplace where money and fraud are involved, plenty of answers are policy, and a paraphrase of a policy is a support ticket of its own.
If you are the person who would have to build this, how much of it is work you do yourself?
Less than it sounds. Connecting a help center is a switch (ours syncs the articles in and keeps them current), and the documents are an upload.
The exact answers and the rules our AI follows are the real work. Somebody on the team wrote them, and worded them well enough to cover the payout policy.
A public help center on its own is thin cover for a support load this varied, which is why the written answers carry most of the weight.
When did the marketplace decide to turn on 'direct replies' to customers?
From the start. They skipped the internal-note stage, where our AI writes its answers as internal notes and a human decides what the customer sees, and had it answer customers directly inside Intercom.
When a conversation needs a person, the agent hands it over and steps back.
The AI Copilot Chrome Extension is not switched on for this account, so the 58% is a number our AI earned on its own conversations.
What was the biggest thing the marketplace did to improve their AI agent's resolution?
They built a named Task for each of the ticket types that arrive most.
A Task is a multi-step workflow written in plain English. Our AI picks the right one for the conversation in front of it, asks the questions that Task needs answered, and works through it with the customer.
Twelve live Tasks cover at least 68% of the tickets that arrived in the last 30 days. Anyone who has run a marketplace will recognize the types: a payout that has not landed, a withdrawal that will not go through, a trade that shows as canceled after the items were sent, a verification check that blocked an account.
Each of those needs its own questions answered before a reply is possible (which trade, at what time, in what state).
The trade-swap case is the highest-stakes ticket the team gets. A customer has sent their items and the trade then shows as canceled, and on a ticket where somebody's items are already gone, our agent asks those questions before it answers.
These Tasks scope a conversation, work out which case it is, and give the customer the correct answer. Where the resolution needs a human action, the Task collects everything needed and hands over.
At least 68% is the share of tickets these twelve Tasks cover. The 58% is the share our AI resolves.
How does the marketplace customize their AI agent setup to work for their business?
Shaping how the agent talks and when it asks
Guidance is our set of plain-English rules for how the agent behaves. This team's rules come in three kinds: how it talks, what it asks before it answers, and when it hands over.
The clarification rules do the most work here. The correct answer depends on which trade, which payout and which account, and those rules are what get the customer to name it.
Handing over the moment a customer is upset
Anyone in clear distress (yes, including anyone writing in all caps) goes to a human straight away, and gets an apology first.
On handover we summarize the conversation and a human takes over inside the same Intercom inbox. The AI does not reply again unless it is handed back.
Escalating early is deliberate here. When the subject is somebody's money, the team wants a human on it sooner, and it shows in the numbers: escalations run at 10% and AI CSAT sits at 92%.
What impact is the marketplace's AI customer service agent having now?
- 58% AI resolution rate over the last 30 days.
- 36,413 conversations, 58% of them resolved by the AI without a human.
- 1,768 hours of agent time saved.
- 92% AI CSAT.
- 10% escalation rate.
We count a conversation as resolved when the AI handled it without escalating to a human, with escalation always one message away for the customer. Insights scores and groups every conversation for this, so the figure is a full count.
Take one deployment's number over one window with a grain of salt (this one included). Our AI resolution rate benchmark study puts 58% against the wider set.
Where does the marketplace go from here?
The next step is live data. Connecting their account data to our agent lets a Task check a payout's actual status rather than explain what usually happens to it. We would expect the next lift in resolution to come from there.
As handovers build up, Self-Learning drafts new articles from what the human agent said after taking over. The knowledge grows out of the tickets the AI could not close.
The tickets that do reach a human are the hardest ones. Our AI Copilot Chrome Extension would draft those replies for the agent inside Intercom, which is the natural next switch to flip.
Some share of money movement and fraud claims will always need a person to decide. The ceiling they are working toward is our AI handling the conversation and a human deciding the outcome.
If your inbox looks like this one, high volume, real money, and a customer who suspects they have been wronged, nothing here is exotic. A help center, the answers your agents already give, and a Task for each ticket type you see most is the whole recipe. Our pricing is per ticket, so the bill does not climb as the resolution rate does.
If you would like to see more customer stories like this one, browse all our case studies.
