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
- What does Inspire Uplift do?
- Which helpdesk does Inspire Uplift use?
- How did Inspire Uplift train their AI customer service agent?
- When did Inspire Uplift decide to turn on 'direct replies' to customers?
- What was the biggest thing Inspire Uplift did to improve their AI agent's resolution?
- How does Inspire Uplift customize their AI agent setup to work for their business?
- Routing most of the inbox to humans on purpose
- Keeping replies on brand
- Asking the right follow-up before acting
- What impact is Inspire Uplift's AI customer service agent having now?
- Where do Inspire Uplift go from here?
Created time
Jun 3, 2026 01:54 PM
Title length (<60)
Author
Ecomm?
Tags
Image
Publish date
Jun 3, 2026
Slug
my-askai-inspire-uplift-case-study
Featured
Featured
Type
Case study
Ready to Publish
Ready to Publish
Inspire Uplift runs about 6,100 monthly Zendesk tickets through My AskAI, auto-drafting around 1,300 replies a month, holding 66% AI CSAT, and saving the team roughly 105 hours every month.
Inspire Uplift is an online marketplace for unique gifts and everyday products from independent sellers. A marketplace inbox is a two-sided thing: shoppers ask about orders and returns on one side, sellers ask about listings and payouts on the other, and plenty of cases land in the middle where a buyer and a seller both need a person to sort something out. At roughly 6,100 tickets a month, they wanted to lift the repetitive load off their agents without handing customers to a bot that would fumble the threads that need a human.
So they did close to the opposite of the usual ecommerce automation pitch (this is one of my favorite kinds of rollout to talk about). Instead of pushing the AI to close as many tickets as it can, they tuned it to answer the easy, repeatable questions on its own and pass everything else to a human with the context already gathered. After the first 30 days, the numbers we saw were around 1,300 replies a month drafted by the AI, 66% CSAT on the tickets it handles, and roughly 105 hours handed back to the team every month.
Here's how it came together.
What does Inspire Uplift do?
Inspire Uplift launched in 2017 around a simple idea: shopping should feel good. Today it connects more than 55,000 independent sellers with over two million shoppers around the world (support at that size stops being a nice-to-have), listing north of 20 million products across handmade goods, home and kitchen gadgets, jewelry, beauty, pet supplies and digital downloads.
It is a genuinely two-sided business. Sellers get free membership and free shop creation, so the support team fields questions from the people buying and the people selling, often about the same order (and that overlap is where it gets interesting). That two-sidedness is the reason most of their tickets need a human rather than a canned answer.
Which helpdesk does Inspire Uplift use?
Inspire Uplift runs support inside Zendesk, in the Support and Tickets inbox. At marketplace scale that inbox is busy and varied: order status and tracking, returns and refunds, seller onboarding, payout questions, listing and policy queries, and the buyer-versus-seller disputes that only a person can settle.
We plug straight into Zendesk Tickets, so there was no new tool for their agents to learn and no second window to live in. The AI reads incoming tickets, answers the ones it is confident about, and hands the rest to the team inside the same Zendesk workflow they already use. For a team handling around 6,100 tickets a month, slotting the AI into the helpdesk they already live in was the boring-but-effective call.
How did Inspire Uplift train their AI customer service agent?
A marketplace can't be answered well from a single FAQ page, so Inspire Uplift connected three knowledge sources (help center, internal docs, and uploaded files) and let the AI draw on all of them.
First, they synced their on-site help center through our website sync, so every published help article became answerable knowledge the moment it went live. This is the front line for the common buyer questions: where an order is, how returns work, what the shipping timelines are.
Second, they connected Dropbox through our Dropbox connector. A marketplace runs on internal process documents that never make it onto a public help page: seller policies, payout rules, escalation procedures, edge cases. Pointing the AI at that material let it reason about the more involved questions instead of guessing.
Third, they uploaded files directly for anything that lived outside the help center or Dropbox (the catch-all bucket). Between the three sources, the AI had the buyer-facing answers and the operational context sitting behind them.
One source they haven't switched on yet is the User Data API, which would let the AI look up a specific order or seller account live (we'll come back to it in the roadmap). That one decides which questions the AI can safely close on its own today.
When did Inspire Uplift decide to turn on 'direct replies' to customers?
Direct replies are on, but kept on a short leash. Rather than letting the AI attempt every ticket, Inspire Uplift uses Handover and Escalation Guidance to define a narrow band of clearly safe, repetitive questions the AI can answer end to end, and routes everything else to a human (in our experience that caution is the right place to start).
The team decides which questions the AI closes and which it prepares for a person, and they can widen that band over time as they get comfortable. On an inbox where a wrong answer to a seller about a payout costs far more than a slow one, starting tight makes sense.
What was the biggest thing Inspire Uplift did to improve their AI agent's resolution?
The single biggest lever has been Self-Learning. Every time a human agent answers a ticket the AI handed over, our Self-Learning feature studies that reply and drafts new knowledge from it, so the next time a similar question arrives the AI already has an answer. Over the last 30 days, articles the system drafted this way fed into around 1,300 ticket responses a month.
Inspire Uplift isn't chasing the 70%-plus rates you see on a single-brand DTC store, because most of their inbox is meant to reach a human. What Self-Learning improves is the quality and coverage of the slice the AI does handle (no one sitting down to write help articles by hand). That widening slice is what turns into hours saved, every week.
Across the wider field of AI support deployments, resolution rates cluster around 70% on average, per our resolution-rate benchmarks aggregated across roughly 55 vendors. Those numbers are directional rather than like-for-like, since every vendor counts a resolution differently. Inspire Uplift sits well below that average, and that's on purpose.
How does Inspire Uplift customize their AI agent setup to work for their business?
Inspire Uplift configured all three types of Guidance, and the tuning is where being a marketplace really shows up.
Routing most of the inbox to humans on purpose
Handover and Escalation Guidance is the dominant rule in their setup. It sends roughly 77% of tickets to a human (disputes, payouts, refunds, anything two-sided). This is the rule that explains their headline numbers, and it is doing exactly what they want.
A multi-seller marketplace is full of cases where the right move is a person with full context. The AI's job is to spot those early and route them, instead of attempting an answer it shouldn't. When it routes a ticket, it gathers the relevant context first, so the agent picks up a thread that already has the groundwork done (the part we care about most).
Keeping replies on brand
For the questions it does answer directly, our Communication Guidance keeps the AI's tone, structure and sign-off consistent with how Inspire Uplift's own team writes. A shopper shouldn't be able to tell whether a quick order question was answered by a person or the AI, beyond the fact that the reply arrived faster.
Asking the right follow-up before acting
Context and Clarification Guidance (the third type we give them) handles the in-between cases, where the AI needs one more detail before it can help. It asks the clarifying question first, which keeps more of the easy tickets out of the human queue without sacrificing accuracy.
What impact is Inspire Uplift's AI customer service agent having now?
Here is where the setup lands after 30 days:
- ~6,100 tickets a month flowing through the AI layer inside Zendesk Tickets.
- ~1,300 replies a month drafted from knowledge that Self-Learning created on its own.
- ~105 hours saved each month for the support team.
- 66% AI CSAT across the tickets the AI handles.
- ~21% AI resolution rate, which is low on purpose. With Handover and Escalation Guidance routing roughly 77% of tickets to people, only a small, chosen band is ever meant to be closed by the AI.
Let's be honest, a 21% resolution rate looks low on a slide. When we say a ticket was resolved by the AI, we mean it was handled without being escalated to a human (we don't pretend to know more than that). We make escalation easy, so a customer can reach a person whenever they want, and the AI hands off the moment it is unsure.
On a marketplace where most threads need a human, a 21% resolution rate sitting next to 66% CSAT and 105 hours saved is the setup working exactly as designed. I think that's the right way to read it: the real question for a marketplace is how much repetitive load the AI can lift while the human-needed cases stay human.
As a marketplace, a lot of our tickets involve interactions between a buyer and a seller, and getting those wrong costs us more than answering them slowly. The value hasn't been automating everything; it's been letting the AI handle repetitive questions so our customer support team can focus its judgment on the conversations that actually need it. My AskAI has been great at either helping customers directly when possible or gathering the relevant context upfront, so when a support agent needs to step in, they already have the information they need to resolve the issue quickly and accurately. - Aaron Wallace, Co-founder, Inspire Uplift
Where do Inspire Uplift go from here?
The clearest next step is the User Data API (the one upgrade we'd push for first). Connecting it would let the AI look up a specific order or seller account in real time, which is what stands between today's setup and safely closing more order-status questions on its own. A lot of the tickets currently escalated only need a human because the AI can't yet see the live data behind them.
After that, Tools and Tasks would let the AI take guarded actions rather than just answer: processing a straightforward refund, say, or updating an address, each behind the guardrails a marketplace needs. And as Self-Learning keeps proving out new question types, the team can graduate more categories from "escalate" to "reply directly", widening the AI's band one safe step at a time.
For a marketplace, I'd say full automation means the team only ever picking up the threads a person has to handle, with the AI carrying the rest. Inspire Uplift is already most of the way there, and the roadmap is about pushing the line a little further each month.
If you'd like to see more customer stories like Inspire Uplift's, browse all our case studies, or take a look at our pricing.
