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
- Creator platform AI customer service results on Zendesk at a glance
- What does the platform do?
- Which helpdesk does the platform use?
- How did they train their AI customer service agent?
- When did they decide to turn on 'direct replies' to customers?
- What was the biggest thing they did to improve their AI agent's resolution
- How do they customize their AI agent setup to work for their business?
- Controlling tone and language
- Knowing when to ask, and when to escalate
- What impact is their AI customer service agent having now?
- Where does the team go from here?
Created time
Jun 3, 2026 09:29 AM
Title length (<60)
Author
Last optimised
Sep 15, 2026
Ecomm?
Image
Publish date
Jun 3, 2026
Video
Slug
creator-platform-ai-customer-support-case-study
Featured
Featured
Type
Case study
Ready to Publish
Ready to Publish
A creator-monetization platform runs My AskAI as an internal Copilot inside Zendesk, drafting answers on ~1,400 monthly support tickets so agents reply faster. Every draft is grounded on their Zendesk help center and a deep library of Custom Answers, and seen by a human before it reaches a creator.
You've probably landed here for one of two reasons. Either you run a busy consumer platform on Zendesk and you're trying to picture what AI in the inbox looks like before you let it talk to customers. Or you've heard people mention running AI in "copilot" or internal-note mode (it drafts, a human sends) and you want to see it at real volume rather than in a demo.
Either way, I've got you. This company sits at almost exactly that profile: a busy inbox full of repetitive, policy-heavy questions about payouts, fees and accounts, and a team that wanted speed without giving up the final word on what gets said.
So we draft, and they send. Today My AskAI writes an internal-note answer on around 1,400 of their Zendesk tickets every month, and an agent reads each one before it goes anywhere.
They send it, edit it, or bin it. Replies default to British English (no stray American spelling), and direct-to-customer replies are off for now, by design.
Here's how it came together.
Creator platform AI customer service results on Zendesk at a glance
Fact | A creator-monetization platform |
Industry | Fintech, creator monetization and direct payments (B2C) |
Helpdesk | Zendesk (Support tickets) |
Support volume | ~1,400 tickets/month |
Key features used | Custom Answers, Guidance across all three types, a British-English default, Zendesk tagging and triggers |
Knowledge sources | Zendesk help center via the Knowledge Base connector, a large Custom Answers library imported from the team's text-expander macros |
Go-live mode | Internal-note Copilot mode only, direct replies deliberately switched off |
What does the platform do?
It is a creator-monetization service. It gives creators one place to get paid by their fans, across one-off payments, recurring memberships, commissions, and a shop for digital and physical products.
That creator-first model means a lot of the inbox is money- and account-shaped: when do I get paid, why was I charged this, how do I change my plan, what counts as eligible. Those are exactly the questions where the wording of the answer has to be right, and their help center is where most of the canonical answers already live.
Which helpdesk does the platform use?
They run customer support out of Zendesk. Tickets land in the Zendesk inbox, and that's where we sit, inside the same ticket view the team already works in.
We run there in Internal Notes / Copilot mode. Instead of replying to the customer, the AI drafts its answer as an internal note on the ticket, and an agent decides what happens next: send it as-is, tweak a line and send, or bin it and write their own. Nothing reaches a creator without a person having looked first.
That's a deliberate choice. For a service where a wrong answer about a payout or a fee is worse than a slow one, copilot mode gets the team the speed (a ready-to-go draft on every ticket) while keeping a human on every single reply. The integration itself is the standard My AskAI Zendesk one; the interesting part is what they did with it.
How did they train their AI customer service agent?
Two knowledge sources do the heavy lifting.
First, we connected their Zendesk help center through the Zendesk Knowledge Base connector. That hands the AI the existing library of help articles to draft from, the baseline of how the team already explains tips, memberships, the shop and account settings.
Second (and this is the part that makes the drafts sound like the brand), the team built out a large library of Custom Answers. These are exact question-and-answer pairs the AI returns word-for-word, and we built the feature precisely for the moments when paraphrasing isn't good enough. They seeded theirs by importing the text-expander macros the support team already used: the canned, agreed wording agents reached for on the questions that come up again and again.
So when a payout question comes in, the AI drafts the exact line the team had already signed off on. The training here is knowledge-led: help-center articles give breadth, Custom Answers give the exact, high-stakes wording.
When did they decide to turn on 'direct replies' to customers?
They haven't.
They run in internal-note mode only. Every draft is read by an agent before it goes out, and direct replies stay switched off. The plan is to graduate one category at a time: prove the AI drafts a given query type well enough that an agent is just hitting send, then think about letting it reply directly on that category alone.
What they're watching first is draft quality on the money- and policy-heavy tickets (the ones where being confidently wrong would do real damage). Copilot mode buys them that confidence without putting it on a customer to find the mistakes. A group-payments platform on Zendesk runs the same notes-only setup for the same reason.
What was the biggest thing they did to improve their AI agent's resolution
The single highest-leverage move was the Custom Answers library, used as the place to pin down the answers that have to be exactly right.
For most businesses, "improving the AI" means widening what it knows. For a payments service it's sharper than that: it's making sure the AI never freelances on the answers where the wording is the policy.
When a creator asks why a payout is delayed, or whether a fee applies to them, a plausible-but-slightly-wrong paraphrase is the kind of answer that creates a second, angrier ticket (and when it's someone's payout on the line, those are the worst kind). Custom Answers close that gap by pinning the response to the exact text the team approved.
Importing the existing text-expander macros made that fast. The team had already written and refined those canned answers for human agents over the years, so turning them into Custom Answers handed the same vetted wording to the AI on day one. The British-English default keeps the spelling in every draft consistent with the brand.
Because they run in copilot mode, "better" is measured as draft acceptance: how often an agent can read our note and just send it. That's the number the Custom Answers work moves.
How do they customize their AI agent setup to work for their business?
Beyond training, the team uses Guidance to steer how the AI drafts and when it should step back. All three Guidance types are configured.
Controlling tone and language
Communication Guidance sets the voice: warm and creator-friendly, the way the team already talks to the people who use the product. Paired with the British-English default, it keeps drafts on-brand without an agent having to rewrite for tone before sending.
Knowing when to ask, and when to escalate
Context & Clarification Guidance tells the AI to gather the missing detail before it commits to an answer (the account, the specific transaction, the plan). Handover & Escalation Guidance does the opposite job: it keeps the AI away from the tickets that shouldn't be near it. They run a separate trust & safety team alongside support, and those tickets stay human-only, as do the sensitive edges around payout disputes and account security.
So the AI drafts the high-volume, well-trodden questions and routes the rest to people. This is also where the staged rollout lives: using Zendesk's own tagging and triggers, the team can point the AI at one query type at a time and widen its remit step by step.
What impact is their AI customer service agent having now?
Here's where the rollout sits today:
- ~1,400 Zendesk tickets a month carry an AI-drafted internal note.
- Copilot / internal-note mode only: an agent reviews, edits or bins every draft before it's sent.
What the team gets is speed: a ready-to-go answer waiting on most tickets, so agents move through a repetitive, policy-heavy inbox faster without handing over the final word. For a service handling other people's money, that trade (faster drafting, human still in control) is the one they wanted.
Where does the team go from here?
The roadmap is a steady widening of what the AI is trusted to do, in the same one-step-at-a-time spirit as the rollout.
The nearest move is graduating the safest, highest-volume query types from drafted notes to direct replies, letting the AI answer a proven category on its own while the rest of the inbox stays in copilot mode. After that, we'd bring in Tasks and Tools so the AI can do more than draft text and actually run the multi-step procedures behind questions like refunds and account troubleshooting. Connecting live account data through the User Data API would let those answers speak to a creator's real payout and account state, and Self-Learning can keep turning the questions the AI couldn't answer into new draft material over time.
For a direct payments service, "100% automation" was never the goal and probably never will be. What it looks like here is the AI taking the well-trodden, high-volume questions directly, with people staying firmly in charge of the long tail: the trust & safety work, the genuine edge cases, the conversations that need a human. Copilot mode was the place to start, and from where we sit it's earning the right to do more, one query type at a time.
If you'd like to see more customer stories like this one, Apartment List and Sofar Sounds run My AskAI on Zendesk too, and you can browse all our case studies from there. If you're weighing up AI for your own Zendesk inbox, take a look at how the pricing works.
