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 the platform do?
- Which helpdesk does the platform use?
- How did the platform train their AI customer service agent?
- When did the platform decide to turn on 'direct replies' to customers?
- What was the biggest thing the platform did to improve their AI agent's resolution?
- How does the platform customize their AI agent setup to work for their business?
- Auto-classifying every ticket with AI Tagging
- Running one agent across chat and email
- Tuning the answers that matter most
- What impact is the platform's AI customer service agent having now?
- Where does the platform go from here?
Created time
Jun 3, 2026 01:48 PM
Title length (<60)
Author
Ecomm?
Image
Publish date
Jun 3, 2026
Slug
high-volume-ai-customer-support-case-study
Featured
Type
Case study
Ready to Publish
Ready to Publish
A trading platform resolves around 73% of roughly 105,000 monthly Intercom tickets with AI, holding a 68% AI CSAT score and saving about 5,650 hours every month, on one of the highest-volume deployments we run.
I've seen a lot of support inboxes over the years, but six figures of tickets a month is rare. This one belongs to a trading platform with a very large global base of active traders, all asking the same handful of things over and over: how the evaluation rules work, when a payout lands, what a particular account status means.
On a low-touch consumer model, every one of those tickets eats into the margin on a customer who may never pay much. And here's the thing: you cannot hire your way out of 105,000 tickets a month. So they didn't.
Today our AI agent handles the bulk of that load directly inside Intercom. It resolves around 73% of those tickets without a human, holds a 68% AI CSAT score, and hands the team back about 5,650 hours every month.
Here's how it came together.
What does the platform do?
The platform is a trading firm. It gives active retail traders a simulated, funded evaluation account, then pays out a share of the profits they make once they pass. It sits squarely in consumer fintech, with a large, global base of traders coming and going at all hours.
That business model creates a specific kind of support load. There are far more users than any team could staff against one-to-one, the questions are repetitive and high-frequency, and a slow reply on a payout is the sort of thing that turns into a public complaint fast. At this scale, deflection is the only way the unit economics work at all (you can't put a human on a near-zero-margin ticket).
Which helpdesk does the platform use?
They run support entirely on Intercom, across both chat and email. Our Intercom AI agent plugs into that same inbox, so there was nothing to rip out and no new system for the team to learn.
Before they came to us, they'd spent about three weeks running a trial of Intercom's own AI agent, Fin. And honestly, it worked: Fin was resolving around 65% of chats off their FAQ content alone, and the team were happy with the quality.
The problem was the bill. Fin is priced per resolution, so you pay every time the AI resolves a conversation.
At more than 30,000 interactions a month, that projected to north of $15,000 a month, and even after Intercom worked to bring the rate down it still didn't add up. We came in at around $3,000 a month for comparable handling, because we charge per ticket (roughly ten cents a ticket) rather than per resolution.
There was a second reason too. Fin was, in their words, a little too advanced for what they actually needed at the time.
They weren't trying to automate payouts or build complex flows yet; they needed solid FAQ handling at a price that worked at their volume. So when the Fin trial ran out, they moved the whole operation across to us.
How did the platform train their AI customer service agent?
An AI agent is only as good as what it knows, so they connected several sources before letting it answer anyone.
First was their Intercom Help Center, connected in a single click through the Knowledge Base connector. Their public website content synced in alongside it, which gave the agent the baseline of everything already documented for customers.
The highest-leverage source was their own ticket history. We auto-generate help articles from a customer's past Intercom tickets, so the agent learns from how real questions were actually answered (which, on a fast-moving product, usually goes well beyond what the help center ever covered). Their team reviewed those auto-drafted articles before they went live, then kept them current over time.
On top of that they built a library of Custom Answers for the questions where exact wording matters: the precise evaluation rules, the payout policy, the account-status explanations. For high-frequency, business-specific questions like those, a hand-written answer beats letting the agent paraphrase a help article.
When did the platform decide to turn on 'direct replies' to customers?
They didn't flip the AI to direct replies on day one. They started in note-reply mode, where our agent drafts a response as an internal note and a human checks it before it goes out, so they could watch the quality on real tickets first.
During the Fin trial they'd only ever tested on chat, never email, because they wanted to be sure before letting an AI answer email directly. Once they'd watched our agent draft in note mode and liked what they saw, they switched it to direct replies across both chat and email. The cutover happened cleanly as the Fin trial expired, so there was no gap in coverage.
What was the biggest thing the platform did to improve their AI agent's resolution?
The single biggest lever was taming the duplication that comes with this kind of volume. At 105,000 tickets a month, the same trader will often open several conversations about the same issue, across chat and email, and those duplicates inflate the queue and pull the agent in different directions.
So our team built them a custom de-duplicate-and-merge capability, specifically to handle their volume. It spots when several tickets are really the same underlying conversation and merges them, so the AI answers the issue once instead of repeating itself across three threads. On a queue this big, that's the difference between an agent that keeps up and one that quietly drowns.
Underneath it sat the continuous training on historic tickets (the boring-but-effective half of the setup). Because the agent keeps learning from how real tickets get answered, the range of things it can resolve keeps widening rather than being frozen to whatever the help center said on launch day. Together, the de-duplication and that self-widening knowledge are what hold a 73% resolution rate steady on six-figure volume.
For context on that number: 73% on this much volume is strong. Across a large field of AI support deployments, the median resolution rate sits around 70%, and almost all of those teams are handling a tiny fraction of this firm's load.
How does the platform customize their AI agent setup to work for their business?
A consumer fintech operation can't run a generic bot, so they tuned the agent around a few things in particular (the small settings that, in our experience, end up doing most of the work).
Auto-classifying every ticket with AI Tagging
The team rely on Intercom conversation labels to track why customers are getting in touch, which is core reporting data for them. Before AI Tagging, agents were applying those labels by hand (slow, and never quite consistent). Now the AI tags every incoming conversation automatically, a few seconds after the first messages are exchanged, so they get clean contact-reason data without anyone spending time on it.
Running one agent across chat and email
Because support arrives on both chat and email, they run a single AI agent over both channels inside Intercom, with the option of different guidance and tone for each. Email answers can be more thorough and formal, chat answers shorter and faster, and we don't need two separate setups to do it.
Tuning the answers that matter most
The Custom Answers library is the other ongoing job. As the evaluation rules or payout policy change, the team update the relevant custom answers so the AI is always giving the current, exact wording on the questions where getting it wrong would cause real problems.
What impact is the platform's AI customer service agent having now?
The numbers from the last 30 days:
- 73% AI resolution rate, on one of the highest-volume deployments we run.
- ~105,000 tickets handled per month, of which roughly 76,500 are resolved by the AI with no human involved.
- ~5,650 hours saved per month, at roughly five minutes of agent time per resolved ticket.
- 68% AI CSAT across the tickets the AI handled.
The hours-saved figure is the one that tends to land hardest internally. 5,650 hours a month is the equivalent of dozens of full-time agents they never had to hire, on a support operation that would otherwise be impossible to staff against.
Where does the platform go from here?
The clear next step is live customer data. They haven't yet connected the User Data API, which would let the AI answer account-specific questions, like where a trader is in their evaluation or whether a payout has been processed, using real data from their systems rather than general knowledge. For a fintech, that's the unlock that turns a big share of today's escalations into things the AI can resolve on its own.
From there, Tasks and Tools would let the agent move from answering to doing: kicking off a payout, resetting an account, or running a verification step through the platform's own APIs, with the right approvals in place.
I'll be straight about what 'fully automated' looks like at this scale, because it's easy to oversell. For an operation this size, it means the AI owns the enormous repetitive core (the same questions arriving thousands of times a day), while the team keep the judgment calls and anything genuinely sensitive.
That's the split they've built, and it's why a 73% resolution rate at this volume comes with a 68% satisfaction score rather than at the expense of one.
This setup echoes what we've seen with other teams: RecruitCRM running the same kind of Intercom support, and Freecash on the same sort of consumer-scale volume. If you'd like more stories like this one, browse all our case studies, or check our pricing to model what this would cost at your own ticket volume.
