2 January 2026

Hey Intercom. We really like Fin. We even named ours Bizy, but I’ll stick to calling it Fin here to avoid confusion. We went all-in on Fin in the beginning of November. It gets involved with most of our email- and chat-based questions (71% of them to be exact). And it resolves about 43% of those questions without human intervention. Fin’s CSAT is ok at just above 80%.

Not a bad showing. (BTW kudos to your design team for creating intuitive visualizations like this one that really help us understand how Fin is doing.)

We’ve integrated Fin with 14 home-grown MCP data connectors and have noticed that each makes Fin a little bit better at answering questions.

Fin’s impact has allowed us to flatten our human support team. We now call it Support Engineering, and they handle the more complex issues that Fin doesn’t handle well. The on-call engineer from our Product Engineering org even has a new job. S/he works alongside our Support Engineering team in Intercom answering customer questions from a shared inbox instead of handling escalated issues in Jira. Ironically, Fin has bought us space to make changes like that to get our teams closer to the customer.

Magic.

But I wish Fin did more. I’m on the Engineering team at Bonusly. And I’ve worked closely with our support engineers to implement Fin here. AI technology has progressed today to a point where Fin could reliably resolve 80% of our issues in a genuine, helpful way. That’s what we want.

So I put together a short wishlist of capabilities for you that I think could get us there.

Use our code as the knowledgebase

We ship to Production dozens of times per day. Each new improvement creates what I’ll call a micro drift between what our Help Center says the product does, and what it actually does. AI is remarkably good at making sense of code to answer questions about product behavior.

We implemented an internal agent, BonuslyGPT, that does just this. And we wrote about it here a month or so ago.

Fin’s knowledge sources are limited to our Intercom-based Help Center and these document-focused options:

I’d love to see “Github” listed here. When I click it, you can ask me for a PAT, which I’ll gladly provide. With it, I’ll give you access to peruse and ingest relevant parts of our source code with the hope that Fin will use it to answer most product-related questions (you’ll need to sync with Github regularly to keep up with aforementioned updates). All of those other documents might serve as reference material that our customers can bookmark for later.

Give us multi-turn reasoning with data connectors

We implemented an MCP data connector called “Answer general questions” that actually does give Fin access to our source code. It’s our most used connector, and we created guidance to instruct Fin to use it instead of documentation to answer questions about product behavior. The results are good.

But they could be great.

While you haven’t said this anywhere (that I can find), the “description” field for a data connector does seem to influence how Fin interacts with our customers. With other connectors we’ve added dynamic response guidance to it based on tool results and see Fin adhere to it in conversations.

Optimistically, then, we borrowed the “PLAYBOOK” part of our BonuslyGPT prompt (that internal agent I mentioned before) to see if we could coerce Fin into doing multiple turns against the MCP tool to read our code more comprehensively. BonuslyGPT does follow the playbook; we observe it query this MCP tool dozens of times with varying input before answering a product question.

But Fin is “one and done, your playbook be damned.” 

I get there are cost implications here. Multi-turn systems use more tokens. Tokens cost money. Your business model is resolution-based, not token-based. Margins matter.

Perhaps time is another reason. Multi-turn reasoning can take minutes. Within voice and chat context, you can’t take minutes to respond. But in email context, you can. Frankly, I’m willing to ask the customer to wait a little bit longer in any case so that Fin can get them the right answer. And maybe Fin can tell them a joke or two while it’s reasoning. (Or, maybe it needs to be ok to put people on hold again.)

Tool chaining, please

Now, this is a wishlist item that, I think, will be answered once Procedures is available. But I’ve been staring at this screen for a couple of months now:

What we’ve seen is that more complex problems can still be solved by Fin if only it could chain tool calls together. One chaining scenario here is related to fact collection. For example, it’s common for our customers to write in with a question about a pending reward.  A reward can be pending for many reasons: 

• Is the company in good financial standing?
• Have we detected fraudulent activity?
• Is there an outage with a downstream provider?
• Did the administrator recently turn off auto-fulfillment?

Imagine if Fin could issue a call to a tool to answer each of these questions, then compile what it learned into a well-informed response. Boom.

Anyway, you all have unlocked a valuable AI-oriented use case with Fin. We’re grateful. And thanks for listening.

Cheers,

JT

SVP Engineering, Bonusly| LinkedIn

‍