Support Stack Episode 11 - Teaching Fin Your Product Language
When teams switch on Fin, they often expect strong results straight away.
But there’s a hidden problem most don’t spot at first:
Your customers don’t speak your product’s internal language.
In this episode of Support Stack, Gabriela Passeri, Knowledge Manager at AutoDS, shares what happened when her team introduced Intercom’s Fin AI Agent — and realised that terminology gaps were quietly undermining resolution.
Customers were saying “products”.
The platform called them “listings”.
“Balance” could mean a feature… or an amount.
“Credits” meant different things in different parts of the product.
Fin wasn’t wrong. It just didn’t have the context.
Instead of patching individual answers, Gabriela and her team took a more structured approach. They built internal, AI-enabled glossary articles designed specifically for Fin and Copilot.
In this episode, we explore:
Why terminology mismatches quietly reduce AI resolution rates
How to structure glossary articles with terms, definitions, synonyms and “avoid confusion with” guidance
Why internal articles can be more powerful than publishing everything publicly
How to help Fin interpret screenshots by adding layout context
Why reviewing real DSAT and low CX Score conversations is essential
The discipline required to manually audit conversations and iterate
This isn’t about advanced data connectors or complex automations.
It’s about language precision.
If you’re implementing Fin in a B2B SaaS environment with complex features, overlapping terminology or integration-heavy workflows, this episode offers a practical starting point.
Before you optimise tasks.
Before you build procedures.
Make sure Fin understands your product vocabulary.
Here’s the doc that Gabriela created for you!: https://docs.google.com/document/d/1Q7eFGcrwHi70SyOltGKw6emiIThh82au9hQqFGLSOxo/edit?tab=t.0#heading=h.u9e4hny14af
If you want more practical, real-world examples of how teams are implementing Fin — and how support roles evolve alongside AI — subscribe to my daily email:
Episode transcript
Conor Pendergrast (00:00)
Hi, I'm Conor Pendergrast and welcome to episode eleven of Support Stack.
So Gabriela Pisseri, welcome to Support Stack. How are you doing today?
Gabriela Passeri (00:16)
Thank you, Conor, I'm doing very well.
Conor Pendergrast (00:18)
Good. So you are knowledge manager at AutoDS and AutoDS has just started using Intercom and Intercom's FIN AI agent. Is that correct? Lovely. And so I saw that you had a post on LinkedIn where you were talking about as a knowledge manager, some of the things that you were doing to get started with FIN AI agent. I'm always interested in those effective strategies and tactics for
Gabriela Passeri (00:27)
Yes, exactly.
Conor Pendergrast (00:45)
for how to set up FinAI agent for success, especially because you can get a whole bunch of great work done without getting into the complexities of data and actions. what I was hoping to talk to you about is, well, ⁓ what were you thinking about? So cast your mind back. What were you thinking about in that first switch to intercom? What was it that you found you needed to tackle first? What were the priorities for you?
Gabriela Passeri (01:13)
Yeah. Our main goal back at the time was to introduce AI into your chat bot in our replies, because we are struggling with the same questions for agents. And it's a lot for them to take like common questions and easy ones for answer the user we have. So we thought that we needed AI for this.
small things like small questions, common questions and everything that we can cover with our own knowledge base. So we decided to change to intercom that is more robust than the previous one we are using. So we decided to change and start working with AI more closely.
Conor Pendergrast (01:55)
Great, excellent. And so while you were doing that, process of getting started with Fin and getting started with intercom, there was something that came up that was a bit of a challenge for you though. And I think this is probably something that anyone who runs a B2B business with integrations with other products is going to come across. So tell us about the first challenge that you saw ⁓ when using an AI agent to talk to customers compared with human teammates talking to customers. What was the biggest challenge there?
Gabriela Passeri (02:24)
I think like as all AI works, they need their own prompt in their own language. So we're struggling to find some ways to work around questions that Fin was unable to answer correctly, like very accurate, you know? So this was the first blocker we found. Like we need to understand AI first.
So to understand AI first, we need to build a document, a knowledge base that has this knowledge of AI first. So we decided to change all the articles we had back at the time. We had like 230 articles on the previous CRM. And now we only have 55 articles because they are more AI-readiness. They have this tool.
Conor Pendergrast (03:05)
Wow.
Mm-hmm.
Gabriela Passeri (03:14)
the way that we are building the article, is for the prompt, it's like a prompt. The articles now, they are almost a prompt for Fin to understand.
Conor Pendergrast (03:25)
Perfect, perfect. So we're going to cover generating and creating those articles in the second episode. So listener, watcher, viewer, click that subscribe button if you want to see the next episode with Gabriela, I am also really interested to share that. One thing that you found though, when you were seeing, so I imagine it was from testing Fin, you were finding confusion between Fin and customers, or Fin and pretend customers in terms of the questions
that were, that customers would ask and the language that they would ask. That's right?
Gabriela Passeri (03:56)
Yeah, it's like, for example, I speak two languages. It is Portuguese and English. So it's not about the language itself. It is more about problem that Fin couldn't map
Conor Pendergrast (04:02)
Mm-hmm.
Gabriela Passeri (04:08)
customer language to our documentation terminology, the one that we use internally the company. So we built the articles with our own terminologies. And during the conversations, we're analyzing Fin. We understand that a customer would say like, my products aren't showing up. And our own articles, the terminology for products were listings. So we identified that the terminology was not the same.
Conor Pendergrast (04:09)
Yeah.
Yeah.
Yeah. Yeah. And what's, what's interesting is I remember when I, when I was, I was a senior leader in customer support at Expensify and Expensify had a classic, had two classic examples of confusion between us. is pre back in my day. This is before FinAI agent. and the challenge would be understanding between expenses and receipts. So customers would say, I've lost my receipts. but what they meant was what we called expenses.
Or they'd say, ⁓ I need to create a report. And we would never know, do they mean a grouping of expenses or would they mean an analysis and that sort of financial analytics of spend? And I can't like, I have definitely seen examples of Fin also struggling with that, just like you did. And this, the approach that you've taken, I think is really novel and a really great thing for people to learn from in that, what did you do?
Gabriela, how did you teach Fin what the difference between all these things are?
Gabriela Passeri (05:35)
⁓ we started analyzing Fin's Conversation to understand how the user is using the terminologies. And then we create a glossary. So everything that's inside our own platform, all the features we have are there in the glossary and also how the user referred to this same feature.
Conor Pendergrast (05:44)
Hmm.
Gabriela Passeri (05:55)
So we build a document for Fin to understand each column, each term, definitions, synonyms, avoiding confusion with the same wording because inside our own platform, we also have wordings that are the same. So we avoid this confusion with the features and also fields we have inside our own platform.
Conor Pendergrast (05:56)
Yeah.
Mm-hmm.
That's super. Okay. So, so now let's, let's show your screen here. So,
great. So we're looking perfect. This is perfect. So this is an example of a glossary. Now, first question, I think maybe other people don't feel this way, but I always find it really useful to know like the precision behind what you've done. So it looks like this is, it's not a publicly available article. Is that correct? It's an internal article.
Gabriela Passeri (06:41)
Exactly. It is internal and also enable it for Fin and copilot.
Conor Pendergrast (06:46)
Of course, of course, because otherwise
Fin wouldn't be able to see it and it'd be a bit pointless.
Gabriela Passeri (06:51)
Exactly, this is the point, right?
Conor Pendergrast (06:53)
Yeah. And so, so that's one aspect of it is you've created it as an internal article, not a public article. The other thing to note it looks like is you've created a set of glossaries. It's not just one glossary. It's you have multiple glossaries.
Gabriela Passeri (07:07)
It's topic related. this is the AutoDS platform and features terms. We also have for billing terms, orders terms, and other teammates we have inside the company. So we covered everything for all the teams we have.
Conor Pendergrast (07:20)
Yeah.
That's fascinating. That's great. So you use topics and is that the Intercom's Fin.ai topics or is it just informally the topics that you have?
Gabriela Passeri (07:36)
So we do create ⁓ each topic related with the Fin categorization. And also from the reports, we saw that we needed for each one, each topic that Fin is analyzing. So we create a glossary for each one.
Conor Pendergrast (07:48)
Yeah,
that's super, that's super. Okay, so you've got this glossary and then you've got a lovely table here which Fin can read. So it's got term, definition, synonyms, avoid confusion with and related terms. So, I mean, I imagine that just the term and definition is helpful, but the synonyms and avoid confusion with is really powerful and should be a really great way of Fin being like, wait, hang on. ⁓
That's so I just talk us through an example, talk us through an example of how Fin have have you seen Fin apply this.
Gabriela Passeri (08:22)
I have one good example, the synonyms and they avoid confusion with, let me just scroll down a little bit and I will show you something that we realized. Yeah.
Conor Pendergrast (08:31)
Wow, it's a lot of terms, isn't it?
That's incredible, and this is
just in one single topic.
Gabriela Passeri (08:39)
Yes, one single topic. Because we have a lot of features. That's why this is the most large one. ⁓ This was one thing, balance. This term, is referring to a tool where the user can purchase balance inside the platform. However, balance is also a word for amount
So we are not using balance in any circumstances that is not related to the feature itself. Only the feature. So...
Conor Pendergrast (09:09)
Okay, yeah, yeah.
Gabriela Passeri (09:12)
The user are telling us, purchase the FBA, but not same balance. So explain to Fin what is managed balance, what is the synonyms the user are using, and avoid confusion with the amount. Balance is not just an amount, it is a feature.
Conor Pendergrast (09:31)
Yeah, yeah, okay. That's great.
Gabriela Passeri (09:32)
And credits,
another wording that is common, right? Credits. So we have two credits types. is AI credits and auto-order credits. And the tool here that you are seeing right now, it is to locate where the manage balance is and where is the AI credits, where is the auto-order credits, because users are confusing because the terms are similar.
Conor Pendergrast (09:37)
Yep, credits.
Mm-hmm.
Mm-hmm.
Gabriela Passeri (10:00)
So we need to Fin to identify where is each one, because here in the platform, you only have icons, not the text. So Fin could not understand that this icon referred to manage balance. So we thought Fin that manage balance, it is this one, AI credit is this one, and auto-order credits is this one. And also not just an image.
Conor Pendergrast (10:00)
Yeah.
Gabriela Passeri (10:26)
We also add a text locating the top right corner of the screen header.
Conor Pendergrast (10:32)
Yeah, you're telling Fin exactly where it
is in that image. Yeah. And that's really interesting as well, because I bet that helps Fin to more accurately use screenshots when sending them back to customers, because when we're recording, we're recording in February, 2026 and screenshots are only about two months. We've only been able to send them or Fin only has only been able to send them through, through messenger and through email for about two months. So it's pretty, pretty new.
But I bet this sort of guidance and this, this glossary really helps Fin to understand screenshots, but also to understand when to send appropriate screenshots as well.
Gabriela Passeri (11:09)
Exactly, we do have this feature inside our own intercom, And the fun story about this one, we identify in a conversation, we are checking conversations with Fin, and a user only send a screenshot and ask him, how much balance do we have? And Fin wasn't able to reply to the user because he couldn't read the image. That's why we create an image and also the
Conor Pendergrast (11:11)
Mm-hmm.
Right?
Okay.
it didn't have that. Yeah.
Gabriela Passeri (11:34)
the guidance, what is, where is, and everything here.
Conor Pendergrast (11:38)
Yeah, yeah. There's so many layers to this. It's so interesting because it is both the language side of it, which Fin is usually pretty good at, but it's getting very precise on what your customers can say and what you say based on your actual product. But also the visual side of it is really fascinating. And just knowing, giving a frame of reference for what an image should look like, because I've definitely seen confusion from Fin on like, okay, here's what they send a screenshot.
and the customer sends a screenshot and Fin is just like, you're looking at this. And they're not, they're looking at something totally different. It's just, you know, maybe it's a third party, a third party tool. ⁓ So yeah, that's fascinating. So I'm curious, this, presume you've done, you've done some testing on how this improved things. Like what were the signs for you that this was working well? Did you use, did you just...
Gabriela Passeri (12:13)
Yeah.
Conor Pendergrast (12:32)
observe it in the wild of customers interacting with Fin? Or did you use the testing groups or the testing suite? What were the like, how did you get to the point where you're like, this is good, this is working?
Gabriela Passeri (12:43)
We need people for that because we need to manually check each conversation. So we started with the DSAT conversations and the low CX score to identify this type of conversation. So it was four people analyzing the conversation and adding what Fin answered wrongly. And then my team and...
We understand how we are going to do this. So we start brainstorming and we created the glossary and this was a second part of the plan, like to present the UI to Fin. So I came to let's try this way. Let's give Fin the instructions on how to read the platform layout. And we did and.
Conor Pendergrast (13:20)
Mm-hmm.
Mm-hmm.
Gabriela Passeri (13:31)
We use the same question that the user did. We are not creating questions because we want to have the user wording, right? The user terminology. So we test it. open a conversation with thing and test it and work it like, like, was amazing to see this working. We tested it out. We just give the...
Conor Pendergrast (13:40)
Yeah.
Mm-hmm.
That's so good.
Gabriela Passeri (13:52)
this one first, the text first, and then we realized we need the image because the other user we found only sending screenshots instead of asking questions. So it was a very good one, very good process.
Conor Pendergrast (14:07)
That's super.
Yeah, that's super. there's, yeah, I think two things to pick out of that. The first thing is you have to do your conversation testing. Excuse me, your conversation reviews. You have to do your conversation reviews. That's the only way to find the problems. And I like the targeting of both low CX scores as well as the low CSAT ratings. So the D-SAT, the dissatisfied satisfaction scores. I think that's great way of finding those problematic conversations.
Gabriela Passeri (14:36)
Yeah, we started because it was a lot of conversations. couldn't like see all the conversation with the capability we have inside now. So we created a process, SOPs and everything for this to make sure we were tagging each conversation and also analyzing how much we did resolve, how much we do have for resolving. It is a lot.
Conor Pendergrast (14:36)
⁓
Yeah.
Yep. Even with four people, it's still, you can have thousands of conversations. Yeah. Yeah.
Great.
Yeah, yeah, that's brilliant. And then the second point to pick up from what Gabriela said is just to reinforce, use your customer's language. Literally copy and paste what they say. Don't imagine it. I found this myself yesterday, yesterday and the day before. So I just set a new task live for a client. And I was like, this is great. I've tested it. I tested it using existing customer language. I tested it using my own words. I tested it in like probably about 15, 20 different ways. And yet the first conversation that happened.
It was just slightly unexpected and Fin just got slightly thrown in it and it ended up not resolving it appropriately. And it was very frustrating. It did get escalated to a human teammate, obviously, because it wasn't working out. But it was just one of those things where I was like, ⁓ I have to remember that the whole point of this is to create great customer support experiences. It's not to create like some magical bot. The purpose of this is great customer support experiences. Nothing else.
Gabriela Passeri (15:56)
Peace is ours.
of course
is our goal. And also we did create the glossary instead of creating articles on Hub Center because we have the terminology inside the platform. So you must use the same terminology as the platform in your articles. You cannot use the user terminology. So this way we identified that we need the glossary as internal article, not as a published one.
Conor Pendergrast (16:15)
Absolutely.
That's super. That's wonderful. So Gabriela, you have very kindly written up this document as well, or written up this idea in a document. So dear viewer, look down, just down there, there's a little link in there to a Google Doc that Gabriela has created for you. That's for both this episode and again, teasing, click the subscribe button, the next episode that's coming out with Gabriela, which is all about the public articles side of it. So I'm really interested in how they went from
What was it, 250 to 50 articles? 300 to 50 articles?
Gabriela Passeri (16:57)
I believe. don't know.
Conor Pendergrast (16:57)
Okay, great.
Super, super. This is great. Thank you so much for being here, Gabriela. Is there anything that you'd like to promote? Like, where should people follow your amazing work online?
Gabriela Passeri (17:08)
Yeah, you can follow me on LinkedIn, Gabriela Passeri and also from AutoDS because I'm working for AutoDS. So it is good for you to understand this amazing platform we have for sure.
Conor Pendergrast (17:21)
Super. Well, thank you very much for joining us for another episode of Support Stack, Gabriela, and thank you, viewer, for watching. You can find my work at customersuccess.cx. You can join my week daily email list where I talk all about Fin AI Agent and the evolution of customer support teams as we get AI agent into our support organizations. You can find that at customersuccess.cx/Daily. And, yep, that's where you'll find everything. Other than that, I'll see you next time for the next episode of Support Stack.
Bye bye.
Gabriela Passeri (17:51)
Thank you, bye bye.