Train Fin

 In Ep. 11 of Support Stack, Gabriela Passeri (Knowledge Manager at AutoDS) walks through how they made Fin actually understand their product.Their key move was to build internal, Fin-enabled glossaries mapping customer language to product terminology. Including synonyms, “avoid confusion with”, and even UI layout guidance.Super practical and very applicable if you have a complex B2B SaaS product with integrations!Full episode: I really hope it helps! 

I'm looking for a streamlined QA process for managing Fin’s performance as a one-woman team. Would love some suggestions for best practices for maintaining quality without getting overwhelmed with all the reports and data Intercom provides for Fin performance.

Is there a way that FIN can be instructed to go to certain "Paths" within the workflow or it's just a dead-end when we let FIN handle it?  I noticed that when a user goes into the “Let FIN handle” node, there is no way out. The only option is to escalate/hand it over to a human agent. 

Hi everyone,Our bot communicates in Hebrew. Because Hebrew is gendered, we’re seeing an inconsistency: Fin sometimes replies in masculine form and sometimes in feminine form.Both of the following Guidance rules are under Communication style: Bot self-reference is always masculine When responding in Hebrew, always refer to yourself in masculine grammatical form. This applies to self-descriptions, actions, and statements about your role or capabilities. Do not change your self-referential gender under any circumstances.What NOT to do: Do not adapt self-reference based on the user’s gender or language. Do not use feminine or neutral forms when referring to yourself.  AND:  Hebrew gendered language (pronoun-aware) When responding in Hebrew, adapt grammatical gender when addressing the user:

I am pretty new to Intercom although I am pretty familiar with other CX tools that are similar - I have been struggling lately with our bot giving completely wrong and false information to our customers on topics that need to be answered with accuracy. It is drawing information from places outside of the articles and snippets we have supplied in our instance. I am struggling with understanding why it would pull in information that is seemingly just completely made up or not in line with things we have published for the site. What would be some suggestions for how to update and change this? Should we use guidance? How specific can we get? Any tips at all? Thanks!

Fin is technically able to read dates, but in practice it often fails to reason about them correctly.ExampleFin currently sends out discount codes that are already expired, because it doesn’t reliably calculate whether a date is in the past, present, or future.Feature request    •    Fin should be able to compare dates to the current date    •    Fin should not mention or suggest expired offers    •    Ideally, Fin should be aware of:    •    “valid until”    •    “expired”    •    “not yet active”Why this matters    •    Prevents customer frustration    •    Avoids incorrect promises (e.g. expired discounts)    •    Increases trust in AI-generated responses 

I would like to set FIN over the holidays so that it informs customers that the wait time for human support may be a little longer. Where can I set what FIN responds when asked for human support? 

Hi folks, I am a product manager on a no-code app building product for manufacturers. We allow process engineers to build apps and deploy them onto a manufacturing shop floor.Since app building is a creative process, and requires understanding of UI, logic and our no-code database, our customer questions tend to be complex- they are not easily answered by single knowledge base articles, and the possibilities for using our features are essentially endless.Has anyone come up with prompts that help Fin answer these types of questions?As an example, I have gotten pretty good results from adding this prompt to the “Other” category under guidance. It took a lot of trial and error.title: Assume complex answers requiredbody: If a customer asks a question that spans multiple features in our product, investigate each feature individually before returning a response. Your answers should comment on the review of each individua

I’ve noticed FIN closing some conversations because the user didn’t explicitly confirm escalation, even though the conditions for escalation were met. The rules I set were: Escalate if the user shows negative or neutral sentiment. Escalate when the user provides all required information. Escalate when the user requests escalation. The issue is that when a user provides all the information and FIN sends the escalation message “An agent will be with you soon,” it later follows with “Is there anything else I can help you with?”If the user responds with something like “No,” FIN ignores the first escalation message and treats the user’s “No” as a signal to close the chat. Instead of keeping the ticket open for an agent, it ends the conversation.

Hi everyone, We’re seeing Fin frequently sending a follow-up message after responses, like:“Did that answer your question?” / “Was this helpful?” In many cases it happens almost after every message, and it can feel repetitive for customers. Is there a way to control the frequency of this follow-up (reduce it, change when it appears, or disable it)? We’d like more control so it doesn’t trigger too often.

I would like to setup for all of 2026 for FIN give advance notice when a holiday office closure is approaching and then set proper expectations for human responses during the holiday.I don’t want to create a separate workflow for each holiday. Appreciate any suggestions, thanks. 

Hi everyone,We’re using Fin with an escalation rule that hands the conversation off to a human when certain conditions are met.Right now, when escalation triggers, Fin sends the customer a fixed handoff message that repeats word-for-word across conversations. The phrasing doesn’t match our support tone of voice, and it feels robotic.Is there a way to control or customize the customer-facing text that Fin sends at the moment of escalation/handoff?  

I want to establish a consistent daily ritual where I review Fin AI performance and make meaningful updates (I currently make content updates based on Fin AI DSAT ratings).  I’ve already seen this post.What Intercom features do folks use in their daily maintenance of Fin? How much time do you spend on this daily? Looking for ideas!  I still do sprints outside of the daily routine to explore certain Intercom features, so this would be in addition to that.

We are a SaaS environment where FIN supports two separate products that are able to be connected to one another.  We are working with audience content targeting and other documentation flows and guidance to ensure the proper resources and content are sourced for the correct inquiries.  We’ve also had issues where changes were implemented before we were able to ALL sourced content updated, and instruction was given that couldn’t be followed. Has anyone come up with a way to track conversations that specifically are giving information the customer is unable to follow?Current Idea going over this afternoon is to make an attribute, and have the key words and things often asked include “its not there” “i’m unable to do that” “that is missing for me” and other example statements we’ve seen in chat, and have that be a specific attribute used in an escalation guidance, and get those customers to reps to verify their environment. But would also then be able to get rep

Hi community, I'm trying to set up Fin to provide different responses to different audiences using audience-targeted Snippets, but Fin is completely ignoring them and not using them at all. What I've configured:• Created specific Audiences based on user attributes • Created Snippets with different content for each audience• Assigned the appropriate Audience to each Snippet using Audience Targeting• Verified that users match the correct Audience criteriaThe problem:Fin is NOT accessing the audience-specific Snippets at all - it completely ignores them and uses general content or other articles instead. Even when the Snippet is perfectly relevant to the question and targeted to the correct audience, Fin doesn't use it.Context:We're a fintech company with complex financial products. It's critical for us to have a SCALABLE solution that ens

Earlier, under Fin AI Agent → Train → Content, there was a Custom Answers tab within Content Sources, where I had added several custom answers.I no longer see the Custom Answers option in this section.Could you please clarify: How can I access the custom answers that were previously added? Are these existing custom answers still being used to train the AI agent? What is the current process for adding new custom answers?

I would love to have an option to involve in only for collecting conversation Attributes. Right now it is not possible and any workaround works.  Even if I create a special tag/inbox for collecting attributes Fin must be involved which means even if it meet the escalation rule, it will reply that the conversation has been passed on. I would like Fin working in the background, collecting attributes from ALL conversations.  Maybe managing the Fin’s steps (in Let Fin handle nod in workflows) would be a solution. I mean turning off using content, guidances and leaving just collecting attributes would do the trick.  

Is there a way to have the Fin AI agent reference information (including hyperlinks to PDFs) on a website but restrict it from linking customers to the site? We have a page with links to user manuals for devices listed on our website that was meant to just be for AI reference rather than for public consumption. We want Fin to be able to answer a question troubleshooting use of a specific device type with the user manual, but don’t want to take customers to a page with a huge table of devices on it.

Hello! I wanted to report a couple of issues we’ve been experiencing with the FIN Suggestions feature: Suggestions tab showing empty results – We sometimes see that a suggestion appears to be available, but when we open it, it’s actually empty (no comments or text). Article changes not saving – When trying to edit an article through FIN Suggestions, we often get an error saying the changes could not be saved. In some cases, the changes aren’t even saved as a draft, which has caused us to lose work and time. Could you let us know if these are known bugs or if there are any current workarounds?Additionally, we were wondering if there are any options to guide or rate suggestions so the system can learn from our feedback. At the moment, many suggestions seem repetitive and often just reword content without explaining why the original version might not be good. It would be great to have a way to indicate which suggestions ar

Hi Intercom Community,I’m trying to fine-tune how FIN handles unresolved issues. Specifically:Is there a way to set a limit on how many times FIN tries to resolve a customer’s question before escalating? Can FIN automatically hand off to a human agent after those attempts?I want to make sure customers aren’t stuck in a loop if FIN can’t help. Ideally, it would try a few times, then gracefully pass the conversation to a person.Has anyone set this up successfully? Would love to hear how you approached it—or if there’s a built-in way to do this.Thanks in advance!—Vanessa

Hey everyone 👋We’re currently using Fin as the first line of support, handling all incoming customer conversations.It works great — once a human agent joins the conversation, Fin automatically stops responding, which makes perfect sense in most cases.However, there are many situations where the agent realizes that Fin could actually continue handling the customer effectively after their short intervention.It would be super helpful if agents could reactivate Fin in a conversation — ideally with a simple button or command — so Fin can pick up right where it left off.Is there a way to do this today? If not, we’d love to see this added as a future feature 🙏Thanks!Tamir

We utilize Fin as the first line of defense in supporting our software. I’d like to explore having Fin handle chats related to known bugs/issues. Right now, a user will message saying “I’m experiencing XYZ issue” and Fin will go through the normal troubleshooting steps. Ultimately these chats often get routed to a human agent who then lets the user know we are aware of the issue and working on a resolution. Users are often annoyed that they had to go through the process of explaining the issue to Fin just to find out we knew there was an issue. Sometimes these issues only impact a small subject of our user base so we don’t want to widely exclaim that there is an issue. Thoughts/Questions:Best way to implement something like this - I’m curious if we can utilize the Content and Sources (beta) section under Guidance? Best way to track these conversations - we track these reports so we can reach back out to the user once the issue is resolved. Ideally, these

I’ve been back and fort with the support team on this for weeks, and made sure to set up Finn correctly. I have this very simple instruction asking Finn to not reply to my customers if it doesn’t know the answer to their question. No matter how I phrase it or how many ways I repeat it, Finn will “behave” for a while and then go back to replying, mostly embarrassing nonsense responses.It’s ridiculous! I keep having to apologize to my customers for the dumb responses it will give to super basic questions. See below my current instructions - I can’t think of another way to repeat this… I keep adding them.Is Finn just not ready for real world usage?Or is it just my account that’s cursed? I can’t keep coming up with different ways to phrase this simple instructions (and the support team can’t either, aparently).  

Friends! I LOVE the suggestions feature with Fin! I want to try and accomplish something that I feel would be table stakes but not sure how to go about it: I’d like to filter tickets we got last year within a specific date range (our customers are accountants, so for tax season), then I’d like Fin to aggregate and analyze the trends of the tickets and suggest where our content gaps are so that we can fill them preemptively and flatten the curve for the upcoming season. Anyone know how I might go about that?! Best, Malachi

Is there a way to train or edit customer sentiment within CX Scores assigned to a conversation? For example, I reviewed a low CX Score conversation and it was determined the customer was frustrated, but that was not the case when reading the customer's response. Any tips to reduce the number of low customer sentiment scores?