Support Stack E15: How Fin's AI Knowledge Manager Uses Operator Memory with Dawn Perrott (Fin)
In this episode of Support Stack, Dawn Perrott, AI Knowledge Manager at Fin (formerly Intercom), walks through the Operator feature that reshaped her role: memory. Before Operator, being a knowledge manager meant carrying a mental map of the entire knowledge base. Now that map lives in Operator's memory — so the basics like your style guide, terminology and standards are applied automatically to every content update, rather than pasted into a prompt each time.
Dawn shows how to save a style guide to memory, how to see and delete what's stored, and how to tell a useful memory entry from a useless one. Vague rules like "always write high-quality content" don't work, because Operator can't act on them or check against them; product functionality and pricing don't belong there either, since they already live in your knowledge base. She also demonstrates the "echo principle" for FAQs on a live article — a small change that makes a big difference to how well Fin retrieves and uses your content. Throughout, the message is clear: Operator handles the retrieval and drafting, but the human judgement stays yours.
If you own Fin content, run a help centre, or you're an Intercom admin trying to lift answer quality, this is a practical look at getting more out of Operator — the how and the why behind the feature, not just the what.
🔗 Resources mentioned:
AI knowledge management workflow — from setup to content readiness
Optimise your help centre for Fin AI Agent (includes a downloadable Style Guide example)
Want daily, practical tips on getting the most out of Fin and Intercom? Subscribe to my daily email list.
Episode transcript
Conor Pendergrast (00:00)
Hello and welcome to episode fifteen of support stack.
So, Dawn Perrot, you are the AI knowledge manager at Intercom. Welcome to the support stack. How are you doing today?
Dawn (00:18)
Thank you, Conor I'm doing really good today.
Conor Pendergrast (00:21)
I just realized I said Intercom, but I meant Fin because we are recording this long after Fin has been renamed from Intercom. So I apologize, Fin. I apologize, Intercom. Apologize everyone, but but this is just your reminder that the company's called called Fin again. Just in case you've forgotten. Anyway, so so don't we we get to talk about something very fun today, which is a relatively newly released feature called called operator. So what
Dawn (00:24)
it's better.
Conor Pendergrast (00:50)
Tell tell me a little bit about Operator and what what's your favorite parts of Operator?
Dawn (00:55)
So today we're talking about Operator. specifically, one of my favorite parts of Operator is its ability to hold memory. previously the knowledge manager role meant having to have an encyclopedic mental map of the content that lived in your knowledge base. But now, because of operator's functionality and memory,
it it is no longer the case that I need to remember the entire knowledge base. It is all now housed in operator's memory and I'm able to make workflows significantly faster because of this.
Conor Pendergrast (01:43)
so operator is a bit like giving you as an AI knowledge manager like an external brain or a superpower.
Dawn (01:50)
a hundred percent. A hundred percent.
Conor Pendergrast (01:53)
That's fantastic. So like I think what would be useful I know I've used Operator quite a bit. I suspect that there are people who have access to the beta and haven't really known like the ways to get the most out of Operator. And that is hopefully what you're here for for the next two or three episodes. So what are we going to talk about today, Dawn? What in particular?
Dawn (02:17)
So, first of all, I'm gonna dive into memory, how it works. And just to give some context on what memory is in terms of operator, it's the difference between using operator as a one-off tool and working with it as a collaborator. When it is set up well, you stop prompting for the basics, the style, the terminology, the standards. They're just there living in operators' memory every time. So I'm gonna show you what I've saved.
why each entry earns its place and what I've learned not to put there also.
Conor Pendergrast (02:54)
That's fantastic. Okay, let's dive in. So let's get you sharing your screen first, I guess.
Dawn (03:01)
Cool. So I think how I approach memory is like onboarding a new teammate. It's everything that I would tell them when they'd start in a knowledge management or content management role. So I suppose to start off, I'll actually just give a quick overview on how to to save memory to operator. So I've created a
AI knowledge manag management workflow article for setup from setup to content readiness. So this is an overview.
Conor Pendergrast (03:33)
And it is
super fresh as well. Cause it's I see it's last updated in the last 15 minutes, and you don't get much fresher than that.
Dawn (03:40)
I updated it just before we met. So yeah. and and I'm sure there will be more updates on this as well when we when this is out of beta and released all. So if you just scroll down the article, it's highlighting what knowledge management was before and now what knowledge management looks like because of operator and how operator has changed the knowledge management role.
Conor Pendergrast (03:42)
Perfect.
Dawn (04:09)
And then I have this section, the setup of memory and style guide. So I think this heading is very poignant: memory giving operator brain that persists. So instead of you having to paste your style guide into the prompt or into the composer every single time, memory is the thing that's stored there, stored in operator. So you don't have to prompt it every time you're making a content update. It lives there, it has access to it. It makes things much easier.
So I have an example of a few things that are saved to the memory. So I have a style guide, I have a content readiness framework, naming conventions, and workspace context.
Conor Pendergrast (04:51)
brilliant.
Dawn (04:52)
And then these are the sorry scrolling down to the steps. So how to save your style guide to operators memory. So initially you'll have to start off by going into Finii Agent and Operator, and you'll see this composer here. And then in the composer, you paste in a message. I just have an example in the article here. You paste in your message, you add your
Conor Pendergrast (05:06)
Mm.
Mm-hmm.
Dawn (05:22)
This is just my style guide and I'll ask them.
Conor Pendergrast (05:25)
Perfect. So that's
that's a PDF that you had on your desktop, but a lot of people will have a style guide already developed and won't have thought, we should get this into Operator so Operator knows our style guide.
Dawn (05:36)
Exactly. And then if you're making content updates, your brand voice is already applied to proposals. So you're not having to go in and change content regularly. It's already going to be in the voice that you want it to be. and then I will just interact with operator and ask it: can you save this to your memory? So operator will start thinking and
Conor Pendergrast (05:47)
Perfect.
Dawn (06:03)
uploading this to its living memory. I've actually already saved it. So
Conor Pendergrast (06:11)
Yeah, the the tricky I know from from experience there's a bit of trickiness when you're when you're working in in test workspaces or real workspaces as well to get those first setup steps demonstrated.
Dawn (06:22)
Exactly.
Exactly. So I do have I have the all the memory saved here. but I'll wait for I'll come back to this in a second and I'll here we go. So you know you'll know when your style guide is saved when you get the message save for future conversations, I'll apply the style guide whenever creating or editing knowledge-based content. So now operator has the rules applied to it for when content is getting updated or created.
Conor Pendergrast (06:28)
super.
Dawn (06:54)
it's super helpful. If you if there's ever a time where you're you're curious as well what is saved in the memory, all you have to do is prompt operator what you have saved in your memory. And it will give an overview of everything that's saved when it was last updated and a summary a summary of what it is.
Conor Pendergrast (06:53)
That's super.
Mm, okay, excellent. So I wouldn't I wouldn't have thought to to like explicitly ask operator. And maybe maybe a small bit of of feedback for the operator product team is to say like, hey, there there should be a way of showing of seeing without having to know to prompt operator somewhere somewhere in here what the memory what the memory is.
Dawn (07:33)
That's really good feedback. Yeah. I think I think this is something that we can potentially expect f when this goes to GA that there'll be a draw for memory specifically.
Conor Pendergrast (07:40)
Yeah.
GA being general availability. Super. So I'm I I'm curious as well, Dawn, this is this saved just for you? Is it saved across the whole workspace?
Dawn (07:44)
Yes, general available to
So this is saved across the whole workspace.
Conor Pendergrast (07:56)
Thank God. 'Cause otherwise if everyone had to s updo the st upload the style guide, that would be very frustrating.
Dawn (08:02)
For sure, for sure. And then there'll be people who will have different versions. So it's it's good to have one consistent memory that applies to everyone.
Conor Pendergrast (08:14)
Absolutely. Okay.
Dawn (08:15)
Yeah.
once you have the style guide saved, you'll get the message. And then like I was saying, you can you can
Conor Pendergrast (08:24)
Mm-hmm.
Dawn (08:27)
verify that it has worked by asking operator to cross-check it against one of your already existing articles.
Conor Pendergrast (08:34)
of course.
Dawn (08:35)
And then it will output something like this. So you'll get what's passing and the issues that are found. I have an example that I previously looked at. So okay, so yeah, this is from the screenshot from the Help Center article. So it exactly that you can see what's doing well and where the issues are found. So AI isn't actually able to tell.
isn't able to read emojis necessarily. So they're not they're not really if they're pointing ones, they're a bit ambiguous. So that's one of the issues that were flagged. And
Conor Pendergrast (09:17)
interesting. so it's pointing so the style part of the style guide that you have set up is to say like you shouldn't use the emojis. Is that correct? Or you should use emojis only in specific circumstances?
Dawn (09:29)
in specific circumstances. We actually have a article in the help center for optimising fin. I can share that with you as well. And there is a list of what to do and what not to do. And we actually have our own interco we sorry, we have our own style guide applied. we have it added as a PDF in that article.
Conor Pendergrast (09:54)
I do remember that. So and this is going to be obvious though, but because you're using Fin's operator, f Fin's operator, I presume already has access to Fin's help content as well and will know like, okay, well this is what makes a good a good resource for Fin, the AI agent, to be able to interpret correctly. Because you have to one of the tricks, I'd I don't know what your perspective on this, Dawn, is, but I think one of the tricks is
AI or knowledge management, content management has traditionally served like one to two audiences, our customers and our teammates, but now we've got this hyper-critical third audience, which is AI agents, and they read content in a very different way. Have you found like have you found similar to the emoji thing? The emoji the emoji, the stripping out emojis.
Are there other things that you've had to do do in the style guide or outside of the style guide to account for the fact that there's a new audience here?
Dawn (10:53)
Yeah, for sh like when I'm creating content, I am always
thinking, how does this flow for the customer experience and how does this work for the AI agent who'll be retrieving this content for answering purposes? What I have found is the human will often push through. So the person reading the article will figure it out, whereas the AI needs to have explicit yes and no, what works, what doesn't work. So there are different
there are different factors that we have applied to ensure that the experience is the experience for both humans and AI is what it needs to be. It's very symbiotic relationship between the two. different things like echo principles in FAQs. So if you have FAQs in your content and you're asking a question, in the answer you need to reiterate what content there is in the question. So when a customer asks that.
the AI will be able to pinpoint
the FAQ to answer with.
Conor Pendergrast (11:58)
Hm. Can can you can you think of an example just to make that really, really clear for for people?
Dawn (12:05)
I can show you an example.
Conor Pendergrast (12:07)
that's even better. Show, don't tell, is even better.
Dawn (12:10)
So I don't have any FAQs in this, but even let me see. So
Any
Conor Pendergrast (12:21)
excellent. So we can show operator running live and going and checking the knowledge center in this in this workspace. And of course this is a this is a test workspace, right, Dawn?
Dawn (12:29)
I'll just
Yeah, I was just going to caveat with that that this is a test workspace. So the content in here is for an example project management tool. I'm not sure if we have a FAQ's article, but let's see anyway. Okay, so let's go with this one. just on those prompts that you probably saw operator generate and I had to choose one. Initially, when they when I was being
prompted to answer those, I was frustrated because I I wanted to be fast. But then I re then realized that it's really important to answer those questions because operator is flagging something ambiguous and it needs to know what is the right thing and what is the wrong thing. So that kind of ties in back into how to make sure AI agents will know what is correct and what isn't. It's
it's those questions that operator will ask that is clearing up any miscommunication or misalignment.
Conor Pendergrast (13:42)
Absolutely. Okay, cool. So we've got two examples.
Dawn (13:46)
So
here I have found two pieces of content with FAQ structure. There is no QA structure at all. It is bulleted troubleshooting list covering field mismatches. Okay, so this wouldn't be our usual setup. this is just because it's on a test workspace, but let's look at this one. Article has a proper FAQ section at the bottom with two QA pairs, neither echoes the question. So
Can I edit project details after creation? Yes. Go to project settings to update the name, description, team members, or other details. So you can see the issue here is that it's jumping straight into the instruction. It doesn't restate the yes editing is possible. So I can now prompt operator to fix those FAQs that they adhere to follow to the echo principle.
also this is just one principle. We have the content readiness framework, which is a list of fourteen factors that we'll touch on later. But just
Conor Pendergrast (14:51)
Yeah, that's for our next episode, which I'm also
really looking forward to. So I guess you're getting a preview of while we're showing you how to use how to use operator, you're also getting a preview of how to of one of the the content principles aren't you, aren't we? Okay, cool.
Dawn (15:09)
So this actually doesn't look how I want it to look. no, sorry, here we are, FAQs. So can I edit your project details? And you can see that it's added in this. You could actually
Conor Pendergrast (15:15)
There we are, yeah.
So
yeah. Go on, sorry, Dorne.
Dawn (15:24)
I was just going to say that you if you're not happy with some wording in the generated response, you can go in and edit it. So I can be I can remove this and then I can save it and then apply that.
Conor Pendergrast (15:32)
cool.
Yeah. that's great. Okay. So that's that's a this is an example of number one, the echo principle, but also number two, using operator to improve your content in the knowledge base in order for Fin to do a better job of interacting with it and retrieving it and answering customers in these pretty simple content questions. That's super.
Dawn (15:56)
Yeah. Mm-hmm. Each memory,
each memory that I've saved does a specific job. So you'll have the style guides, the terminology rules, and the content readiness framework. there also are bad examples of memory. So yeah. So when you are writing instructions,
Conor Pendergrast (16:14)
Interesting. Okay. Tell me more.
Dawn (16:21)
it's best not to be too vague.
always write like using an instruction like or using an instruction like always write high quality helpful content that customers will find useful. This feels like an instruction, but it isn't. Operator can't do anything with high quality, it's not specific enough to act on, and it's not specific enough to check against. If you can't imagine operator failing this rule, it's not worth saving. So I've actually created a test.
Of these memories. So these are the ones not to go for. So here's an instruction for your memory: always write high quality, helpful content that customers will find useful. It's too ambiguous. it's not specific enough to actually be helpful for your operator. And then
So another bad example is saving your product functionality to operator's memory,
which isn't necessary. Operator doesn't need to remember what functionality your product does, or operator doesn't need to have saved in its memory, this is the pricing plan for this customer, because that's already a living.
artifact in the actual knowledge base operator has access to it. When you add that into the memory, you're only clogging up the space in there. And I'd also advise people to have audits over their operator memory every month to make sure that everything is aligned because that's being checked when content's being created, procedures are being created, guidance is being created and monitors. So it's important to make sure that it aligns with
it aligns with your company values and how you want to
communicate with your customers.
Conor Pendergrast (18:16)
Great. Absolutely. So okay, so if someone has been using Operator a little bit and realized that they've given some memory s instructions in memory that they that didn't now that they realize it doesn't make sense. For example, they've said always write high quality help content that's useful for customers. And now they realize, well they mean they need something more specific, like the 14 factor content readiness check that we're gonna talk about in the next episode. How would they you've shown how to how to find
your your memory, you can just ask operator what what are my memories? Is that also how you get rid of memories?
Dawn (18:51)
Yeah, so in this example, I had added the two bad examples for instructions. And then I simply said, can you remove them? And it operator said, Do you want me to remove both these? And I said yes. So delete both. And then I got the confirmation that both of those were deleted. if if you want to see what's in operator's memory, all you have to do is ask what you saved in your memory. And
Conor Pendergrast (18:57)
Mm.
Super.
Dawn (19:20)
It will give an overview of everything. So in this test working working space, I have the content readiness that we'll touch on later. I have knowledge base do's and don'ts. So this is where it's more specific rather than write high-quality content. This is me telling operator, I want your writing for both humans and AI. you're to use terminology and title heading and headings. So rather than having ambiguous headings like
Conor Pendergrast (19:27)
Mm-hmm.
Dawn (19:50)
the feature name, phrase it in phrase it in how a customer would ask about it. So how do I assign a task, for instance? Using numbered lists for steps, bullets for items, never describe describe sequences in flowing pros, making every section self-contained. That's just an example of the do's. And then the don'ts is telling operator to never construct or guess URLs in in generated content.
Conor Pendergrast (20:20)
Mm.
Dawn (20:20)
using vague spatial references, so this screen, pointing emojis, using bare imperative or bare noun headings. So the this is just an example of what not to do.
Conor Pendergrast (20:32)
Yeah.
Yeah. And this and there there might be other companies who look at this list and they think like, No, we we are okay to say we when we when we mean the product, for example. That that's there are parts of these that are stylistic as well.
Dawn (20:48)
Of course. It depends on how you want to what tone of voice you want for your for your company.
Conor Pendergrast (20:54)
Yeah. This is Super Dawn. Is there anything else that you want to show us about operator today?
Dawn (20:59)
for memory, that's really everything. so that's just the content right in my in my memory that we have saved is the content readiness framework, the style guide, the do's and don'ts. we have other rules as well for internal when when in creating internal content. and that's just how to communicate for the customer support team, what structure to follow. It's
It's essentially another style guide but for internal content.
Conor Pendergrast (21:31)
Okay, brilliant. Brilliant. Well Don, this has been really useful. I I'm I have lots of ideas for how I'm gonna go to to some of my clients and set up their operator to work a little bit better as well as my own my own operator, so I'm excited to to to get some good memories out of this episode.
Dawn (21:49)
Very good. It's a
it's a great tool and it's only getting better.
Conor Pendergrast (21:52)
Yeah, absolutely. I mean you use this every day and and I'm I'm excited we've got two more episodes to record and to to share with people and all about how how you how you find operator useful to to improve AI AI knowledge management at at Fin. So that's brilliant. Dawn, where where would people find more about you? What what what do you think that they should do next?
Dawn (22:17)
If anyone wants to follow along with more AI knowledge management tips, they can follow me on LinkedIn.
Conor Pendergrast (22:24)
Super, and how will they find you on LinkedIn?
Dawn (22:29)
my Google Google Works or I believe you're gonna add the link to my LinkedIn in the description.
Conor Pendergrast (22:29)
Have a Google, maybe?
Perfect. I will. I'll put a LinkedIn a link a linked a LinkedIn link in the YouTube description, just in the down below. Just cast your eyes or your fingers down down below people and and follow Dawn on on LinkedIn. She's a very good very good, very helpful, very valuable follow. Dawn, this has been great. Thank you very much. thank you to viewers. You can find me at customer success.cx slash daily if you want my weekdaily
in weekdaily tips and hints, which actually have talked about Operator quite a bit in the last couple of weeks as well. all about getting the most out of Fin, AI agent and and Fin the the company. And yeah, find us next time for for the next episode. But for now, yeah, have a wonderful rest of your day. Bye.
Dawn (23:27)
Thanks. Bye.