Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Spotify | Android | TuneIn | RSS
At Trusted Shops, Claudia Francesca Mueller and her colleagues have built an AI-powered chatbot called Piuma that lets non-writers access content guidance through a natural-language interface.
It took just a few weeks to launch the initial version of Piuma, building the chat interface with Voiceflow and using the LangChain development framework to access both their content design guidance and OpenAI’s API.
Even though the chatbot’s functionality matched their users’ expectations almost perfectly, they still find that they have to constantly collaborate with their partners to fully understand their needs and communicate the benefits of the product.
We talked about:
- her work as a content design and localization lead at Trusted Shops
- Piuma, the AI chatbot they have built at Trusted Shops
- how Piuma arose from research and discovery work they did around how to best share their content-design guidance
- how they developed Piuma using Voiceflow with guidance from a conversational design expert
- the learning curve around incorporating LLMs into a chatbot like Piuma
- how they decided which parts of their voice and tone guidance to include in the chatbot
- how Voiceflow works with the OpenAI and Langchain
- the need to sometimes adjust the source documentation that the LLM is consulting to get the answers you want in the chatbot
- how her multilingual background helps her understand computer languages
- the challenges of getting designers to adopt a new tool like Piuma
- her ongoing communication with designers to understand their needs and how to address them
- how she balances evangelism and outreach with collaboration around improving Piuma
- the tendency of humans to stay with familiar patterns and routines
Claudia’s bio
Claudia Francesca Mueller is a multilingual content designer living in Amsterdam. She speaks Swiss-German, Italian, German, English and Dutch daily, and feels at home when languages are mixed up in one sentence. That’s how she has also learned to bridge culture gaps with style and the right tone.
Her love for languages, words, culture, shapes and colours brought her to content design. A discipline that became her passion and that she loves to live as an expert, leader and coach.
Her background and career in multiple content roles have strongly shaped her thinking. She believes content is a holistic discipline where words are only one of many tools to convey a message.
Currently, Claudia works at Trusted Shops as Principal Content Design and Localization.
Connect with Claudia online
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Content and AI podcast, episode number 7. On any one digital product team, there are never enough content designers or UX writers. So when interaction designers or engineers have to write interface copy, they typically have to consult content-design documentation. AI creates new ways to share this kind of content guidance. At Trusted Shops, Claudia Francesca Mueller and her colleagues have built an AI-powered chatbot that lets non-writers access content guidance through a natural-language interface.
Interview transcript
Larry:
Hi, everyone. Welcome to episode number seven of the Content and AI podcast. I am really delighted today to welcome to the show Claudia Francesca Mueller. Claudia is a multilingual content designer and a localization lead at Trusted Shops. Welcome, Claudia. Tell the folks a little bit more about what you’re doing these days.
Claudia:
Hi, everybody. Thanks a lot for having me, Larry. As you said, I’m a content design and localization lead at Trusted Shops. Trusted Shops, for the ones that don’t know what Trusted Shops is, it is a German e-commerce software as a service company. We certify shops. And if you’re a trustworthy shop, you will get a badge and you can start collecting reviews. We offer a review system management to the customers or shop owners. And for the consumers, we offer a buyer protection. And I’m part of a UX team at Trusted Shops. We are seven designers at the moment. And in the content localization team, we’re five people.
Claudia:
Together, I do the math, we’re around 13, 14, 15 people in the UX team. And I’m responsible for the content design craft at Trusted Shops. I jump into, let’s say, projects that have a high impact on business, but I am also responsible to develop the craft. That means helping others in the team to write UX copy. One of the things that actually was part of this lately was developing a UX writing body called Piuma. That was one of the projects I worked on it this year. I can tell how we came actually to build Piuma the chatbot.
Larry:
That’s why I wanted to have you on. And also, something you just said, we talked about this a little bit before we went on the air, but your official title is principal content designer at Trusted Shops. And I think what you just described is a classic principle role that, “Hey, there’s this new thing. Got to do it. Somebody super experienced has to take this on.” I just wanted to get that back in there because it sounds like you’re genuinely doing principal-level work. But I’d love to hear the origin story of Piuma, especially the name.
Claudia:
Piuma is my cat. And she’s always here in my home office. And when we started working on this project, and I worked together with a lovely technical writer in our team, I said, “Well, we need a name.” And he said, “Well, we just call it like your cat, Piuma.” And I was like, “Yeah, that’s a good one because I really love Piuma and she’s always here.” That’s how we actually got a name for the project, but also for the chatbot. And the whole idea of this chatbot … Well, I have to go back in time a little bit because when I started Trusted Shops, we were talking about consistency. There was no consistency. There were a lot of voices, casual, funny. You could see whatever you wanted to see just in our content, in our UI. We thought, “Well, we need consistency, so let’s develop a voice and tone guideline.”
Claudia:
And that’s also what we did. But after just doing all this work that actually took a year because, as I said, I have other tasks, that was just side project, we implemented those, so this voice and tone guideline, in the design system, in the existing design system, so we could call it content and design system. Also, actually for advocacy reasons, of course, when it comes to content design. The next step was like, okay, so now we have this documentation, but it’s just a documentation, so how do we get people using this documentation? Because we all know it’s not really attractive to have a documentation lying somewhere. And people just forget about it. How can you do that that just designers that actually are busy mostly with other things, more the visuals, how can we get them using this documentation? We thought about a lot of different softwares and we involved the designers as well in this process.
Claudia:
We thought about Ditto and Frontitude. And somehow, all of these softwares were not convincing. Designers were like, “Yeah, we don’t know. It seems like a lot of work.” And so we were like, “Okay, so if this software are not the solution or are not actually what you need in your daily life on your daily work, let’s find something that works for you.” We did a survey and we did research with the designers and OpenAI came up. One of designers just said literally, “Wow, it would be so great to just have a plugin in Figma and we can just ChatGPT.” And then I was like, “Hmm.” And then at the same time, I saw somebody from Expedia that was posting on LinkedIn about a chatbot. And I was like, “Okay.” I just combined … I was just making the connection. Just have our documentation as a knowledge base and data that we can use actually for the chatbot and we build a chatbot in Figma.
Claudia:
That’s actually how this whole idea started on this project. And everybody was like, “Oh, yeah, that’s such a great idea. Let’s do this.” Of course, there was missing knowledge in this whole project that I couldn’t fill because I’ve never worked … I’ve never built a chatbot. I know OpenAI, yes. I slightly knew what the LLM is, but how do we connect this data? How do we get a chatbot, the user interface? We needed an expert. And luckily enough, I had one at hand here in the Netherlands and a conversational designer expert. And so we asked her, “Would you be interested to help us for this project?” And she was like, “Oh, yeah, that’s such an interesting project. I really want to help you.” And so we hired her and she helped us with the first MVP and the concept of talk to your style guide.
Claudia:
And so she recommended to use Voiceflow. That’s one of the programs out there. It’s a commercial program, of course. Platform that you can use and that you actually … It’s actually really accessible because you’re going to work especially on the conversational flow with blocks. And it’s really intuitive, so it’s easy actually to … I would say for content designers that are not really technical yet, that you just work on the conversation, let’s say. And of course, there’s prompt engineering involved because you have to tell, let’s say, the LLM and the whole data what it has to do, that you get the answer you want. But still, for starting up with conversation AI, I think, and building a chatbot, it’s a great program. We started working on that together also with a technical writer that helped me with the data. It’s really easy.
Claudia:
You just upload a file or website. It’s like the most easy you can think of. And then you’re just busy actually with the conversation. And of course, when you start working with it, you get curious because you want to know how it works because if you know how it works, you probably can tweak the data better and you get a better result. Of course, it’s easy to use, but I think it’s also really natural that you get more interested in the backend of this whole chatbot. What I can tell is that Voiceflow has LangChain under the hood. It’s basically LangChain is a open source software together with OpenAI that actually in the end, you have an LLM that you can train with your own data and that you can actually tweak to get the right response you want.
Claudia:
And we built this chatbot, I think, in two or three weeks. It was not like the building part. And if you can work on it continually, then I think you’re quite fast with building because you can copy and paste the blocks. If you know once how one flow is working, you can just duplicate your blocks. And what we did is, we didn’t build the whole voice and tone guidelines now in the chatbot. We thought about, “Okay, which are the most used content components that designers need?” Also thinking about the process being efficient and enabling the designers. Also, it’s not a chatbot like ChatGPT where you just write, “Hey, give me an error message,” because I also wanted them to understand what they’re doing. What is it? Is it a specific error message? Or is it an unspecific error message? What is an error message? I wanted them also to have an educational part in this chatbot and not only, well, “Give me the answer, chatbot.”
Larry:
And you’re prompting a lot of questions as you go. That’s a great overview, but I want to go back to a couple of things that I got curious about as you were talking. One, the integration of your writing guidelines. And so Voiceflow is the tool that you’re using to build this into blocks or the model that they use, is your components to chunk all that together. But where does OpenAI and LangChain come in in terms of are they the chat interface, the conversational part of it? Or how … Tell me a little bit about how those parts work together, Voiceflow, your own corpus of writing information, and OpenAI. And quick thing on that, I assume you’re using the OpenAI API access, not just ChatGPT?
Claudia:
Yeah. Well, Voiceflow is using it. Voiceflow, the conversational block part is what you, in the end, build for the UI of the chatbot, so that’s, let’s say, the visual parts. Of course, it’s what’s happening in the background of the conversation, but LangChain and OpenAI is behind all that. It’s more the data. The data … You upload the file. And basically, that’s converting into vectors and it’s all about numbers. And so the numbers … and OpenAI together … I’m not an expert, but OpenAI has, of course, much more data. Together, I think they’re capable to see more patterns and actually get a better result. But don’t ask me in detail how this whole data is coming together and how they really … I don’t know about that.
Larry:
No. I’ve had other guests who are the whiz kid engineer on the backend who can explain vectorization and embeddings and then the math behind it. Actually, conceptually, it’s not that hard. But boy, the math is … That’s why I got into journalism, to avoid all that stuff. But hey, the other thing you mentioned is tweaking, making adjustments as you go. And you mentioned prompt engineering. Are you also doing fine-tuning on the backend? Are you in the model itself doing fine-tuning as well as prompt engineering? Or …
Claudia:
Yeah. The thing is I go back into the document of our voice and tone that is actually on the Google Docs, and I just tweak information in there. And then in Voiceflow itself, I just have a look and then test. Does it now get what I mean? For example, there was a big problem with sentence case. We use sentence case. And somehow, it never got that we use sentence case. And I was like, “Why? It’s written in the documentation.” I went back to the documentation and I just added there, wherever I could, “Hey, we use sentence case.” Really stupid in the sense of you would never write this in a documentation. I also had to tweak actually the data source and make it, well, more simple or more understandable for the model because in the end, what’s happening is that words or sentences get numbers. That’s how it works. I tweaked that part, but then also, my prompt engineering or the prompts in Voiceflow had to be changed.
Claudia:
I had to specifically perhaps say, “Hey, based on the query you get, or request, we also always use sentence case.” I added just the sentence case as a word, as a term. And probably, that made it easier to match or to pattern, to see a pattern actually in this whole system for the machine. It’s about tweaking on the data source and actually also to make it more understandable for the machine. And I think it’s this kind of translation that is happening. And that’s also really interesting as a translator myself. I’m a trained translator and journalist, like you. Also, I was more into journalism than numbers and this kind of stuff. I found it really interesting to, again, translate things for a machine. In the end, you have to learn another way of thinking so that the others, like the machine, is getting what you mean. And I found that really interesting in this whole … The project itself is very interesting because I think there is just so much to learn. And I love to learn new things. And it seems I can learn a whole new way of talking.
Larry:
Well, I’m really curious now too, because you’re at least tri … how many … you speak a bunch … Because you were born in Switzerland, German parent, Italian parent, work in English. You’re just … I wonder, it’s kind of early days. There’s not a lot of people doing this kind of work, but I wonder the deeper we get into this, if people with a background like yours, then like you just said, well, just another the language. I’m talking to a computer now instead of an Italian person or a German person.
Claudia:
And it’s just really interesting because every language has just its structure and you just need to understand the structure and the basics and then you can play around with. That’s actually how I think about language. Also, I picked up Japanese again because I just like it because it’s so different how the communication … How we communicate in Japanese is just really different than how we communicate in all the European languages. I just think of it as another language I have to learn. And I found it super interesting. And also, if you go deeper, of course, I’m thinking, “Okay, how is this working with the numbers?” But there, probably I will stop. But thinking about coding as well because most of these models are based on Python, for example. I was thinking perhaps I should just also get into that to learn that as another language, but, well, first, I have to figure out better the prompt. And I think also just really how to talk to the machine together-
Larry:
I have a role model for you, one of Bala’s colleagues, our mutual friend Bala who works at IKEA, one of his colleagues, Dan Porder. About two or three years ago, he was teaching poetry and now he’s doing Python programming and doing deep AI stuff at IKEA. It’s not unprecedented, the path you’re curious about. Hey, I wanted to ask another thing about just the day-to-day usage of this. You go to … You’ve learned all this whole new skillset and tooling, and you’ve done a lot of design work ahead of time with the designers. And so you put it out there and everybody uses it just as it’s designed and planned, correct?
Claudia:
Yes. That’s what we wish for.
Larry:
Sorry.
Claudia:
That’s the ideal case, yeah. But no, unfortunately, that’s not happening, of course. And I think it’s just also something I have to learn, that everything I do … also, with the voice and tone guideline, when we got it out there first, I was like, “Okay, my work is done,” but no, the work is starting again and it’s a different kind of work, but it’s just never ever … There’s no finish line, honestly, on this. There’s just like … There is a lot of communication and advocacy actually in this. Also, when we had the voice and tone guideline, we started with workshops. I did a lot of workshops with the designers just to get them familiar with the content components. And so it’s the same with Piuma. We agreed that this would be actually a great solution for their needs in their daily work routine. And then we build it, and now it’s there, and nobody’s using it, so I have to start again and think about how do I advocate for Piuma. And also, to find out what are the reasons why you’re not using it.
Claudia:
I just talked to designers yesterday, and it was really interesting to hear that they really think this is the ideal solution, but somehow, and that’s actually not surprising, they say, “Yeah, I just prefer to just get a consultancy hour and talk to somebody.” And I was like, “I understand that, that you prefer human, the humans, instead of talking to a machine. On the other hand, we have to be efficient in what we’re doing. And we built this so that you can be more efficient in your work.”
Claudia:
And also, seeing the resources we have, we have to think about putting the resources out there for the projects that have a bigger impact and not, let’s say, a really content component that is just a basic, I don’t know, error message or something like that. I’m not saying error message can’t be tricky. I’m just saying there is some basic kind of content components that would be great that we could master with the help of Piuma instead of somebody as a content UX writer or content designer that is actually putting time in it because we don’t have so many. That’s also something that I think it’s a change of mindset.
Claudia:
And also, you have people that embrace new technology and you have people that don’t do that so easily, or they might be thinking, “Oh, I don’t want to change my routine,” so it’s hard to say. There are different kind of aspects to this. It’s not like, “Oh, this is a fail,” and we stop the project. I think that’s the wrong way to go. And it can happen, of course, that, let’s say, in four months of now, I’m saying, “Okay, Piuma was a total fail. Nobody is using it. Let’s think about something else.” But I think until then, there is still a need of communication, advocacy, and also helping people to think that what we have here is something, a tool that is helping us. That’s all it is. It’s not taking our work. It’s not something that’s going to … It’s there as a tool, as an extra tool. And it’s actually a gift and we should embrace it and see it as something that is making our life easy.
Larry:
As you talk about that, I’m reminded of … I was talking to a friend the other day who did a presentation at a taxonomy conference about, “Hey, look, here are all the common excuses for not adopting AI or not being fond of this. You better get over that.” That was sort of his point in the slide. But I’m curious, have you run into that, what you just said, that sort of … It’s not really objecting to it, but just it is asking them to change their behavior and adopt new ways of working. Do you feel like it’s … I know it’s early days. It’s almost like you’re in the middle of a proof of concept kind of thing. But do you feel like … Is it just a matter of, not hacking the behavior, but just understanding better the user needs and going like, “Oh, if we did it this way, maybe you’d be more likely to use it”? Or …
Claudia:
Yes, of course. One of the designers was saying, “Yeah, but I feel it could be frustrating to use it because it’s not behaving like it should or it’s not giving me what I need.” And I was like, “Yes, of course not, yet.” But it’s like the chicken egg thing. I think we also talked about that, Larry, earlier on, saying, “Well, no, it’s not perfect yet, but it’s not going to get perfect if you don’t use it.” I really need them to understand that in the end, it’s like this is … It’s going to stay like this and I can’t put more time and effort in it. And of course, I want to meet their needs, but now I have done this work, now they have to do a step. It’s like it’s a team work in the end also. It’s not me being … Providing a tool. And I’m not piuma.com and that’s my company and I’m selling you this product.
Claudia:
It’s like, “Hey, we want to create a better product, for Trusted Shops, we want to create the best possible products we can. And this is what we build to help, to be better, to be more efficient.” We are all in this. And I think sometimes that’s really hard to understand that all we do, and that’s, again, also UX designers and content design, that all we do is we want the same thing. We want to build great products. And this is just helping us to be better and to focus on other things that are more exciting.
Larry:
Right. That’s the point of … that’s why we’ve always … That’s been the exciting thing about computers from day one, is like, “Oh, great, I can do the more interesting stuff here.”
Claudia:
Yeah.
Larry:
As you were talking to us now, was thinking about the … You talked a lot about the communication and advocacy around this. In that sense, you’re almost like a product marketing manager or something like that. But the way you were just talking, it sounds more like, no, we just need to collaborate better around this. Does that make sense?
Claudia:
Yeah, I think so. And I also think working more together on it and think about perhaps also involving them more in the next steps, that’s certainly something I’m thinking of. That’s why I also took the time yesterday to talk to them, and I will take the time next week to talk to them about it. Also, I don’t want it to become this red flag, like, “Oh, now we have to talk about, again, this chatbot.” And so it has to become also something they like and they enjoy working with. And I also hear, “Yes, it’s really cool. It’s exactly what we wanted. Yes, we want it in Figma.” And I’m like, “Why are not using it?” “Yeah, it’s just like … Yeah, I forgot.” Some just forgot about it. They just forgot that it existed. That’s also say something about how we work as humans, that we just are comfortable.
Claudia:
We just go back what we know. And the new things, oh, yeah, that’s extra effort. I might not do it because I want to be safe. I’m faster. So I just… And it’s an extra step and it’s an investment, but on the long term, it’s going to help us. And I think that sometimes that’s hard. It’s a step out of comfort zone, and it’s all what it needs. And that’s not something you can … That’s a mindset thing. And I think you can only talk about it. And it takes time also. It’s not something that’s happening tomorrow to change.
Larry:
In this world, it could be happening. That’s the thing.
Claudia:
Yeah.
Larry:
The whole AI world, things are so quick. Hey, Claudia, I can’t believe we’re coming up on time. But before we wrap up, is there anything last, anything we haven’t got to, or that you just want to make sure we cover before we wrap up?
Claudia:
I think I just want to say we should embrace what’s happening right now and see it as an opportunity to actually get better in what we’re doing. And also, just seeing that … And I’m really honest, I’m a really lazy person. And I think that’s why I embrace every technology because I think I’m just faster and I can get rid of the things I don’t like. We can get rid of – automate – things that are just boring work. Why not doing it? Just like some things are, for me, really obvious. And I think we should see it as something that helps us in our everyday life. But also, we have access to a world, or for me, that’s how it feels, that I haven’t before, so there’s a lot to learn. And by learning, we also get better in what we’re doing and probably we also will keep our jobs.
Larry:
That’s the ultimate goal because all this work is in service to something that leads to your employer giving you a paycheck. Thanks so much, Claudia. Hey, one very last thing. If folks want to follow you or connect online, what’s the best way to reach you?
Claudia:
LinkedIn. I think it’s the best way to reach me. That’s where you can find me.
Larry:
Excellent. I’ll put that in the show notes as well. Thank you so much, Claudia. Really fun conversation.
Claudia:
Thank you, Larry. And have a great day.
