Microsoft Teams Recording and Transcription Uncovered, Ritika Gupta, Microsoft Group Product Manager

Ritika Gupta, Microsoft Group Product Manager, delves into the power of AI in Microsoft Teams, focusing on recording, transcription, and intelligent  recap.

• Ritika highlights the shift of recording and transcription from convenience to essential for AI capabilities like intelligent recap and Copilot.
• Overcome language barriers with Interpreter Agent
• Real-time transcription challenges and AI improvements for more accurate language processing and cultural understanding.
• Compliance controls and policies for IT to ensure secure and effective use of recording and transcription.

Thanks to Luware, this episode's sponsor, for their continued support of Empowering.Cloud

Transcript

Ritika Gupta: And our job as product managers on these types of experience is to bridge the gap here where, uh, people who have to set guardrails, they have good enough granular controls to manage those guardrails. And then the day-to-day users, we don't, uh, enable, uh, increasing levels of friction. To, to get to you know, what they want to do, uh, and be productive in their environment on a day-to-day basis.

So,

Tom Arbuthnot: hi, and welcome back to the Teams Insider Podcast. This week we are going deep on recording transcription, live captions, and all that data being used by AI in things like facilitator, interpreter, and intelligent meeting recap. Really interesting conversation with Ritika Gupta, who works on the team that are involved in all those technologies.

Great to understand how it all hangs together, how the different LLM models are being used and also some of the Admin controls around recording and transcription. Many thanks to Ritika for taking the time to jump on the pod and also many thanks to Luware, who are the sponsor of this podcast. Really appreciate their support.

Great contact center attendant and recording platform as well. Hope you enjoy the show. Hey everybody, welcome back to the podcast. Really excited to have this conversation. I feel like this is the technology that is the, uh, the of core importance as we get into this Copilot and, and AI world. Got Ritika on the show and, uh, we'll get into what her and her team look after.

Uh, but should be a really interesting conversation. Ritika, you just wanna start by, uh, introducing yourself and, and your role. 

Ritika Gupta: Yeah. Thanks for having me, Tom. Uh, really excited to be on your show. I'm Ritika the group product manager for recording, transcription, interpretation, and intelligent recap for teams meeting.

Uh, I have been at Microsoft for 13 years. It's been an exciting journey because this wave of AI is just an amazing opportunity to be a part of and to write history together. So, very excited about this conversation and this right, uh, this AI wave. 

Tom Arbuthnot: Super interesting. I mean, like, I've been in this space for a long time.

You've been at Microsoft for a long time. I'm guessing, were you in the recording part of teams before the AI wave when it was just more, we need recording for recording sake. 

Ritika Gupta: Actually, I'm, uh, I joined this team just two years back, so just when the AI wave. 

Tom Arbuthnot: Okay, so you were just on the, on the start of the wave.

Awesome. I, I feel like it's really, uh, it's really interesting, like we've had obviously recording both for compliance and for convenience for Forever in UC, but suddenly there's an immense pressure on making sure that that is transcribing accurately because it's feeding into the, like, the intelligent recap and, and the Copilot stuff.

Um, is that, is that what you've seen as well? 

Ritika Gupta: Absolutely. Uh, it's been very exciting to watch how recording and transcription went from a feature, which was a good to have to like a necessity, uh, with this wave of AI because these are the two most important pieces of grounding data, uh, for any AI to be able to process a teams meeting, collaboration environment.

Uh, of course the quality of transcription is a very, very tough. Problem space to be in, uh, because there are these elements around people's accents, so many different languages that we have to support, uh, the tonality. Uh, even hardware plays a very important role depending on what type of microphones are people using.

So it's an extremely complex space, but, uh, we are all learning together as a team and writing history together because we, we are supported by our Azure Cognitive Services team who have a lot of experience in the speech domain. And we like given the interpretation experience that we recently launched, where there is a real time speech to speech translation happening, especially for meetings where people would need a human interpreter to be present because there is no common language that is shared, uh, that has been the best experience that we have built in the last, uh, recent past, together with the cognitive services team.

Tom Arbuthnot: Yeah, that's an amazing experience. I've, I've, I've had a play with it and, uh, the demos are coming out now and it's, um, like, like, uh, I feel like it's the first thing in UC we've had for a while that's genuinely like a net new ability and a real unlock. I, I work with lots of teams that are multinational and, and actually one of the scenarios is a lot of the global enterprises run.

It in English is the default language, but actually it's really great for people to have the option to turn interpreter on and have it back in their own local language to get full, full understanding of the conversation. In some cases, 

Ritika Gupta: I. You bet. Yes. Uh, I think when we started thinking about this problem space, initially we started off with a lot of features that we have in the product, which is like captions, interpretation or transcription.

However, very soon we realized as a product team that the real problem that we are solving for here is communicating without the language barriers. And as we become like more distributed orgs, uh, global orgs, uh, people have a preference because they definitely feel challenged, uh, speaking and conversing in a common language or even expressing themselves in a very comfortable way.

I. So some of the deep insights that we leveraged to in start investing in this domain was around the fact that how do we enable more equitable participation in meetings, especially in a remote, global way. And the, the beauty of a remote team is we can use these assistive tools real time. Uh, had it all been like we are all in the same meeting, sitting together, then it's a really hard challenge to work through.

However, since we are remote and we can use all of these tools in personal scope or group scope, it opens up the opportunity to bring in more, uh, equitable collaboration, uh, and communication going on in the meetings. 

Tom Arbuthnot: Yeah, I love it when the product team are living the product. I feel like teams got a huge boost, uh, during the pandemic, obviously, mainly because everybody had to use it, but suddenly.

Everybody at Microsoft had to use it. And I, I've heard stories of a lot of executive pressure being like, like now the, everybody at the exec level's using it every day. This needs to work. This needs to work. So it, it, it, living the product is definitely how you, you get to feel for, is it, is it working for real?

Ritika Gupta: You bet. Yes. And I think, uh, I especially work with a team in China. In Israel, and, uh, a lot of people, like, they don't, like, we, we don't feel comfortable talking in English because it's not our native language, including myself. So, yeah. Uh, it's a built habit that we built to acquire, uh, proficiency in the language.

And, uh, we believe that nobody, uh, should be holding back from contributing to these important discussions. Uh, where language is a barrier because that is right up the alley for AI to solve for, uh, where we can just converse in a language of our choice. And then AI will make the conversation much smoother, uh, for everybody else to follow along and contribute.

Tom Arbuthnot: So one of the things I've noticed using teams is there's, uh, the, the, there's the captions and then there's the kind of real time transcription, and then there's obviously the transcription after the event. Are those different things? I'm, I'm guessing technically there's an, it's easier to do. To take a complete file and produce a transcription versus doing it on the fly, particularly with some of the languages where context is relative to the end of the sentence and you're trying to transcribe in real time.

Ritika Gupta: You bet. I think that is the single most challenging piece because when we are doing real time, uh, I think for languages, as you rightly identified, transcription quality is better. If we have a longer context that we can transcribe on, it could be full, complete sentence or a couple of sentences. Uh, as we keep increasing the context, uh, the, the quality shows an automatic upward trend to begin with.

So we do offer both the capabilities, as you rightfully pointed out. Uh, we have captions and real time transcription and, uh, we also have the capability to kind of just, uh, retranscribe it after the fact, which is something, uh, that we haven't really exposed to the end users, but we are contemplating on it just so that we can improve the quality of transcription.

Especially now that it is used as the grounding data for all things AI Uh, however, it's a very tough choice. We still have to support real time because a lot of AI capabilities, especially like Copilot, is used in the meeting real time. 

Tom Arbuthnot: Hmm. 

Ritika Gupta: We have to have a reasonably good solution at good quality in the real time scenario as well.

Uh, so we have different optimizations that we do. But to your original question on are these capabilities. Extremely different. Well, the answer is not really. They use the very similar speech stack because whether it's captions or transcription, it's quite about, uh, speech to text. Uh, just that some of these are real time, some of these are configurable versus not.

Uh, we are also trying to see how we can streamline transcription and captions in the meeting. Because from a meeting consumption perspective, they provide very similar experiences with some subtle differences. But we want to bring these two experiences together so that the users don't have to choose between which one to consume, uh, during the meeting.

And then over a period of time, we, we see that transcription is going to become more and more a tool that AI leverages as opposed to somebody that would have to consume it real time because real time consumption should idly happen on captions. Because you want to follow along the conversation, but you're actually in the meeting as well.

Uh, 

Tom Arbuthnot: so yeah, I, I actually have a, like a, i I turn on transcription almost by default now because I want the summary. Um, but I don't really want transcription on my screen, so by default it pops up. I, I actually find it distracting having the transcription ticking along, so I'd rather have the default be.

On, but, but hidden to, to your point. 

Ritika Gupta: Exactly. And I think that is the insight that we also gathered, that the reason why people are using it real time is because we, we provided an option to do so. But as we dig into more insights into user behavior, uh, it's increasingly clearer for us that people actually appreciate captions.

They want a slightly longer context in the captions as well, so that. In case they missed something or their reading speed is different, then they can just go, uh, scroll up and down a little and gather the full context and the captions. Uh, and then in the future, we want to kind of move to a direction where transcription is going to be an artifact, which is purely leveraged for AI related workflows.

Um, and we can also make it more enriched and detailed where we can do a lot more annotation in the transcription. To make it more beneficial for the AI tools to use. 

Tom Arbuthnot: Yeah, I mean, in the case of, uh, teams by default, when you are joining on a, a laptop, the, the, you know, in the, the, the process on the MCU who the speaker is, so identifying is really easy, and we're seeing that conversation come into the, the room space where we have to do the voice enrollment.

Face enrollment, because you want the transcription. Again, I think at some point that will just be an expected default that you are enrolled to these services because it's, it's so important to have. Who said what in the transcription? 

Ritika Gupta: You bet. Yes. Some of these are extremely critical, especially in the room settings.

Uh, and we are trying to identify, uh, ways where we can prompt the users because some of these are capabilities which might seem like small capabilities, but they have a huge impact. On the quality of the call or any AI tool like summarization and stuff. So we are trying to identify sleek ways to prompt the user to 

Tom Arbuthnot: maybe we need to get facilitators to just jump into the meeting and be like, excuse me.

Who said that? Okay, I've got you. 

Ritika Gupta: You bet. I think that is exactly something that we are exploring because facilitator is going to be that companion in the meeting who will ensure that, you know, everyone's time is better leveraged and their configurations and settings are appropriate. To make this meeting truly productive 

Tom Arbuthnot: for, for the interpreter agent and facilitator, are those models being fed the audio, are they multimodal and they're doing their own transcription, or are you feeding them the text and then they're doing the, the, in the case of facilitator summarization?

In the case of interpreter flipping the language and, and playing it back? 

Ritika Gupta: Uh, great question. Uh, I. I'll share the details that I can share. So essentially, yeah. Yeah. What 

Tom Arbuthnot: you can, obviously, it's really interesting, so, uh, yeah, if it's, if it's secret source, then uh, no problem. 

Ritika Gupta: Uh, not really that much, but I'll definitely share a high level understanding here.

So definitely we use a model for speech to text, uh, and then we focus a lot on the quality of speech to text. Uh, that model is optimized for just transcribing or getting the text output there. And then the text output is then fed into. Another layer of models where our facilitator Copilot or recap summarization work, uh, because each one of those are actually optimized for a different purpose.

Uh, what I mean by that is Copilot has to do things real time. The the speed at which it can respond and then the context that it should respond with is fairly different than what we can offer with, say, post-meeting AI summaries. So essentially we have a multi-layered modeling, uh, architecture here, uh, so that we can optimize for the user experience.

Tom Arbuthnot: Yeah, I'm really interested to see if and when we get to, um, interpretation of tonality and, you know, like that's, that's the kind of a gap in the process. If we go just pure transcript and we feed the model of the transcript, you know, that someone can say something. Very sarcastically, I'll do all the work like, and clearly not meaning it.

And, and humans can pick that up. Obviously if AI is for just text, then it's gonna be like, okay, Tom said he'll do all the work. Um, have you, have you found scenarios like that with the, with the summaries where it's like text is, uh, is, is challenging in some cases, or I guess language in some cases as well?

Ritika Gupta: You absolutely are. Uh, spot on with that, Tom, because there are scenarios, especially when we work with. Global cultures, right? Like, uh, uh, text to text translation does not offer the best communication because certain, say references may really fall flat for somebody in a totally different cultural context.

So that is where AI has a huge opportunity to bridge the gap. Because we can process these things after the fact, or even real time in some scenarios to really offer a more meaningful translation. Uh, it's because this problem space is not just about word to word. This problem is about, uh, understanding and communication and bridging the communication gap where it cares a lot about the context, whether it's the cultural context, the choice of words, the tonality, or the accent.

Everything plays an important role to make a global communication more effective. 

Tom Arbuthnot: Yeah, I hadn't considered the cultural one. That's really interesting. The models are big enough and smart enough to actually know. Tom's in the uk. Therefore, this type of phrasing or this, this, this colloquialism means this thing.

So it can actually help beyond just word to word translation into contextual. Oh, in, in the, in the uk. That culturally means this, we're gonna take this break or we're gonna do this thing. 

Ritika Gupta: Yeah, you bet. I mean, when we were diving deeper with some of our users in deep, like in research, right? A very interesting feedback came up where things were like when it's actually sometimes not.

Even colloquialism is just about some choice of words, so it could really, when interpreted or translated into a different language, can can be perceived as extremely. Like rigid to like soft or like, you know, is, is is this very, being very pushy. So yeah, and we kind of learned about some of these examples, which were almost always out there.

It's not like we used to initially, before AI was there, a lot of you, us would sign up for some communication courses to understand and learn the tricks of like global communication or managing a team or working with the global team to be sensitive. Around these things. Uh, however, now, uh, sometimes like we are all humans, right?

Like we, we have our intuitiveness at one place and then learned or acquired things, uh, at the other end of the spectrum. But AI is a great way to kind of just bridge this gap where we can, uh, intuitively share how we speak, uh, without having to worry about how it might get perceived or interpreted in a different language.

Tom Arbuthnot: Yeah, I really like a, uh, a potential future where the AI can, can help you with that in, in real time. It's like, oh, okay, that person is, is trying to say this. Or coaching you potentially as a, a presenter to be like, bear in mind you've got multiple people, multiple cultures here. Think about phrasing it this way or that way.

There's so much, we are just at the start of the potential of what the models can do to help us both kind of. Post review and in real time, I think. 

Ritika Gupta: You bet. Yes. It, it is two way where it helps us become better at, uh, communicating and collaborating globally. And then it also helps us show up, uh, better, uh, in scenarios where we, we will probably.

Speaking or picking the right choice of words, however, unintentionally in some language, it may have, uh, been interpreted as something which either, uh, did, did not communicate the, uh, essence of that word or like, was completely, uh, like different than what, uh, the, that means in that language. Uh, because we, we are no masters in all of these global languages, and a lot of productivity is actually hinged on the effectiveness of communication, so, 

Tom Arbuthnot: mm-hmm.

And have you found that certain languages have been, I, I presume certain languages have been harder than others, either because of dataset and training size, or because of the complexity of the, the language and the structure of it? 

Ritika Gupta: Uh, absolutely. The, the challenge comes from several different. Directions.

Like sometimes it is the data set, uh, sometimes it is the quality of the data set because when we train our models, we are limited to public domain data sets, right? Uh, however, uh, it's very different than how a natural meeting conversation happens or even there are so many different formats of meetings.

Like a formal structured meeting is very different than an open conversational meeting. So. That definitely quality of the data set or access to a good data set is definitely one key factor into how we develop our models. Uh, secondly, uh, to your question of like, are some languages very difficult than the others?

You bet. Yes. Because the way the grammar works in some of these languages, like especially let's say Japanese, like it's very hard to do word to word translation. Because the grammar is totally flipped there. Right? So you, you have to know the full sentence, more or less to be able to appropriately, uh, translate it.

Yeah. Uh, so, so that is a very different problem, but we are lucky to have like a great, uh, science team or a cognitive service team that has been working into the speech domain for several decades. Like, uh, even before, like all of this became like the centerpiece of 

Tom Arbuthnot: Yeah, I, I spoke to somebody on that team, uh, an event earlier in the year at Microsoft, and they were like.

We've been doing this stuff for quite a while. People have just woken up to it now. It's like, yeah, okay. 

Ritika Gupta: Yeah, exactly, because, so that has been a huge lift to the experiences that we've been able to build and how we've been able to innovate in this space in general, because we are supported by a great science team, uh, and speech scientists, uh, where at least a lot of these things are.

Like almost known to us, or we have already optimized our models, uh, to work that. Mm-hmm. However, uh, no level of optimization that it's very hard to say that we are done because we're never done in this space. Like, you know, the expectation that our users come with, come up with, and then the opportunity that exists is just, uh, unmatched.

So we kind of rise up to the challenge every single day and, uh, try and make our models and our, uh, experiences more better. 

Tom Arbuthnot: Do, do you think there's a future? And again, don't, don't talk about it if you can't talk about it, but like I've noticed, we've just had Microsoft build and Microsoft have announced, uh, custom training of models for individual enterprises.

It feels like there's a possible future where like, actually we are a legal firm. We could make this translation way better if we were allowed to augment. Train the model on our terminology, our, our approach to meetings, our, our text. Do you think that's something that might come? 

Ritika Gupta: Uh, it's a very interesting thought, uh, to kind of consider.

However, I feel, uh, there are going to be some real challenges because I. One thing is to kind of feed the model some data, which is very contextual and local to your, you know, day-to-day working set or whatever. So that piece definitely would be extremely helpful if there is a way to provide, say, local to tenant models, because then you can, you know, do a retraining however.

Doing it from scratch is a totally different cookie to crack because, uh, speech modeling or any of these, they require an extremely deep understanding of like how LLMs work and how do we train LLMs, and then the choice that would exist. For all of our customers is do they really want to invest in talent and resources, or would they want Microsoft to provide a reasonably good starting point where they can fine tune and customize?

So I, 

Tom Arbuthnot: yeah, that's why we, I guess that's where I would see it coming in is the fine tune scenario. Like if I could feed it some samples of stuff we've done or feed it at least a dictionary or something. Yes. Uh, I, I've definitely had conversations with customers where they use. They'll use like internal acronyms and, and the translation will persistently get it wrong because it hears something different.

Or it's just, it will spell it out rather than doing it as, you know, the, the, the letters for example. 

Ritika Gupta: Yes. And our approach here is that we want to make this, uh, system smarter because there is a lot, uh, that we can do or provide out of the box. However, there will always be a need for, say, a human in the loop or some kind of fine tuning that the customer can do and offer.

So along similar lines, like we have been contemplating and we are going to share, like we have already started evangelizing. Some of this is we are considering dictionary based approach where you can bias the model a little bit. Towards, uh, certain things. There are certain references which are very commonly used in your environment.

Uh, some classic examples have been that the same acronym in, in say, an accounting reference versus like in a medical domain can mean totally different things. So, LLM can fall, uh, really flat if it is using just public domain context. So definitely. Domain based, uh, models or optimizations or biasing is definitely going to provide another, uh, quality lift to this whole domain.

Tom Arbuthnot: Yeah. That's really interesting. I think that's gonna have some huge, huge Outlook for people in terms of the, the quality of the output and therefore the quality of the, the AI summaries and the, the insights. 

Ritika Gupta: Yeah. I think the challenging part for the, for that would be how many of our customer, especially IT teams, are going to sign up for that challenge.

To maintain that dictionary and stuff, because I think what we hear from our customers is they're saying, Hey, why don't you just go and make it smarter? Right. So we, we definitely are working on both fronts. 

Tom Arbuthnot: Oh, I mean, somebody's, somebody's gonna build an AI agent, do that. Right. It'll scrape, scrape the, uh, scrape the SharePoint, scrape some previous meeting recordings, do something clever.

Like there's a, it's, it's gonna be AI to help train the AI no doubt. 

Ritika Gupta: Exactly, yes, because I think it. It does not have to be a blank slate. We, we definitely strive to provide our customers a better starting point, a smarter starting point for these types of solutions, and then definitely give them the ability to fine tune it because no matter how smart we make it, there is still going to be a very specific to tenant environment or, uh, like specific to their domain context, which they should have the ability to fine tune and manage.

Tom Arbuthnot: Yeah. Have your team been involved in the policy control of transcription versus recording versus in meeting? There's been a lot of work I've noticed done there. I, I work with a lot of legal enterprises and their default was everything off before the AI wave, and now they're like, oh, okay, actually we're gonna allow transcription, not record, or we must have a record for every transcription.

A a lot's been done there. 

Ritika Gupta: You bet. I come from an IT Admin background in the sense that before I joined this group, I was working in the IT Admin domain, building the Teams Admin Center from scratch. So it is a very interesting problem space because, uh, it truly are enablers, like they are. They are our friends who want to enable like the best productivity tools, the safest tools.

However, they also have a hard job to ensure that they configure the tools. The best way to avoid any. Uh, unwarranted, uh, incidents. So that is where, uh, we, we kind of, uh, have been kind of beefing up our compliance and manageability for recording and transcript because, especially with this wave of AI now that, uh, it's not only the quality of the transcript, but it also is around.

Who can download the transcript, what gets transcribed, what should be omitted from the transcript. Like, these types of things are becoming increasingly important because, uh, like there is a legitimate need to have good level of, uh, compliance and control on it. Uh, to be able to ensure that it's not like the wild playing field out there for AI to just yeah, uh, you know, uh, leverage something, uh, which it should not have.

Leverage. So responsible AI is, is like front and center for us. We go through several rounds of reviews, uh, with a specialized team, both internal to Microsoft, where we do these reviews around responsible AI in terms of, uh, Hey, is this something safe to do in the context of say, uh, how we leverage ai or even, uh, very exhaustive compliance reviews with, uh, regional compliance teams like say European Union Workers' Council because mm-hmm.

The, the threshold is extremely different. It's not that, you know, there is a global threshold to these things. Uh, there is a real reason why some companies, some environments, some regions are very open to saying, yes, transcribe every meeting. There is nothing to lose there, or, I'm okay with that risk.

However, some would want to be extremely cautious because their, their local laws, legal laws. Uh, or like their, uh, environment or domain is where it warrants, uh, more control. So we, our thought process here is that we spend the last one year to just, uh, beef up some of our manageability story for transcript and recording.

We invested in things like, you know, implicit consent, explicit consent. How, how do we give that power to the IT admins to decide what is the best configuration for their environment, for their users? Uh, because there is no right or wrong here and 

Tom Arbuthnot: No, no, with a, with a global audience, you, you have to have so many policies and toggles to, to meet every require.

I feel a lot of really good work has been done in that space. Like I've seen a lot of customers unblocked in terms of. How the recordings are now stored in the meeting, organizers, OneDrive versus the person who happens to hit record. The explicit consent you mentioned is a really good one. Like, uh, as much as, as a user, I, I, I feel like that's a, a, a painful threshold to hit the button.

I can absolutely see industries where. That's, that's right. You've got the explicit per user approval for that meeting to have the, the recording and transcription. 

Ritika Gupta: You bet. Yes. And I think it requires like a reasonable good level of craftsmanship in product making because, uh, we have to find a good trade off between how the user experience shows up for these types of, uh.

Legal implications that certain workflows can enable. And then what are the constraints that the, or the guardrails that an IT Admin or a compliance officer wants to put in place? They of, and our job as product managers on these types of experience is to bridge the gap here where, uh, people who have to set guardrails, they have good enough granular controls to manage those guardrails.

And then the day-to-day users, we don't. Uh, enable, uh, increasing levels of friction to, to get to, you know, what they want to do, uh, and be productive in their environment on a day-to-day basis. So we, uh, we are lucky that we have a lot of history as Microsoft to build extremely compliant products. We have a lot of help that we get from experts in these domains.

Uh, and compliance has been like front and center to our product development narrative, uh, all along. So. Uh, plus it's a, it's a craft, I would say. Uh, and we are all learning because AI is the new thing here. And that is where, uh, uh, we, we have a lot of conversations with our customers because everyone is leaning into understand.

Uh, what it means to have some legal boundaries around what AI can do and not like. There is no set global rules and laws written. 

Tom Arbuthnot: No, no, no. In terms of like, um, do we have to banner and announce it? Do we like, it's really interesting 'cause some of the, obviously I'm in the UK so a lot of Europe influence here and like, it, it, it's like at one point there was a, a consideration of announcing when AI is in use, it's like, well actually.

Spell checks. AI like turning on my computer is basically half AI now. Like, like, like what's using my phone is, is AI Apple just announced battery management via AI Like, so it's, uh, it's interesting as we find out what the agreed standard is in different, in different countries and cultures and regions.

Um, another interesting one for me is the, uh, you can now use Copilot without transcription, uh, during calls. 'cause uh, 'cause a lot of people have a hangup about, uh, or, or a legal requirement to not have. A pert permanent transcription. So essentially it's transient during the call or meeting. So you've got the benefit of Copilot.

Uh, I, I, I just struggle incredibly now, if someone took away my intelligent recap of my transcriptions, because I've got so used to that being always having that reference. 

Ritika Gupta: You bet. And I think like these waves are very rooted. In building that trust and confidence together as a team that is building the product and, uh, as a user group of these products or these tools, right?

Because we, we have to do it together where we have to first build trust. Like, of course, like things, uh. Can go wrong with AI Uh, there is, it's extremely hard to kind of just prevent a lot of things. But, uh, I think as long as we are convicted and committed to the approach of saying that yes, we will provide appropriate guardrails, uh, to minimize the risk, I think it's all about risk, appetite and risk assessment at that point.

There is, it's going to be hard to have an absolute. Uh, threshold here. Uh, given all the circumstances and, uh, regional requirements, as you rightfully mentioned. Uh, but then we, we always optimize for, uh, basically enabling, uh, responsible compliant solution. And, uh, of course the users would want more productivity.

The users, uh, sometimes are not as worried about. Uh, what gets saved or not. That is because they completely rely on their compliance and IT team to say, Hey, don't gimme a tool, which I can't use, right? So if my IT has enabled a tool, I feel so safe just using it because I trust them that they have done the due diligence and set the right card drills.

And for us as product owners of these types of experiences, we, we built for both, not just the end user, but also for the IT or the compliance officers. Yeah. 

Tom Arbuthnot: Yeah, and there's a real, there's a real challenge on it as well to enable enterprise approved tooling and make sure it is there. Because actually if they lock all this down and turn all of this off, you are really.

Uh, incenting people to go and use consumer tools or other tools that they shouldn't be. So I, when I'm having this conversation with it orgs, it's always striking a balance between, well, at least if you give them the tool that is compliant and you can monitor and you can do purview and all your good stuff around, uh, you know, kind of security and info second, compliance, you are not encouraging them to, to go off to a, a consumer tool.

Ritika Gupta: You. Yeah. And that's absolutely right because it admins have been like our strongest supporters and collaborators in how we are evolving the product thinking because, uh, we have only encountered like, extremely positive push. They really want us to enable these experiences and they, they're sharing really great insights with us in terms of how we provide the configuration so that in the.

Case of some event happening or whatever. They, they have some control and they can course correct or they can set the right card drill. So I think it's, it's really rewarding to work in the current times because it's not, um, any longer a case of like, you know, uh, designing for the user or designing for the it like everyone is committed to saying, how do I leverage ai?

Everyone believes and has faith in the. In the benefits that AI can offer. Yeah. Also all have our reservations because this is new. We are all doing it, uh, for the, for the first time in our respective roles. But the community is extremely thriving because I really feel great having these conversations with the end user, with the IT Admin, because it's all about how do we find safe solutions, uh, as opposed to saying, how do I restrict everything?

Or how do I overly, uh, index on being extremely risk averse. 

Tom Arbuthnot: Yeah, definitely. Awesome. Well, Ritika, thanks so much for giving us a bit of insight as to how all that stuff hangs together, both on the technology side and on on the usage side. And again, I think it's such a, such an exciting space you are in, in terms of the team's product of, uh, and, and I think we're still on the early days of how, how this data's gonna be used with things like facilitator.

So, uh, maybe we'll let those things develop a bit and uh, can have you back in the future and we can talk about ev everything that's being leveraged, uh, next time. 

Ritika Gupta: You bet. Yeah. Thanks for having me, Tom. And it was a wonderful conversation with you on the space that we work with, um, or we work in like, thank you so much.

Tom Arbuthnot: Great. Thanks.