Episode 1 of the ZohoDay24 Development Stories series
In this installment of the Intellyx video series, Zoho Development Stories, we replay our most interesting interviews from ZohoDay 2024 in McAllen, Texas. This episode, hosted by JE, features Ramprakash Ramamoorthy, Director, AI Research, Zoho.
Transcript:
JE: Hi, I’m Jason English. I’m a director and principal analyst at Intellyx. We’re an analyst firm and we specialize in digital transformation, so I like to cover especially vendors like Zoho who are really doing some interesting things across many different technology spaces. So today I’m with Ramprakash Ramamoorthy.
Ram, he’s the director of AI from Zoho Corp, and just great to talk to you today.
Ram: Thank you, Jason. Thank you for having me here. It’s a pleasure talking to you.
JE: Yeah. So, Ram, how do you think the state of play has changed for AI in general over the last 12 months?
Ram: Yeah, Jason. So I have been working with AI for the last 12 years, and I have seen the hype cycle, you know, go ups and downs.
I mean it all started off with the notion of consumer software being subsidized by data. I mean we don’t pay for our search engines, we don’t pay for our social networks, but instead We pay back with our personal data and then these companies target ads on us. So it, it started there. I mean, this wave of.
AI summer. It started somewhere around 2011 12 where data driven business models and cheaper data collection sensors. I mean, all of us have smartwatches and that gives us a lot of metrics about us and faster data processing thanks to the infinite computing power of cloud. So these three things led to the rise of this version of AI somewhere.
And it’s been 12 years now. Interestingly, the late 2022 we have been seeing generative AI, we have been seeing AI being able to summarize, AI being able to paraphrase, AI being able to generate images, generate music. This is a very new trend that we have seen and, and the hype cycle has gone through the roof.
Everybody is talking about AI, everybody is consuming AI in some form. I mean, if you consume software, you’re naturally consuming AI because Most software today is AI enabled. So the last 12 months have been phenomenal not just with the hype cycle, but also we are seeing actual value out of AI.
JE: Yeah, that’s very interesting. I mean, Zoho in particular has been doing some form of AI for years. You have this Zia, the personal assistant has been part of the software. So how has your thinking kind of evolved on that area now? And, and what’s different about how you apply it?
Ram: Great, Jason. One thing that we have always seen is these consumer grade models really don’t fit the enterprise. So we saw that in 2017, 2018 when we launched our grammar and translation assistance. So we saw that enterprise users were using tools that were built for consumers to translate their content to, to grammar correct their content.
Again, this could have a lot of sensitive information. You wouldn’t read through the privacy information of the consumer grade translation tool, consumer grade grammar correction reduction tool. Our notion has always been bringing AI into the enterprise so that the context remains. Even with the large language models, we see a lot of consumer grade large language models outside.
These are standalone behemoths. They don’t know any context. Am I talking to Sridhar, my CEO, am I talking to Sridhar, the colleague in my team, or am I talking to Sridhar, the customer? There is a context in everything that we do. And Zoho, as you know, we have been evolving as the operating system of business.
Where, right from the journey of a customer, right from the time he or she landed up as a lead, to all the support tickets that they have raised through the process, right? We are able to see the full. CX journey full employee experience journey right from recruitment to payroll to projects and tasks.
They have been involved with their expertises. So given we are. We have a huge amount of data or information that we can make use of. So we are calling it contextual intelligence, right? This contextual intelligence becomes super powerful. It’s not standalone. I’ll give you an example.
So we have had our expense, Zoho Expenses, an app where you submit expenses. And, and nobody ever says, oh, I love doing expenses, right? It’s, it’s one of the crazy things that nobody ever wants to do. So the first thing that we did sometime around 2018 19 we launched AI powered expense receipt digitization.
So you just take a photo of the receipt, the AI is able to extract information. Now, in the language model era, where we are seeing emergent behavior out of these models, we have used these models to enhance the accuracy of the receipt digitization. Because a lot of times, in receipts, you don’t have the name of the merchant in plain text.
It’s just probably a logo. A human would look at the logo and understand, oh, this is Starbucks, this is a shelf, right? But the models were not able to understand. Now, with the power of these bigger models, we are having multi modal capabilities. We are able to understand that this is the merchant name. Now it does not stop here.
Now I have submitted my receipts and 100 others who have traveled with me have submitted our travel reports. There is an administrator who goes through all of it. Now what can we do to reduce the time taken to bring down the approval of this expense receipt? So here is my travel policy. Here are the receipts that this person has submitted.
Can you see, is there any anomaly? Instead of that person going through hundreds of receipts, he’s just going to go through three receipts. And the time taken to approve that receipt gradually comes down. So this is where we are seeing AI. AI as a productivity booster. AI that is contextually and tightly integrated into everything that we do.
So, overall, again, we’re not talking about replacing people, we’re not talking about, we’re talking about enhancing their productivity, we’re talking about free up their brains to, to not you know, waste time in mundane repetitive things. So, go do something where you are productive, that’s where the narrative is heading, and the context AI strategy.
JE: Yeah, contextual. That is very interesting. I mean, we’re users of a lot of the Zoho suite ourselves, and in particular, I’ve used Expense, and it is getting easier all the time. And it’s kind of interesting, but like, outside of even those productivity tools, there’s this kind of new resurgence of how to, like, better utilize the human potential for creating things with AI as part of that contextual process. I mean, Zoho has a huge corpus of information to learn from and for AI to learn from. So how are you kind of applying that to helping bring forth more of that kind of creative spirit of the people within these teams? Yeah,
Ram: Creativity has always been very important. So, I think this is a good example of how you can apply career function in whatever role you do.
I mean, you just don’t have to be a marketer to be creative. You can apply creativity in any kind of roles that you are doing. And now, with all the context, with all the information, I’ll give you another example. So, you have, let’s say somebody sends you a legal document to sign, and it’s from a different country.
Of course, businesses are operating across the countries. Now, it’s in a different language. The first thing that we do is, yeah, ok, here is a Spanish agreement that I’ll have to sign. Or use AI to translate it. Now that the translation is done and legal agreements are at least 30 pages long. Can I summarize what the agreement says, and can I give out the key highlights?
This is where AI helps. And then the next step is, okay, these are all the agreements that you have signed in the, let’s say last six months. And I find these clauses to be very different from the usual agreements that you sign. So you look at the behavior profiling and find out that something is anomalous.
Again, no rules configured no pre-configured configurations, nothing at all. It just happens. It just looks at all the agreements that you have signed in the past and says there is a deviation in this. And then use generative AI to recommend what could be the corrected statement. What are you expecting?
Right? So it doesn’t stop there. Now, let’s say I’m going to recommend that changes and I send it back to the person who signed me. That triggers a workflow, right? So, so it is deeply ingrained that the person who is signing will really not know it’s AI. I mean, you don’t have to open up a chat bot and go ask somebody that can you help me spot the anomaly?
No, it’s all deeply integrated with the process. That’s what we call contextual intelligence.
JE: Are there places where things like heuristics or rules based, the kind of things you’ve always excelled at in the past, are there other places where AI is really just not relevant and it’s actually better to apply those kinds of work?
Ram: Yeah, Jason. So, if you look at our AI stack today, we have one, the rules, the very basic rules, and then we have narrow models, which do one thing at a time, and then we have models with domain expertise. So, again, we classify that into small, medium, and large language models. The small language models in the 3 to 5 billion space.
It goes and tweaks these parameters that they set when the business started, or when they set, when they install that software in their system. So, we see AI as a combination of this narrow model and the small, medium, and large language models.
And, the problem is, larger the model, the better it is. But, I mean, we’re not running on infinite computing power yet. So, and ultimately, at some point in time, the GPU tax has to be passed on to the customer. So, that’s why we have the notion of smaller model capabilities, smaller language model capabilities, which So, they exhibit the emergent behavior as of the larger language models, but at the same time, they, they, they are able to run on CPUs for inference instead of GPU hardware.
So, I am not passing that cost on to the customer where they would need specialized compute devices like GPUs to run their AI inferences. The training is anyway on GPU, but then if I am able to do the inference on CPU for something like identifying the name of the merchant, so it’s just one task and it’s really good.
It’s much better than the narrow model. What we see is all of these combined together each playing their role and use the larger language model only in areas where you need that higher, where you’ll have to generate an alternate sentence to the legal agreement that you want edits on.
That is where we use these bigger models and the smaller models. The context is the combination of these models.
JE: It’s almost like the contextual AI is like an AI that determines the appropriate sizing for the task at hand. It’s pretty interesting. Something I’ve been super interested in — basically you’ve been doing this for a while, but also internally.
How do you kind of feel the knowledge necessary to generate? to, to produce your AI strategy. I’m kind of curious. I mean, it’s not developed in one team or one place. So where does, where does that innovation come from in your org?
Ram: So, we follow a hub and spoke model. There is a central AI team that looks at all of these parameters that where we try and set the basic framework that teams are going to use.
And then AI has become so commoditized that it’s important because teams, the individual product teams, they have the knowledge on what is necessary for their customers. They have the deep domain knowledge. Basically, the central team might not know the business of the product team. So the product teams knows the business of their customers.
We have a very decentralized model. Where all the privacy, the hardware aspect of it and the broader platform is taken care of by the central team and each product team has their own AI setup where they consult with the central team and then we work together on getting these features. But the important thing we, we convey across all our team members is that don’t get intimidated by AI because the rate at which things are changing.
I have seen it since 2011 and, and every other day there is something new. Everybody wants to share that cool article on AI that you should read. The progress could be intimidating, right? And it was a lesson for me, the first 3 4 years I found it very difficult to comprehend because things change so fast.
And you would have spent months together to understand something. And by the time you finish it, there is something totally brand new which is 180 degree opposite of what this is.
One advice to all my colleagues and teams who are involved in AI that I have always personally given is, okay, don’t get intimidated by the math, don’t get intimidated by the rate of change, there are certain things that will not change, stick to that, focus on that, and of course, you know Zoho, we have offices across different parts of the world.
For example, we have an AI team based out of Querétaro in Mexico, we have a five member AI team, we started that initiative last year, and, So we are pretty much decentralized, and there’s a lot of ground up learning happening across these locations.
So, not being intimidated has really helped me and my team to stay focused.
JE: Right, it’s not as scary as we’ve been led to believe, is it? Yeah. Huh. Well, thanks, Ram. This has been interesting for me. Where, where can people watching this show go to find more about what you’re currently doing?
Ram: Yeah, Zoho.com is your place. We have a all our AI efforts are umbrellaed under the term ZIA. I mean, you might want to look it up. And yeah, the, the notion is it’s contextually, it’s not about the AI. It’s about your CRM or it’s about your finance tool. It’s about how we have contextually weaved that into the process.
So, really great conversing with you. Yes, and I enjoyed it. Yeah. Thanks,
JE: Thanks Ram, and thanks (for watching).
Watch the whole video on YouTube here: https://youtu.be/aGPxKnMN_oI
©2024, Intellyx B.V. Intellyx retains editorial control over this story. At the time of writing, Zoho is an Intellyx subscriber.
