00:05:29 · Siddhartha Ahluwalia, Bhaskar Ghosh
[00:05:29] S0: It can [00:05:29] S1: be it is massive. Silicon Valley is obviously technical people like us, whether we are builders or investors. Reason we have suddenly started looking at this very closely is the Why now? Question around tech. The LLMs and the models coming in along with the tooling around it are helping us do a few things really well suddenly. One is about unstructured data extraction and handling and management. If you look closely at a lot of these services, if you look at the human centric workflows in these services, it's about doing stuff with unstructured be spread... [00:06:06] S0: For our audience, can you explain as the terms that...
00:06:06 · Siddhartha Ahluwalia, Bhaskar Ghosh
[00:06:06] S0: For our audience, can you explain as the terms that... [00:06:08] S1: Oh boy. Unstructured data could be a spreadsheet, could be a handwritten note, could be a PDF, could be a Word document. And the... I mean the handwritten note part is extremely important actually. And you do stuff with data, look at it, you extract something, you write it into something, you read it from something, then you do something with it. And so human in the loop data management and doing stuff with data which is unstructured is a massively important and significant part of this of these companies. Number two is, there's a lot of human beings involved in what we call workflows. Which is you know, you do something, you write something, you decide something and you require domain knowledge to do it. And you have lots of hops in these workflows. So humans, you know, remember back in the past life industries like RPA tried to do this automation well, but for various reasons they could not. Leaving that aside, so unstructured data handling, the multi hop workflow automation stuff, and then there are other areas like you can handle domain specific languages which is a more software aspect in the context of Salesforce and others that we talked about. All of this can be automated with a very high degree of robustness, correctness, and speed with these new models. That is the central reason that you can now get a lot of this multi hop workflow functionality done automagically. With some right set of tooling around it which we can discuss separately. [00:07:42] S0: Can you give some examples?
00:07:42 · Siddhartha Ahluwalia, Bhaskar Ghosh
[00:07:42] S0: Can you give some examples? [00:07:44] S1: I mean, you can think of... If you take a software example, you can take a managed service IT provider. What does a managed service IT provider do? It sets up IT and solves IT problems for say small to medium businesses. And you get tickets. Somebody's password has to be changed. Somebody's Google Drive has to be set up. Somebody has to be added to an email group. So there are various levels of difficulties of these tasks. If you do look at just creating a new account and changing somebody's password, it's actually very simple workflow. You can do that actually automatically. And meaning with some human supervision, so you can do it in what we call a co pilot form or an autonomous agent form. But it can be done with a very high degree of precision, correctness and speed with an LLM today. On the... If you now go and look at... And [00:08:35] S0: before today, why can LLM do this? How can LLM do this?
00:08:39 · Bhaskar Ghosh
[00:08:39] S1: I am not a statistician. LLMs, the new models are all about something called Next Token Prediction. If you ask me the statistical reason why they can do it, I will not be able to answer. You should... We should just keep in mind that this class of models based on transformers can do very clever reasoning based and pattern matching based stuff for multimodal data. You know, so they can do better stuff for images, can do better stuff for text. Lot of these models do amazing stuff with text. Even though the core logic is around next token prediction, they can figure things out in a much more realistic correct way. So that was a partial answer. So anyway, I gave an answer about the software product, software business MSP. You can go look at and think about you want to do invoice processing for supply chain. Well, it so happens that a lot of these invoices are written by hand and they go and it is written by somebody, then it goes to somebody who is in charge of accounts payable, accounts receivable, they have to look at it, copy it into a spreadsheet, put it into some financial software. A lot of this done including the logic of the step by step. So that is the workflow. And this extraction of data can now be done automagically. So I'm giving very simple examples. Then there are much more sophisticated things called agents, which actually with the right set of inputs and watching inputs and outputs can also reason about what's the right thing to do. That's an even deeper level of sophistication. Anyway, coming back to the services industry, a lot of these basic workflows can now be automated. Now the question is what does that do? Well, at the minimum, all the human beings who are assigned to the services industry, in those workflows, they can become much more productive. So they can do... The easier things happen much faster, maybe some harder things have to be done with more human in the loop environment. So in the context of the services industry, first and foremost, AI can help increase productivity. So you can do more with the same set of people. Maybe you don't even need that many people. That's a very contentious point that that, you know, there's the there's the there's the numerator question of more business being done. There's the denominator question of how much resources do you need.
00:11:01 · Bhaskar Ghosh, Siddhartha Ahluwalia
So to make things very simple, if you sit down and think about it, with the judicious use of generative AI, you can increase margins in these industries. There's the technical why now and then there's the outcome of what we are calling AI enabled services. I'll just stop now for a second. [00:11:19] S0: Okay. So you are saying human dependent work. Services what? Human dependent. In the loop. Human dependent work. Where a human had to execute certain tasks, maybe across the same workflow that's doing on your spreadsheet or on multiple things, spreadsheet, a word doc, calling people, this all can be now done with the new set of agents that are getting built and that's why the services economy is very important as we step into building new companies with the AI agent framework. Is that right? [00:11:56] S1: Indeed, it is about redefining what the services economy and these companies will do. It's also rethinking how will this massive existing economy work. So if you look at any services business, you will have a front office aspect, you will have a mid office aspect, a back office, you will have an IT aspect. It so happens that a lot of the headcount in these services businesses is in the front and the mid office, where they are actually dealing with customers. Anyway, I'm overgeneralizing. It is likely that a lot of these offices within these businesses can be made much more productive and can probably even be made much more lean. That's kind of the central thesis. [00:12:39] S0: So front office means a doctor's front office. Can
00:12:42 · Bhaskar Ghosh
[00:12:42] S1: mean Yes, a doctor's front office, dentist's front office. Where in offline, you are sending reminders to patients, you are sending an email to patients. That's all happening automagically. You're interacting with them through text, you can interact with them through WhatsApp, any channel. And you are getting right answers, you're being able to reschedule stuff, you're being able to ask PII related questions maybe as long as you have the right security and stuff. All of that. Now, of the people in the mid office sitting in a doctor's or a medical office who are having to get data after a visit and deal with insurance, deal with follow ups. All of that, if you think of it as a co pilot setting, can be done automagically now, to a large extent. Which is why we are seeing so many companies getting started in each of these areas in a very, very high demand driven way because Applied AI founders, when they go into a domain, are finding many of these workflows are ready to be mechanized, maybe even disrupted. So that is kind of the software aspect of service. I think you and I have spent some time today talking about and where we have constructively disagreed on some things is, where... What happens next in these services industries? If you bring this Silicon Valley and Bangalore view of SaaS, you bring in, you write good software, you start a seed company, you start selling software to say these dentist offices, as an example, or these invoice processing offices for supply chains, second answer. What happens next? Do you build this out of the software business? One of the key thesis that we are all thinking about is we're going a little bit out of the box when it comes to software. It's reasonable to start reconsidering whether a pure software business will scale here. It is reasonable to start asking, if you sell pure software into these industries, how much of the economic pie will you get? It so happens if you look at most of these industries, they don't spend most of their money on software. So the question to ask is, are we envisioning that these services companies themselves will be mostly software and AI? That... So that's where things get murky. That's where things get exciting. It is reasonable to assume at some point in time in the five, seven, ten years that might happen.