[Music] [Hina Naqvi] Hello, everyone. My name is Hina Naqvi, and I'm a group product manager for Acrobat and focusing on GenAI.

So I wanted to kind of get a quick feel for how comfortable people are using GenAI tools. So just a quick show of hands of people who have played with Copilot.

All right, nice. People who have tried maybe, perhaps ChatGPT? That's a winner here.

And then maybe Gemini? All right. You certainly got that. Or anything that's very specific to your function, like Litera for legal function, and there are very specific ones. Are there any other ones that are...

All right. Well, we got one or two people. Awesome. So, in terms of agenda for today, we'll start off with the uber question, as to why are users seeking out PDF tools today. And then we'll move over to talking about the use cases across different functions because it's very function specific. And then we'll move over to talking about what Acrobat has just launched in terms of the two features that we have. And then we will also talk about what does it mean to design GenAI features for enterprise customers as well as non-enterprise customers for that matter. And then I will also end with sharing the broader vision that Adobe has around PDF related AI in our roadmap. And then, obviously, we'll finally end up with questions and answers. So some of us can relate to the image here. Lots of information. Where to find it? I think, like I said, that everybody here, I can imagine, there's a special folder on your machine where you put your PDFs and your other documents. "Someday I will read them because those were good documents to read," but we never get around to reading them.

Today, PDFs are consumed the way they were authored. You have to manually make sense of the overall structure. You have to comb through multiple pages to get to the information that's specific to you. And whether it's a single pass or a multiple pass, you have to, like... It takes a really long time to get to the relevant parts of the document. And sometimes the document is long. Sometimes it's complex. And users just don't have time to read the whole document and extract the information that's really, really specific to them.

We can categorize information into two buckets, structured and unstructured data. Structured data is like the ones that we see in databases. And despite its own challenges, there are certainly tools that help you pick out data inside the databases. But unstructured data is where a lot of time is spent at, 80% of enterprise organization data is inside unstructured data.

And some trivia that was shared at the keynote today as well that there are 3 trillion PDFs in the world, and millions are being created anyway. So this problem is not getting easier, and we need to solve this.

So PDFs are goal minds of critical business data. And they can be created from anything. And that is the beauty of PDF, that you can create it from practically any file format. And PDFs are used to communicate information. Generally, they are either in the final form of the document or the near final form of the document. But PDF itself has no structure. And it's a lot of text, tables, images. And how do you basically make sense of this PDF programmatically? I'm really excited to share, I think this was shared in the keynote as well. So a little bit of...

Shantanu had already shared that with you, that we have two new features. One is the Acrobat generative summary and the Acrobat AI assistant. Literally, before these two features came along, I would sit there, open my PDF, and depending on the type of PDF I was reading, I would literally read it from top to bottom. Or I would spend some time just kind of popping through different parts of the PDF and figuring out which parts were pertinent to me. And whereas now I can just basically say, what relevance does this have to me and particularly my role? What areas should I be focusing on in terms of my role? And the context in which somebody reads a document is very much dependent on what they're looking for. Often, the author knows it, but the audience might want to consume it differently. Like I said, maybe they want to skim through it or just read it from cover to cover.

And these two features are precisely there to get you to the information that you want. So how does consumption and reading of the PDF help to the creation of other content is also a very critical part of the journey because we just don't read documents for the sake of reading. We're generally creating some other content as well.

Last week, I think this was also mentioned in the keynote, but I'll restate here that we also have announced our longstanding partnership with NVIDIA to explore new ways for enterprises to harness this business intelligence that's stored inside the PDFs.

Basically, we're taking the power of PDF services and combining them with the NVIDIA technologies to advance what's possible for the large language models. And we're doing it in kind of three fronts. One is using PDF technology and enterprise to support in the training and finetuning of the LLMs. The second way we're working together with them is for, basically, utilizing the NVIDIA NIM and NeMo microservices and Adobe PDF services. And then the third part is to be able to develop open data sets to promote LLM research and development.

So I'll walk us through a couple of how this is applicable across different departments, within organizations, and how we can save many hours by using these tools. I'll kick off with a marketing department. I think many of you in the audience can relate to that. Extracting insights from corporate disclosures, whether you're reading white papers...

Competitive landscape documents, marketing reports, and I can go on and on, but there's a lot of sample use cases we have here. And some of these documents can be pretty long and detail intensive. And you are creating and reading these messaging documents, creative briefs, product fact sheets. And sometimes we're all, you know, short on time. So meet Ryan Lytle. This is the marketing persona. He's a director in a midsize global marketing organization and works on digital marketing solutions. So Ryan established himself as a thought leader in his company. And research shows that for blog posts, generally, the frequency that's expected is about two to four blogs per week. And he needs to come up with the content ideas often quickly, so drafting blog posts and working with multiple collaborators like his product team and his legal team, this is a very, very, ongoing frequency. So Ryan is writing a blog on digital trends and customer experiences and focusing on topics as customer expectations, importance of creativity, and data utilization. So I will switch over.

All right.

So let's see what Ryan can do.

I'm going to lean over here because I have a second laptop where I'm doing the demo, so I would not be facing as much. So bear with me, please.

So he's opened a couple of PDFs here. He starts off by clicking the GenAI summary.

And basically, the concept of summary is that we are synthesizing information from a collection of documents and creating a summary across that. Right now, we're starting with a single document as part of the beta that we've just launched. But the idea is that we'll be able to open multiple PDFs and do this across those documents. And Ryan wants to get key insights and summarize them and create his blogpost.

Typically, in the interest of time, he wants to consume five or six key points that would resonate with his audience. And he wants to create them quickly. So as you can see that this generates a very nice, quick, you know, top level, one line summary of that document. You have a very nice table of contents here as well. I'll click through different sections. And basically, what I really like about these results is that Acrobat not only just gives me the headlines but it goes deeper and provides a little bit more framing for those ideas. So if there was a specific area that you were interested in and you wanted to write an email or a quick message to somebody in your Slack or Teams group, you're able to just copy this information. So if you notice this, there's a Copy button right here.

And we're taking this document and nicely summarizing it into basically a one-pager. So here's a summary for each of the sections here. This is a 24-page document. And now that Ryan has nicely oriented himself in this document, he can now go deeper into certain areas. And this is where the AI assistant comes into play. Select that. And summary basically plays as a springboard for the AI Assistant. And what we do is, to get you started, we start off with three suggested questions that are listed here. I'll select one of them.

And one of the things I want to bring your attention to are these numbers. If I click on them and it takes me specifically to that part of the document that talks about that particular point. So Adobe's custom attribution engine, and this is proprietary technology that we have, helps us generate these attributes within the PDF document.

So since I'm in Acrobat, I have all the other tools that are available to me. I can highlight. I can read comments. I can do all these things. So as I'm consuming these documents, I have a whole collection of tools that are available to me. And that is the power of working and opening this document in PDF. So let's see if I could pick one of the prompts that I had prewritten.

Come here, and I'll ask this question.

And I think I just want to make sure that, you know, since we have all tested some technology or other, we're all learning to write good prompts. It's not something we all grew up with. But I think the deeper, the more well thought out our prompts are the response, the quality of the response, obviously, matches that. So again, notice the different numbers here.

I'm just gonna jump around the document here. And now you guys will see the magic. I'm just gonna go ahead and ask Acrobat to create this blogpost for me.

So Ryan has written a pretty detailed prompt for that, which is asking it exactly what they want in the blogpost and where is it going to be posted, the tone that they want that to be, and how long it should be in terms of reading. So we'll just wait for it to come back.

All right. So as you can see that it created a title for me. It has the kind of tone that Ryan was looking for. It's formal but not too formal. It highlights the specific parts of this report that I was interested in highlighting. You can click on it.

And it also talks about how creativity is important for a retail organization. And then I can go and click on any of the attributions. So this kind of builds trust in my reading as well, right? So it's not that this information is coming from outside the PDF. This is exactly in the PDF. I can get to it. So as a reader, this is like, "All right, I get it. This is inside the PDF." And then, you know, my favorite, it even generated the hashtags. I really didn't even have to spend time generating those. So I'm gonna go ahead and copy it. And then open my LinkedIn. And I have an image that I've already set up. I copy paste it. I click Next. And if I wanted to make quick edits, I can just do it directly. But it's ready to be published.

So as you can see that...

Being able to generate this blog this quickly, I was able to do something that would generally take me 1.5 hours, or maybe an hour, to pretty much less than 15, 20 minutes. And that is productivity for you because if you're doing this two to three times a week, this definitely saves you time. So what we envisioned in terms for the assistant is that we're starting out with consumption. And then we are definitely going to bring in experiences that cater to creation of other content as well as review process. And I'll talk about that a little bit later.

So let's start with another example.

So this one, I have two documents. One is a 100-page annual report for a fictitious organization called Altura.

And basically this annual report focuses on Altura. And then my other document is a sales play for the organization where I could figure out what is the sales play for my sales team to go ahead and sell my solution. So one of the key elements to the customer journey is discovery. And how do we determine specific discovery questions to ask your customer? And definitely, the AI assistant will help me...

identify that from within the document. So basically, what has happened is and before I kick it off, like, what I've done is that in order because multiple documents is something we're working on, I've actually combined those documents here using our combine tool. And let me just go ahead and showcase that as well.

So it's combined the document, and I will go ahead and type my question here.

All right, so as you can see that it replies with a few discovery questions. Some of these are great. I definitely want to get a little bit more specific than what it has showed. So let me start to type another prompt.

So in this one, I'm asking specifically what are the strategic drivers that Altura have that enterprise could solve for.

So as you can see that it has identified the different drivers that I mentioned in the document. And I can see that there are excellent insights here. And I am starting to see some themes around digital transformation, compliance with accessibility, integrated workflows, etcetera.

So these questions certainly help me draft my discovery conversations with key personas. But finally, I'll have another question to ask...

Which is discovery questions based on the strategic drivers that it identified just now.

Okay. There you go.

So here you can see that I can understand the pain points. And as a salesperson, I'm on my way to understanding and developing a compelling solution for my customer.

So now I'm going to go ahead and show you a couple things. So we're back in the annual report for the company. And I am going to go and showcase some really cool things that AI Assistant can do around formatting. So let me go ahead and ask it to summarize our 2023 financial highlights and format them in a table.

So as you can see, this is a fantastic way to visualize a document. It came back with the table, basically that streamlines information, which was represented graphically. And it would be very difficult for me to go and copy paste it within the document. And the best part of it is if you click on any of the attributions, it takes me exactly to where that information was. Again, it's all about building the trust as you're consuming the document. And I can click on different parts. And I know exactly where it picked up specific year over year change. It's just going to all the specific details of the document. So next thing I'm going to show you, which also is pretty cool and helps me save tons of time, is...

So what's my annual percentage growth of net sales year over year from 2019 to 2023? I go ahead and it helps me basically do a calculation as well.

All right, so as you can see, this is something that would have taken me, again, an hour or two. And it's able to do these basic calculations of the annual percentage growth of the net sales. So I just wanted to demonstrate that, again, it does depend on the quality of the prompt. We're all getting used to it. And I think we've definitely heard the feedback that the experience itself could help write better prompts as well. But I just wanted to showcase the power of the AI Assistant. So now I will switch over.

All right, so we just talked about the marketing use case with Ryan. And within the HRC use cases, we're talking to a lot of customers, we're seeing a lot of use cases across that as well. We're seeing it across policy consumption, drafting, creating and consuming onboarding, offboarding documents, employee agreements. And the one that we just heard is who's HR of Diversity and Inclusion. And as a diligent HR lead, she's working with the legal and compliance department. And she's responsible for basically navigating through complex regulation updates that impact her organization in various geos. So this requires her to meticulously read pretty extensive regulatory documents. And she has to read them from start to finish. So she can't skim them like the rest of us do, maybe.

So she needs to make really, really... She needs to comprehend those changes and basically be able to communicate that to the rest of the organization. So at this point, this whole effort is manual. It's effort intensive. And with AI Assistant, she's able to ask these questions and she's able to consume and send out emails and other, documents she may need to author for this. So the other thing I wanted to do is quickly demo another HR specific use case. And this one is more from a consumption perspective.

So let me share my mobile device.

All right. So what I have open here is the Adobe Acrobat Reader mobile application. I have Altura's employee handbook loaded on this document. I mean, on my application. And I open that.

And I have the AI Assistant on my mobile device as well. So again, this whole technology is available across web, desktop, and mobile. I select it.

And now I'm going to go ahead and ask the question.

What is Altura's parental policy? Oh! Didn't pickup my...

What is Altura's parental policy? All right. So as you can see, it went ahead and picked up the response here. And again, on the go, I'm able to click on the attributions that's there and I can go exactly in the document. So between liquid mode and AI Assistant on the go, you're able to consume a lot of content, things that you don't wanna go back to your desk, and you need the answer immediately, so definitely highlights that particular scenario.

And then going back to, again, the finance department. Like I said, as we're talking to a lot of customers, we're going through department by department and we're seeing these use cases. So within finance, we're definitely seeing company earning transcripts, audit reports, compliance reports, your classic 10k, 10q, and the earning transcripts that perhaps an investor relations person might be interested in. And within the banks, we're also seeing some use cases where loans and etcetera are being evaluated against a certain criteria. So I think for the last several months that we've been in this sort of, like, let's discover what use cases are possible, they're definitely appearing across a lot of functions. Again, in terms of adoption, where AI is a critical part of the function, IT is even past limited adoption. They're definitely trying to move past the piloting point. And in terms of certain use cases that we've seen within IT, we're seeing them across being able to provide intelligent scripts for edge interactions, for reading manuals, enabling natural language troubleshooting are some of the examples, being able to look up incident reports, and vendor agreements. And being able to read these PDFs at scale definitely empowers this function, too.

Again, example of procurement. We are seeing this in things like vendor onboarding. That's a critical part. Supplier compliance is another one. Purchase order, finance agreements, NDAs. And pretty much we're finding that every industry and business function is able to find a custom application for this. We've spoken to a couple of universities as well. And research is very, very critical, where they want to be able to search across. You go to the libraries, there's, like, tons and tons of PDFs. You don't even know which five PDFs you're supposed to read. So we're seeing that use case pop up as well. Manufacturing is another industry as well.

And then even within the local government and state governments, we're starting to see some of these use cases appear.

So let's talk about responsible AI. So Acrobat has a longstanding history of maintaining authenticity and integrity of our documents through a variety of technologies that we have. Digital certificates is one of them. If you're a user of Acrobat Sign, you must have seen our audit trail. There's other PDF mechanisms, such as encryption, disallowing people to copy and print, etcetera. And Adobe has been investing in AI for over a decade. And we are committed to the safety of our models and to accountability, responsibility, and transparency. And you may have heard about content authenticity initiative where we formed a coalition with several tech companies. And it is right now available in Lightroom, Photoshop, Firefly. And we're looking into bringing it in Acrobat as well. So for Acrobat, again, the idea is to take an agnostic LLM integration approach. We want to be able to curate the best of technologies, patterns, models, and deliver the right output for your needs because like I said, PDF is a special document. There's a lot of structural information. And with our 30 years of expertise, we're able to bring you a proprietary, custom-built attribution engine. And we are able to go through the structure of the PDF, the tables, what's inside the tables. Images is another one that we're going to be looking at as well, being able to get... Extract information out of infographics.

We are right now leveraging Azure OpenAI from a public element perspective. And we're also using their OpenAI's content filtering service to moderate, hate, violent, self-harm content.

For both the features that I demoed today, we're extracting all this content from PDF and caching it for about... Not about, for exactly 12 hours. And the reason we do that for 12 hours is so that customers have a seamless experience. It would be weird for you to type a whole bunch of questions and for us to wipe it within minutes. So that's why the session duration right now is 12 hours. And then we obviously are only focused on getting the content just from the PDF that you have opened. And in the future, when we do support multiple documents, the idea is to stay within the boundaries of that set of PDFs.

At the heart of this whole thing of liquid mode is the proprietary AI/ML models that basically help us identify the different elements inside the PDF, whether they're headings, paragraphs, images, lists, tables, and much more. And it helps us understand the hierarchy of the PDF as well. And the reading order is a very particular part of it. And, like I said, this is a journey for us. We'll be able to do more and more of the PDFs.

So far, the demo I shared with you is what's available in the Public Beta. And just to give you what else is... We're cooking around PDF consumption. We want to be able to support scanned documents. I think there's many people in the audience who might be coming within organizations where your digital transformation, you had a lot of scanned PDFs and scanned content. And you want to be able to run the system in those. We also are working actively on supporting multiple documents. And where you would be able to ask the questions across, like, a set of files, like seven, eight, ten. And you'd be able to do that. You can ask for a summary as well. And then the other thing is that, as you saw on the right-hand side, there was a persistent chat history. And talking to some of the customers, they want to be able to store that persistent history. Whether it's in the cloud or in their local machine, they want to be able to store those questions. Maybe they want to come back to it three days later when their session has expired. Or they want to do it from an auditing perspective as well, where they're like, "Hey, two years ago, I asked these questions. What was the answer?" So these kind of sort of enhancing the feature.

We're also looking at the ability to select text and suggest edits to it as well. So that's another one. And then moving into sort of the review and collaboration aspects of it, being able to take someone's comments and make suggestions and provide edits for it. That's like the next natural progression in terms of the feature. We are also looking at being able to, like I said, identify action items as well from commentary that comes in the PDF. And basically, our vision is that we take the Assistant and provide support throughout all your PDF needs and to make the whole experience much more conversational.

So what I'm about to show you is not part of the beta. So it's just to kind of give you a quick vision as to where it's taking some of the things. So, like, the disclaimer is that the UI is not final, but I'll still play the video. All right, so let's see.

[Woman] Writing a marketing brief involves distilling a lot of documents into a simple actionable point of view to help you seize new opportunities. With Generative AI in Acrobat, individuals and businesses can engage with their documents using natural language to significantly accelerate their time to knowledge. You've opened a report in Acrobat and your trusted assistant is here to help you. It starts by providing a summary and key insights upfront. You can either ask your own questions or use suggested questions to help you quickly start understanding this document. With just one click, you've got the answer to your question. And with sources cited, you can easily see where this answer came from in the document, like this table on page 4. Now you wanna learn about a topic from different sources. Instead of having to go through many individual documents, your assistant can do it for you. For example, you ask, "What are best practices for CTV advertising?" They respond with information across multiple documents and document types, like PDFs and web articles. You can even click to view and verify sources. You or your organization can define permissions or data access controls. It can also help you identify patterns, draw connections between documents, and highlight other details that may be relevant to the report you're trying to write. Instead of taking multiple hours or days to review and comprehend these long documents, you've now synthesized, brainstormed, and developed a point of view in much less time.

After synthesizing information across multiple documents and coming up with a point of view, you are ready to create a report. You can simply ask the assistant in Acrobat to create a report based on your conversation. And within seconds, a richly formatted document is generated. It suggests a white paper format, but you can choose to change it to a presentation, summary email, or web article, and then further personalize it. Next, you're prompted to apply a style using design templates to help advance the visual appeal of your draft. Once you select an option, it's automatically applied, instantly transforming the document. And your linked Brand Library, which includes your company logo, is also incorporated. If you want to replace the image in the center, you can select an image, enter any prompt into Firefly, and choose from multiple image options. As you review the text, you see that the overview is quite long. With Acrobat, you can rephrase and shorten the copy effortlessly. The document now looks perfect and you are ready to share it.

So this video kind of just basically encapsulates the vision that we have across creation, consumption, as well as collaboration.

So currently, as I mentioned earlier, that we are definitely available for a Public Beta for Acrobat Standard and Pro, for customers who have Teams and individual subscription. And we also have an Acrobat Pro trial users as well. And in parallel, we are running a private, by-invite only enterprise customer private beta. So if you are interested in that, please come and speak to me after the session, and we'd love to have you added to that. [Music]