back to indexBuild Dynamic Products, and Stop the AI Sideshow — Eliza Cabrera (Workday) + Jeremy Silva (Freeplay)

00:00:00.000 |
All right. Well, thank you for joining us. We are here to talk AI products and specifically 00:00:21.120 |
dynamic products, which we'll unpack in the next 20 minutes or so. A little bit about us before we 00:00:27.960 |
jump in. I am Aliza Cabrera. I'm a principal AI product manager at Workday. I'm currently building 00:00:34.840 |
with an incredible team, our financial audit agent. I also led go to market for our policy agent, 00:00:41.080 |
as well as early access for our assistant, which is more like a co-pilot, as well as some of our 00:00:46.440 |
early days kind of Gen AI features as well. My name is Jeremy Silva. I come from a data science 00:00:52.620 |
machine learning background. I've been building language models since the dark ages, pre-GPT3, 00:00:57.180 |
that is. And now I lead a product at a company called FreePlay, which exists to help teams build 00:01:01.820 |
great AI products. Awesome. Let's get into it. So if there's one thing that we want you to take away 00:01:10.220 |
from this session, it's to stop the AI sideshow, which I know sounds a little bit counterintuitive. 00:01:16.620 |
We're all at an AI conference. All of us are talking about agents and AI. It's all over the place, 00:01:23.180 |
right? So what exactly are we talking about here? Having AI leading your products, your go-to-market, 00:01:31.340 |
your strategy. This was a really great approach when we were all trying to communicate that we were at the 00:01:36.700 |
forefront of this technological disruption, right? But now everyone is really kind of saying the same thing. 00:01:45.660 |
And if we look at the different products that have resulted in hindsight, which hindsight's kind of 2020, 00:01:52.380 |
right? We're able to see that what we've been doing is building product to try to figure out what these 00:01:59.020 |
different technological breakthroughs can do for us. So let's unpack what we're talking about here a bit. 00:02:08.140 |
So let's go back maybe post-chat GPT, maybe for some of us in the room pre-GPT, but whenever your 00:02:14.940 |
sort of aha moment was with LLMs, trying to figure out what you can do with the tech, what you can't, 00:02:20.380 |
what the boundaries are, we ended up using chat UIs, content UIs, existing applications, right? To be able to 00:02:29.180 |
really test the boundaries of these LLMs. We were also using multimodal to see what different kinds of inputs 00:02:35.980 |
and outputs we could use the technology for. Then we realized we could ground the models. 00:02:43.820 |
We had vector databases and RAG. We were trying to get to accuracy and truth, if we can agree on that. 00:02:51.020 |
We had larger context windows and increased memory. We also weren't super, I would say, comfortable with 00:03:01.340 |
having AI do work for us. So everything was a co-pilot, right? A buddy next to us who can help us get 00:03:07.980 |
things done, but we don't want to be taking anybody's jobs away. We don't want to be sort of automating work 00:03:14.380 |
until we realized that might actually be kind of nice to be able to have agents that can do things for us, 00:03:23.500 |
to reason, to be able to use tools and various APIs to orchestrate across different business problems. 00:03:31.100 |
And this is the state and sort of space I would say that we're in right now. We're not saying that these 00:03:37.420 |
different approaches are wrong, but they're an approach to understand the technology. And it's not going to 00:03:44.540 |
build you a differentiated strategy because everyone is doing the same thing. 00:03:48.460 |
So why do we see these kind of like bolt-on, non-differentiated AI products persist? 00:03:56.700 |
By working across dozens of enterprise companies at Freeplay, we've noticed a common trend emerge, 00:04:02.940 |
which is companies know they rightly need to prioritize AI, but the way they do that is by creating 00:04:07.660 |
this sort of centralized AI strategy. And what happens is this centralized AI strategy starts 00:04:13.500 |
running as this sort of sidecar to their core product strategy, rather than the two be deeply 00:04:18.140 |
integrated. There's different initiatives, sometimes even different teams. And then naturally, these sort 00:04:23.180 |
of like bolt-on non-integrated AI features and products start to proliferate. So what are some of the 00:04:29.340 |
causes of this sort of AI sideshow that we're talking about here? The first is that companies seek to 00:04:35.180 |
mitigate the risk associated with AI by quarantining into specific corners of the product, albeit there 00:04:41.020 |
like is new risk here, right? There's this new reliability question you have to ask yourself, 00:04:45.180 |
which is like, can I even get this feature to work reliably enough to drive value for customers? 00:04:50.140 |
Second, we see teams prioritizing the technology over their customer needs. They become the hammer in 00:04:59.100 |
search of the nail. Rather than trying to solve their customer problems by harnessing the technology, 00:05:03.660 |
they're just trying to find any manifestation of that technology. And we see this manifest in a bunch 00:05:08.460 |
of predictable ways. We see teams building chatbots because chatbots demonstrate AI capability, 00:05:13.740 |
not because customers are actually struggling with support. We see companies building document 00:05:18.060 |
summarization, again, because it demonstrates capability, not because their users are suffering 00:05:22.620 |
with information overload. And finally, we see companies creating this kind of top-down, 00:05:29.820 |
the pushing solutions out from the top-down rather than setting that top-level strategy and letting 00:05:34.620 |
the bottoms-up discovery process be the manifestation of that priority. 00:05:38.540 |
So how do you avoid the AI sideshow here? The key is to integrate and align your AI and your product, 00:05:47.580 |
and integrating AI risk into planning is a critical part of that. There is this new risk we're talking about, 00:05:55.180 |
but instead of being shying away from that risk and trying to quarantine AI to specific corners 00:05:59.980 |
of the product or specific teams, you need to deeply integrate that into your product planning. 00:06:04.700 |
And this will require like some new muscles here, right? Like you need to kind of build these systems 00:06:09.100 |
for evaluation, for testing, because if you're doing good prototyping and testing, you can at least kind 00:06:14.220 |
of wrap your arms around that risk and know how to handle it. And then second, start with the customer 00:06:20.220 |
problem. If you're inventing new problems to go solve with the advent of AI, you're probably gone 00:06:24.700 |
astray here. And finally, like we talked about, enable that bottoms-up discovery process for AI products. 00:06:31.260 |
It's likely your product folks who are boots on the ground every day, who understand the right solutions 00:06:36.540 |
here, give them the space to experiment, prototype, and importantly, fail fast, but set that top-line strategy, 00:06:43.820 |
and then allow the bottoms-up discovery process to take place. This is how you ultimately manifest AI 00:06:49.820 |
products that feel like a natural and cohesive part of the product experience, rather than feeling bolt 00:06:55.420 |
on. And that's ultimately the like, the hallmark of good successful AI integration are AI products that 00:07:02.540 |
need not announce themselves as AI, but rather just solve the customer problem better than what came before. 00:07:08.940 |
So the north star that Aliza and I are talking about today are AI products that are deeply and 00:07:13.900 |
dynamically integrated into your product ecosystem. But the only way you get there is by aligning your 00:07:19.340 |
strategy, your teams, and your roadmaps accordingly. And importantly, avoiding the AI sideshow. 00:07:25.900 |
This is admittedly like an audacious north star. And especially if you're kind of stuck in this sideshow 00:07:32.700 |
model, like how do you find your way out? This is where we think this crawl, walk, run approach comes into play. 00:07:38.140 |
We're all new to building generative AI products. Like to some degree or another, we're all building the 00:07:45.020 |
plane while we're flying it. The most successful teams we see here are those that crawl, walk, run 00:07:52.060 |
their way into this new era of kind of generative AI products. Because what that allows you to do is it 00:07:57.740 |
allows you to sort of build the capability iteratively while laying the foundation of that AI functionality 00:08:04.380 |
throughout your product suite. So I want to like walk through an example here. With an example, we'll 00:08:10.540 |
take like a customer support SaaS company, let's say they have like a shared inbox feature that customer 00:08:16.460 |
support teams come on to work out of mature product, but they want to start integrating AI. So in this 00:08:22.300 |
crawl phase, you're starting to build embedded AI experiences. You're likely in this phase, not building a 00:08:28.940 |
whole lot of new product surface area. Rather, you're just like adding AI on the back end and 00:08:33.740 |
starting to kind of accentuate and accelerate the existing functionality you have. If we take that 00:08:38.140 |
customer support example, that might look something like, you know, building a feature that uses semantic 00:08:43.180 |
search to like surface previous similar questions to help the user ground when they like are responding 00:08:48.780 |
to their customer. And then in the walk phase, this is where we're starting to build more contextual and 00:08:55.340 |
personalized AI experiences. Here we might actually we are starting to build like new product surface area, 00:09:01.980 |
but we're probably not at the point yet where we need to like fundamentally rethink our core app 00:09:05.900 |
architecture and our UX. If we go back to that example, that might look something like, you know, 00:09:12.140 |
building a feature that will like suggest a draft ahead of time so that when the user comes in, 00:09:17.180 |
there's already a draft there ready to go for them to start from. And then finally, where we land when we really 00:09:22.140 |
start to run, this is where we're building those dynamic, interoperable and integrated AI experiences 00:09:28.060 |
throughout our product suite. This is the stage where you do start needing to like re fundamentally 00:09:34.060 |
rethink your UI, your UX and your app architecture. Because now your AI features like if we go back to 00:09:40.540 |
our customer support example, it might look like an autonomous agent that can triage issues respond to 00:09:45.100 |
customers. But importantly, it's operating across the product and and feature set. And in order to incorporate 00:09:50.860 |
that kind of functionality, you do need to start rebuilding core surface area and like starting 00:09:55.580 |
to revisit your UX. But importantly, along the way, you're not throwing out functionality, 00:10:01.180 |
it's building on top of each other, it's that functionality is building as you go, right, 00:10:06.060 |
you're just extending on it. And importantly, even at the crawl phase, you're still building embedded 00:10:10.940 |
functionality, not this sort of like bolt on non integrated functionality. So I'll pass it to Aliza now. 00:10:17.580 |
Yeah, so let's walk through a tangible example here, because there's a lot to unpack. So 00:10:24.140 |
this problem space, I feel like everyone knows this, I've been living and breathing it for a few years. But 00:10:29.500 |
HR service delivery or employee self service is all of us work in jobs, or you're running a company, 00:10:36.540 |
your employees need to be able to get their questions answered quickly. And if they can't get those questions answered, 00:10:42.380 |
they need help from a support person. So through a case, this could be a live agent, etc. So we've spent 00:10:48.780 |
a lot of time working in this space. This is also where some products have found product market fit, 00:10:53.580 |
especially with early sort of gen AI solutions. So where we started to, I would say, crawl with the 00:11:00.940 |
technology, this was within our help product. So help has two components, there's a knowledge based solution, 00:11:07.820 |
there's also a case management solution. And so early days, we took a look at the tech and said, 00:11:12.940 |
where can we use gen AI really to affect change with customers. And so I know knowledge base has become 00:11:19.020 |
like the back end for GPTs and just a best practice. But at the time we said, okay, we've got content gen in 00:11:24.620 |
here, we've got translations in here, this is the content that's fueling the answers to all of those 00:11:29.500 |
questions that employees are asking. And so there were two key features. So one was actually content 00:11:35.020 |
authors, so they might come into an editor like this, they are going to upload, say, a policy doc. So 00:11:40.060 |
imagine a benefits policy, like 20 plus pages long, they don't want to necessarily write that article, 00:11:45.660 |
right? But they could have the AI ingest it, create an employee FAQ. In this case, we had talking points for 00:11:51.980 |
managers, and they're able to get a consistent format. So the other thing I would mention is, 00:11:56.540 |
we're thinking about content at scale. So this isn't for small sort of SMBs, this is large enterprise, 00:12:02.300 |
who have content teams of say, like three to 15 people. And so you need to have a united sort of 00:12:08.060 |
voice around that content that's coming out. So on top of that other feature, we put this translations, 00:12:14.460 |
which you can see in the GIF here. In just a couple of clicks, I can go in and translate into one of the 00:12:21.340 |
34 different languages that we support. And you see we added on the left hand panel here, 00:12:26.460 |
and the ability to actually manage versions as well. So I might have my base article, 00:12:31.020 |
I'm generating talking points in English, and then I want to translate into French and Spanish, 00:12:35.420 |
maybe Japanese. And you can see that you're managing those versions as well. 00:12:40.620 |
A couple of things I want to call out here. Yes, we're using Gen AI and translations. But this isn't in 00:12:48.940 |
in your face sparkles and chatbots and text fields all over the place. This was built for users who 00:12:56.140 |
didn't know about Gen AI, this is 2023, wanted to be able to kind of get in and use the features without 00:13:02.780 |
actually understanding the functionality. And it also, you know, keeping that human in the loop, we want to 00:13:08.460 |
have the disclaimer around AI. And so we make sure that we've got enough little purple sparkles to 00:13:14.140 |
let them know what they're using. But it's not the entire experience here. So this allowed us to go GA 00:13:21.340 |
in 2024, or August, I should say, and sort of, I would say, kind of crawling with the functionality. 00:13:31.740 |
So on top of that, so that's our content teams. So then we moved into what I would say is walking. 00:13:38.940 |
This was, now we have our content drafted, but we actually need to solve the self-service problem. 00:13:44.780 |
So as a manager, I might need to come in, Elaine in this case is trying to do a location change to San 00:13:51.100 |
Francisco. And she knows a lot of the fields, but not all of them. And so she now has this sort of 00:13:57.020 |
contextually aware co-pilot Workday Assistant that lives across Workday that she can sort of prompt. 00:14:03.820 |
A lot of us are familiar with this functionality, but a couple of points I want to make here. 00:14:08.220 |
One, we have the contextually aware suggestions, so it knows what's happening when I'm on the page. 00:14:13.500 |
Also around the data processing, if you're looking at a help article, it's generally customer content, 00:14:19.340 |
which is sensitive, but not nearly as sensitive as PII or personally identifiable information. 00:14:24.700 |
Think about these tasks more on safe pay or compensation, things that are really sensitive, 00:14:29.340 |
where employees are putting really sensitive information in. So this is the next level of 00:14:34.540 |
sort of walking with the capabilities. The other piece I'd mention is that this was a platform 00:14:40.380 |
capability, meaning that we had to be working across our suite. So we have HCM and financials, 00:14:46.300 |
think benefits, procurement, core HCM, etc. And so there's a higher level of sort of top down and 00:14:53.180 |
bottoms up alignment that had to happen to get these capabilities out the door. 00:14:57.420 |
Then finally, running. So extending the same use case here, you may have seen a few months back, 00:15:05.900 |
we announced our agent system of record. A subset of that functionality targeted towards, again, 00:15:11.900 |
those employees and managers was really around the agentic capabilities behind the Workday Assistant. 00:15:17.980 |
So again, our users don't necessarily want to know or have the sort of technical expertise around agents, 00:15:26.540 |
but we still have that work happening behind the scenes, where our assistant becomes a lot more autonomous, 00:15:32.620 |
proactive, listening to policy changes, notifying us with suggestions as well. And so you can see just 00:15:39.100 |
thinking through this at scale, there's a much higher level of, I would say, sort of top-down strategy with bottoms-up execution that then happens, 00:15:47.660 |
threading the needle across these different product experiences. 00:15:51.820 |
So you can see here we've kind of gone from a single product within a SKU all the way across our sort of core platform, which some of you may know or not, 00:16:02.460 |
but we serve like 60% of the S&P 500, so a pretty broad group. 00:16:08.220 |
So where we would land with all of this when we talk about not making AI a sideshow, 00:16:13.980 |
we're not telling you to stop working on agents or stop caring about AI, but understand that these are 00:16:21.900 |
stepping stones in terms of teaching your organization, training up your organization on what it means to 00:16:28.620 |
actually be building impactful AI experiences. And so as you sort of, I would say, mature as an organization, 00:16:36.700 |
ideally where we want to get to is building dynamic products. I'm hearing some of this today in some of 00:16:42.620 |
the talks, if you heard Sarah or Brian talking earlier, about building purposeful sort of vertical specific 00:16:49.900 |
products. I think it's really interesting when we start thinking about dynamic products in terms of 00:16:56.620 |
new problem spaces. I don't know if anyone else feels this way, but sometimes I feel like we're 00:17:01.180 |
solving yesterday's roadmap with just like a much more powerful technology. And so as we digitize 00:17:07.340 |
new data, new inputs in terms of our environment and spaces, we can see the problem space of the products 00:17:13.900 |
that we're creating really sort of extend. I think especially with multimodal, this is where this 00:17:19.980 |
gets really compelling as well. When we have frictionless multimodal experiences that interoperate, 00:17:25.900 |
I would say interoperability and RL are still pretty relevant within the agent sphere. 00:17:30.780 |
But when we think about dynamic products that are sort of responsive to your environment, 00:17:35.740 |
this is where we really start to see, I would say, the next generation of products come into play. 00:17:42.460 |
So hopefully this sparked a few thoughts, maybe some questions. If you want to connect, feel free to scan 00:17:50.940 |
our QR codes. Happy to connect if you want to drop us a note. We'll also be around the rest of the week.