LIVE SaaStr AI Day: AI Insights and Lessons Learned with Google Cloud and Thread AI

or cutting the amount of time where humans used to perform very repetitive or manual tasks.

That's great, Maya. You know, one of the things also that you and I discussed, because I know we've been able to spend some time together, is the revolution that we're in, in the midst of AI. There's early on a lot of experimentation going on, but especially large enterprises that had large legacy applications or datasets may have felt a little bit more constricted. They didn't exactly know how to bring these things to life. And I think what you and your team at Thread.ai are doing is, as you said, you're unlocking that potential. You're helping people realize that they don't have to move every application, to move every dataset. Right. And I think that's such a powerful, powerful element of what you and the Thread.ai team are doing.

Exactly. So one of the things— so since we know that the space is changing rapidly, so one of the key things that we invest in is how can we embed AI in a way that's like experimental where you can run different models and always take advantage of the latest and greatest and cheaper models without basically blasting things that you've worked. We make it easy to embed different integration points into your various source and sync systems and bring AI in a way where you can run different processes in parallel and in a way that's like observable and repeatable. And like you said, folks don't have to re— basically don't have to change their entire stack to take advantage of the latest and greatest models.

That's fantastic. When you— as I've told you, you're in kind of a unique spot because you're not only solving really complex business needs today, you're also looking into the future a bit. And so I'm curious, when you look ahead, when you envision kind of how AI is reshaping the way organizations operate and manage their workflows, You're so close to that space. So can you help me and the audience here today understand how do you envision that AI reshaping as we look ahead?

So at Thread AI, our main thesis has been that AI should never be a workflow in and of itself. So we were less bullish on just pure AI applications. So we, our main thesis is that AI is always part of a workflow. And it's, it's there to augment and enhance and in some cases maybe replace how things are done, but it's never a workflow in and of itself. And we've seen kind of like some companies who have kind of like followed that same philosophy where you're— it's more— it's easier now to, whether you're in your sheet or doc or continuing to work as you do here, you're able to take advantage of AI at the most in a very natural way without jumping around between applications. So for us, basically, that is our main philosophy and the way it— and the way based— and the way we think about it is how can we have these different kinds of connectors and also things like mentioned, like ideas of human in the loop where you want to provide these intuitive insertion points into workflows.

That's great. I love your comment of you're kind of an— AI is a natural part of the way that you're getting work done instead of stopping working the way you're used to working, go use AI for a little bit and then return. You're kind of fusing those two things together with the capability. And your point is you see that only accelerating then as we look more and more into the future, that AI will become even more natively as a part of the way we're going to work.

Yes, exactly. And of course, we're, we're very bullish on infrastructure and applications that make it easy to see these things in a way that's observable and repeatable. And so things like workflow orchestration systems or like observability tools are going to be extremely important as folks start integrating AIs into, into both like backend operations or, or even mission-critical workflows.

Right, and I think that definitely from your experience, as I'd said, doing really hard things in hard regulated industries, I think you and your team clearly have a lot of expertise on exactly like security, trust and safety, observability, all of the requirements that I think we all know are going to become even more and more, right, a part of what's going to be happening in the future. So yeah, appreciate your thoughts on that. One of the things that I know you're asked quite often, but I think it would be amiss if we didn't use this chance to, to learn from your mistakes, right? You've done some interesting things. You're a founder in a really interesting time. And so I'm curious, Maya, when you reflect on this journey that you've been on, would you mind sharing openly some of the mistakes that you've made so that ideally we can maybe help some of the folks on this line avoid those same pitfalls.

Yeah, I think there's a few angles to this. There's the general key kind of like failure mode, which isn't necessarily new to AI or GenAI. And I think it also applies to automation in general, is whether or not something should be automated with AI, not simply can it, but should it? And I'm sure folks are familiar with a lot of historical cases where topical AI was used to make decisions where AI shouldn't be making decisions. So things like insurance claims and like predictive policing. And there's many, many of these examples. And I think the same, the same philosophy also applies to GenAI. Now it is easier than ever to potentially produce sloppier outcomes. Some of the tooling has made it easy to do that. So I think the same kind of like evaluation process and philosophy should be applied, um, to that. Um, there's also, um, obviously, um, when we work with a lot of companies, people are thinking about the, the general risk of embedding, um, different kinds of— these models are obviously like probabilistic machines, so they're— even as these models continue to improve, there's still kind of like a risk inherent in using them. So folks have, have spent a fair amount of time thinking about which processes can use AI and which ones can't, and also the risk tolerance that's associated with that. I know earlier on, I guess like there was a lot of hype and we've seen some cases where folks have maybe automated too much, where you have these like chatbots sometimes offering customer service solutions or insurance claims and for products that don't, for example, exist. So I think it's just like taking a step back And making sure that folks have the right checks and balances and guardrails. Obviously, that's super high level. And these things can mean very different things and using very different tooling. But the idea is inherently understanding the risk that comes with these different tools.

It's fascinating to hear you say that not every problem can be solved with AI in the midst of everything we read and hear, right, is kind of AI in our faces. But that's a powerful observation for all for all of us to take away is that, especially in times of innovation and revolution and a lot of excitement, we can tend to say, "I've got a hammer, now let me find that nail." So I appreciate your perspective there of AI is powerful when used responsibly and in the right way. And I think, again, based on your experience and what you're doing at ThreatAI, that's a great example of that. So thank you for sharing that pitfall. And I think that's something we can all make sure and try to avoid. If we can talk a little bit about kind of how you think about partnership, obviously we've been able to work closely with you and the Threat AI team, making sure that we're doing our best to kind of help be an innovation partner with you. And I think as a founder, you obviously think critically about who you're going to partner with, right? Who are the companies that can help you be successful? Can you reflect a little bit on that journey when you think about the collaboration with companies like Google Cloud as a founder, kind of what influences your decision and how do you think about partnership when you're trying to get so much done in such a compressed amount of time?

Yeah. And for me, I guess I've been working with Google products personally for a very long time. I remember earlier the New York Times around 2015 when GKE, for example, was still fairly new. And it's been super exciting to see how Google Cloud has very quickly caught up and with a heavier emphasis on developer kind of like first tooling. So obviously, like besides kind of like my bias for Google tooling and as a founder, I think for us, for me and my co-founder, I think it's been working with Google has been very, very exciting, very interesting in that everyone that we've talked to has really understood our multi-cloud strategy and has basically supported us in ways that, that are very tailored. Obviously, as a startup in the automation and workflow orchestration space, our needs are very different from some of the other startups in the Google ecosystem who are like foundation model shops. Their needs, for example, like GPUs, are very different from our needs. And I think having this very intentional working relationship with Google where folks understand where exactly in the value chain you sit and connecting you with other, for example, ISVs is very important for, for our growth stage as a startup. I think really it's like this intentionality of like not just like applying a cookie cutter kind of like model for every startup, but understanding where you can add most value and how you can come up with this more nuanced and symbiotic go-to-market relationship that's very different from other AI infrastructure companies, for example.

Yeah, thank you for that. I appreciate that feedback, and your words obviously are mirroring what it is that we're trying to build every single day, right? You made a couple of key comments, and one is our obsession with developers and architects, right? We believe people that are dealing with applications and data are solving the world's problems, and so we've always been a company that's obsessed with developers, and I'm glad to know that heritage with you there has been helpful. Also really importantly, not just Google Cloud, I would say for all the founders that may be on this line learning from you is just the opportunity to find technology partners that are willing to not treat you as kind of a faceless customer, you know, with kind of using programmatic APIs or just technology APIs, right? It's kind of putting humans together to solve hard problems and being able to meet you where you are on your journey is obviously suited super important for us as well. So thank you for that. Let's do, if we can, before we kind of open up to some of the questions that I know we're getting from the group, I'm curious, back to the advice. Keep in mind the folks that are joining us today are CTOs, some of them may be Chief Revenue Officers or Chief Marketing Officers. I know there's a lot of builders and inventors that are wanting to learn from you today. So if you had to think of some really great advice, no pressure, Maya, but some really great advice for other startups and inventors specifically on how to innovate with AI. What would that advice be, Maya?

I think honestly studying the open source ecosystem and learning from the giants who have spent a lot of time thinking about some of these problems. So things like durability, durable workflows, or other interfaces. Just because AI is new and exciting doesn't mean that the fundamentals have really change or they're not there. There's a fair amount of literature and like, and for us, the key thing is don't reinvent the wheel. There's a lot of great abstractions that you can enhance and fork and augment to build something special. And also learning from the shortcomings of some of the tools. And I know these shortcomings don't have, are not always like a product or technology shortcoming. But for some companies, it's simply a go-to-market or a fit or nuance that comes with what markets do you have access to as a founder based on your background, for example. But I think, yeah, it's, it's an exciting space and where we're learning also as we go. And yeah.

I think that's super helpful. I know for the years that I've had the chance to work closely with founders like yourself, I keep hearing the most scarce resource that we have is time, right? We can raise more money. Ideally, we can hire smart people. But I think what I hear in you, what I'm hearing you say there is lean into other people's successes, lean into other people's failures as well to try to compress time, right? We're all trying to do quite a bit in such a short amount of time. So that's, that's excellent guidance. Before I turn it over to Amelia, who I know has a few questions for us, I do want to ask you one more question. And that is, I know we're talking a lot about what excites us and all the positive influence and impact that we think is going to be coming from AI. As a startup founder deeply embedded in this, in this revolution, Maya, what are your biggest concerns regarding the future of AI? Would you mind just kind of sharing how, how do you look ahead and think about that concern or risk vector?

I think, yeah, we've touched a little bit on it, but the biggest concern is seeing how sometimes some half-baked solutions make it into maybe the wrong hands or without the right checks and balances. So a lot of successful open source libraries have been battle tested in the past when you think about things like Kubernetes, for example, but But now we're seeing kind of like a little bit of a rush to push things to market when we haven't thought about all the different things like security and authorization mechanisms. So we're seeing a lot of half-baked— I don't want to say the word like agents, but it's exciting that everyone is excited about the space. But my biggest concern is sometimes these solutions might not always be in the hands of folks who are, who are, who come with a mindset of like doing no harm and making sure that they're the right custodians of these technologies. But yeah.

Yeah, completely hear you. The concept that I know we've spoken a lot about and is deep to our value system is the responsible application of AI. And so much we all, I love your term custodian. We truly are all ambassadors and representatives of how this innovation is going to change the world positively. So excellent, excellent point. Amelia, I could go on and on with Maya asking her questions, but I know that we received a number of questions from some other folks. So keeping my eye on the clock, why don't I turn it over to you and we can answer a few raw Q&A if that's all right?

Yeah, let's do it. I think it's perfect timing. Raymond, I'm going to ask you to bring your question if you want to come off mute. Otherwise, I'll ask it for you.

Absolutely. And thank you for the amazing presentation and for sharing your insights, Maya. So my question is, As a founder and builder with the speed of innovation and how these models are changing, how do you keep up and how do you build around that without disrupting the, the, the, your own business, your own business model, if you will, which is also based on some of these LLMs that are changing so fast?

Yeah, that's a great question. So for us, We try to bake some of that into the product itself. So for example, without digging too deep into the product, we have a concept of, say, like a functions registry. And that we, we kind of like took the approach of focusing on like protocol as integration. So things like adding gRPC bindings and, and REST bindings and GraphQL bindings so that we can easily absorb the latest and greatest models. Without necessarily integrating with every single SDK or every single library that's going to come out. And the idea is to make that self-service so that our customers can, can do that kind of updates and integrations where we don't even have to know about it. Knowing— so I think like focusing on the fundamentals of inputs and outputs. Yes. Like some of these multimodal kind of like models throw a monkey wrench in some of that. But, but having these primitives baked into the product so that, so that users can take advantage of the latest and greatest models. And then also knowing that these models are going to continue to be smaller and cheaper. So without taking, for example, like a hard dependency on a single provider is going to be critical.

Perfect. Thank you.

Oh, unmuted. Okay, thanks for the question, Raymond. I'm going to bring up Anjay, who has a question as well. So I'm going to bring you up. I know you're used to this from Workshop Wednesday.

Thanks, Amelia. I'm on a slow internet connection, so I won't be on video. First of all, thanks for bringing on such a good guest. You have a very good background. My question is around the use of GenAI in the enterprise, especially around automation tasks. The expectation in enterprise is very high degree of repeatability and reliability. Are you running into issues with GenAI not being that reliable, given how GenAI works? And if so, how do you deal with it?

So yeah, there's a couple of things. So we've been fortunate, yeah, like Darren mentioned, to partner with Google, where there are certain kinds of security and data guarantees on some of the models in the Model Garden Vertex AI ecosystem. But obviously, to your point around just general nondeterminism, the way we think about it is we make it very clear where, if you are using a model in a particular part of a workflow, that that risk is flagged. And we do kind of like have other ways to augment. We add things like citations so that you can go back. And if you wanted to have like a manual human reviewer go back to where that answer came from, we make it easy in the product to do that. Obviously, the human review check can't make it worse, like it can't be a brute force process. So like we, we're starting to think about some of these like UX and DevEx integration points where first of all you have to surface the fact that there is, you are, whether or not you're using GenAI. And in some cases some enterprises want to know exactly what model, but then you make it easy to pause different workflows and have a human reviewer. And then you add like things like you don't fully, some workflows should not, should never be automated completely. So we basically add all the checks and balances so that folks can bring in the right set of reviewers who might have a very different technical profile from the folks who created, say, the workflow.

Got it. Just a quick follow-up. Have you toyed with or considered different independent models being given the same, same output or the same input to see if they match? If they match, then it's more reliable, and if they don't, then then can human involved?

Yeah, so our, our infrastructure is model agnostic. So you can— our unit of work is something called the worker, which is made up of different steps, and you could potentially bring up as many of these with as different, with different models running side by side. But there are folks who have invested obviously in like in the evaluation space, which hasn't— which there are a set of tools to allow with some of that. But from our workflow automation perspective, we simply make it easy for folks to run different workflows, like you said, with models in parallel so that they can decide on which output makes sense for their particular use case.

Got it. Thanks, Maya.

Thanks, Ajay. Okay, Latif, we're going to try and get your question in as the last speedfire question here. Okay.

Hi. Thank you, Maeda. That was very, very interesting and insightful. My name is Latif Hamlani. I'm the founder of an AI agentic startup that will automate channel management workflows. And, you know, I was very interested in what you said about embedding AI in workflows, that AI is not the workflow, right? There's human in the loop. And I would be very interested in following up and doing a deeper dive into that conversation and see how your technologies can help us build our product. How do we— how do I engage you?

Yeah, of course, would love— would love that. And we work with all kinds of companies, some technical like yourself, some more on the vertical side of things. And happy to share with you what our human-in-the-loop interfaces look like. But the idea is for us, our infrastructure makes it easy to have these guarantees where if you do have like a long-running workflow, you can easily pause it. And we take care of all the, all the infrastructure around the correlation IDs and making sure that other folks can come in and either review, edit, or reject the results of a model easily through different kinds of mediums, whether it's like through the phone or like leaving a voice note and/or logging into our platform, and then the workflows are basically resumed. So yeah, happy to do a deeper dive and share with you.

Cool.

Thanks, everyone. I know, Darren, you answered in the chat just what I was going to ask you for the parting advice. There's a few folks asking how they can get help on Google Cloud. So yeah, connect with Darren on LinkedIn. That's the best way to find him.

Yeah.

Yeah, absolutely. We have a team of folks that can help across an entire slew. So absolutely feel free to send me a LinkedIn request. I'm very, very happy to engage in support directly or get a member of our team who's super excited to do so. So thanks, Amelia.

Yeah.

Of course. Yeah, Darren and Maya, thank you so much for doing this. So, so great to have you here for AI Day. I know Google Cloud will be at SaaStr Annual talking more about AI and all things AI and SaaS. So we'll see you guys on the next session, and we'll see you at SaaStr Annual. Thanks again, guys, for coming to AI Day.

Thank you, Emilia. Thanks everyone.

Have a great day. See you on the next one.