How Writer is helping Fortune 500 companies become AI-forward

0:00Starting a business comes with its share of ups and downs, which is why staying true to your vision is essential. a non-negotiable for Romeo and Milka Bregali, Capital One business customers and co-owners of Ra's plant-based restaurant in New York. Romeo and Milka took a leap of faith when starting their own restaurant, gutting an empty space and building it from the ground up, every pipe, every wall, every detail. But building from scratch came with a heavy financial burden, which is when they turned to their Capital One business card. With the flexibility of the card's no preset spending limit, they were able to spend more and earn more rewards while bringing their vision to life.

0:36Today, Raz's success is proof that with passion and the right support, it's possible to make your dreams a reality. Learn more at CapitalOne.com slash business cards.

0:52You know, one thing that AI never asks is, why did you ask me that question? AI is always like, oh, I'm so glad you asked, or you are brilliant, or thank you for asking me that. Let me tell you, blah, blah, blah, right? It's incredibly subservient. That's Mai Habib, CEO and co-founder of the AI studio, Ryder. And I think the meaning of being human is to not be, right? At Ryder, we've got this saying, break rank, break glass. And that's what I think it means to be human compared to being AI. Mai and the team behind Rider take no issue breaking rank in Silicon Valley. They're shaking up the status quo with an army of AI agents that can handle all sorts of enterprise tasks that humans normally do.

1:46From writing website blurbs to drafting business proposals. These are often time-consuming menial tasks and automating them is saving companies millions. Companies like Salesforce and Uber, to name a few on their client roster. Rider, which Mai founded in 2020, is now valued at$1.9 billion. On this week of Pioneers of AI, Mai and I are digging into the nuts and bolts behind one of Silicon Valley's rising disruptors. We'll talk about human-AI collaborations, data privacy, and building a unicorn startup in the age of AI.

2:29I'm Rana El-Khalyubi, and this is Pioneers of AI, a podcast taking you behind the scenes of the AI revolution.

2:45Mayi, welcome to Pioneers of AI. It's so great to have you on the show. Thank you, Rana. It's so nice to be here with you. So we have a lot that we share in common. We're both of Arab descent. You, I believe, were born in Lebanon and then grew up in Canada. Is that right? Yeah. Yeah, that's exactly right. We went back and forth a lot. So I definitely feel very Lebanese. Do you still go back? I haven't been in a while since before COVID. You know, you always say next year, but inshallah now, actually next year. Inshallah, inshallah. That's great. We're also both young global leaders at the World Economic Forum.

3:22Yeah, it's such a great community. It is. Yeah. Okay, my most burning question for today, actually, is that you and my daughter, Jenna, both graduated with a major in Near Eastern Languages and Civilizations. Yes, indeed. Yes. But somehow you ended up in tech. So I feel like there's hope for my daughter. Yeah, I mean, it's going to be the only industry left. It's so exciting. After years of working in venture capital, Mai founded her first company, Cordoba, in 2015. It started out as a software company helping their clients write clear, consistent content across languages. I was a machine translation company, and we did everything from human translation to machine-generated translation, all in a platform that helped enterprises really build highly localized products and doing it in an engineering-friendly way.

4:21Cordoba allowed companies to enter new markets by translating all of their content into new languages with ease. But when Transformers hit the scene, the kind of neural network architecture that bred LLM's, Mayi saw new opportunities. Her work didn't need to stop at translation. Generative AI opened the door for all kinds of business automation. In 2020, she co-founded Rider. So you describe Rider as a full stack AI powered platform that helps your customers and your, you know, your enterprise customers become AI first and agentic first. Unpack that for us. What does it actually mean? Yeah. So in 2023, we said dominant design and generative AI looked like LLMs and retrieval and guardrails and an AI studio.

5:16An AI studio because you really needed business and subject matter experts to collaborate. And we called it a dominant design architecture because we said this is the only way to build highly reliable, effective AI systems on top of highly non-deterministic technology. And, you know, two years later, everybody has an AI studio. Everybody has an agent builder. But compared to writer, they're still very thin wrappers on top of the LLMs. And what we've been able to do year over year is really invest in those AI native primitives that make it really easy to actually connect LLMs to data, to workflow, to guardrails to produce highly precise use cases, insights, content in highly regulated environments.

6:07So, you know, agentically, we are doing client onboarding. We are doing KYC. Know your customer. Yes. We are doing orchestrated digital marketing. And, you know, it's agents that just work in a way that is just not happening on any other platform. Basically, Maia is saying that Ryder offers a lot of capabilities to their clients, whether that's integrating their AI agents into marketing strategies or using them to help automate fraud detection. She says that her clients are seeing results. So it's been, you know, super incredible to see our customers really lead their markets. You know, customers like Uber and Salesforce and Accenture and NVIDIA are using Rider internally to really streamline all sorts of operations.

6:59So, Mike, can you walk us through some examples of how your customers are using Rider in their workflows? Yeah. Uber, we've got an incredible support team there that is using our agentic AI in their documentation and support processes. So they use writer agents to classify JIRA and Slack requests. They assess urgency and they generate updated responses and content, right, based on that. The AI agent is staging the updates in the CMS, and they're publishing it upon approval. And so there's this very artful orchestration between the deterministic and non-deterministic aspects of this workflow. And it's incredibly amplifying to what the human teams are doing.

7:47Tons of manual handoffs reduced, right, reducing content maintenance and production costs, et cetera. RFPs and business proposals are also incredibly rich for agentic AI. A company called Commvault, a cybersecurity company, AI agents from Writer, we're automatically pulling RFP files from Salesforce. We're associating them with the right opportunity. We're setting up a dedicated Microsoft team channel where teams can actually collaborate on the AI-generated responses that are based on previous RFPs that the teams have been successful in. And then once they're verified by people right there in the channel, we are automatically actually producing the final PDF of that proposal or that RFP and then sent to the client.

8:31This is really interesting with my investor hat on one of the areas that I'm very excited to be investing in is this idea of an AI employee, right? Like an agentic AI, which is basically embedded into a workflow and is able to start with like, you know, a number of tasks, but increase its ability and its scope of responsibility over time as it learns and evolves. But it's working alongside human teams. And your example of the kind of working on the RFP proposal, I have this picture in my mind. There's the Microsoft Teams channel with all these humans and the agentic AI. Is this kind of how it works?

9:09So I don't like to anthropomorphize AI, but the mental model of managing agents like employees is actually more apt than managing agents like software, right? But you actually do need a agentic PM that can think about how you get these agents to do things reliably. Our head of product, Matan, likes to say, yes, agents don't follow rules, but you can get them to behave. And really taking the build process from, you know, this very deterministic software-based approach to how you actually align behavior of an agentic system. Using the mental model of an employee is actually really helpful. Like in Rider, we build blueprints.

9:57We don't build workflows because that's very deterministic and that's not what we're really doing here when we orchestrate these agents against a goal, right? These are really goal-oriented and can take lots of different paths to achieve a goal within guardrails. But, you know, your blueprint for that agentic system is kind of like your job description, right, for an employee. Your prompt setup, your grounding, that's like how you onboard an employee. Your escalation paths, your fallback agents, those are like your managers, you know. How you do performance reviews, how you do feedback and coaching, that's your outcome tracking.

10:33That's your evaluation setup. That's your retraining. And then we've also built just lifecycle management for Agentex. So how do you deprecate them if they cost too much relative to the value or folks aren't using them anymore or more or they've been replaced by something else? That's the equivalent of like firing? Yeah, yeah, exactly. Termination, right? Yeah. And so the mental model of them as employees, I think we can, you know, think of them as a super employee, right? An employee that really needs to be a member of multiple teams to be able to get their objective achieved. We're going to take a short break.

11:11When we come back, how Rider is increasing enterprise visibility in a post-SEO world and the new job opportunities that Agentic AI is creating. Stay with us.

11:37If you've spent any time building AI products or leading technical teams, you know this. Transformation doesn't fail because of ideas. It fails because teams can't move together. Enter Atlassian's Teamwork Collection. It has planning in JIRA, documentation in Confluence, video updates in Loom, and now AI agents in Rovo, which connects the dots across your work so nothing gets lost. It's one AI-powered teamwork platform designed for how modern teams actually build. Learn more at Atlassian.com slash TeamChanger. That's A-T-L-A-S-S-I-A-N dot com slash TeamChanger.

12:26When you onboard a customer, what does this look like? Yeah, so we are actually a verticalized platform. And so when we onboard a CPG customer, it looks a little bit different than we onboard somebody in retail, onboard a technology customer. We have libraries of hundreds of prebuilt agents, and they're quite powerful. So, you know, a Qualcomm that does really sophisticated ABM, that's account-based marketing, or a Salesforce that does incredibly powerful comms and PR, right? Their use cases are going to be different. And what we're able to do is onboard them into libraries of pre-built agents that are ready for configuration for their systems, their tooling.

13:15If you are doing GEO, that's Generative AI Engine Optimization. What does that mean, actually, by the way? It's a new term. Yeah. AEO, GEO is the new SEO. You know, when you really think about how much of the consumer market is turning to LLMs for search, right, really thinking about your web footprint as a brand and what it's going to mean for the visibility of your products, your SKUs, your viewpoints. You absolutely can structure, agentically restructure all of your sitemap to be much more amenable to the kind of web search patterns and the training data collection, right, of LLMs. I'll give you like the tiniest kind of tactical thing that we do.

14:08so much of conventional search is keywords-based, but LLMs really like nice, comprehensive paragraphs and sentences, right, in their responses. And so actually being able to mirror your content automatically, right, it's exciting to get to take solutions to customers where generative AI is a solution to a problem that generative AI caused. Right. Right. But, you know, it really gets to the change management complexity. Yeah. So I kind of want to like simplify this transition from search engine optimization, which is essentially search based on keywords, which is we've had for the last, I don't know, like 20, 25 years to this transition to generative AI engine optimization.

14:53And I love that term, which is essentially modifying or adapting your content so that it could be found by a gen AI search engine, like a perplexity or whatever's out there. That's super cool. And I think that's going to also create new business models and monetization, right? Yeah, absolutely. I mean, I think the consumer LLMs are likely to do, you know, the kinds of things that Snapchat did, right? Like brand takeovers, exclusivity in certain categories. So there's definitely going to be an explosion of that. And I think the brands that really get there first will have first mover advantage.

15:31So a lot of the work you do, I imagine, with your customers isn't just about the technology, to your point. It's about change management and bringing kind of the organization and the employees on board with all of these new tools and technologies. I imagine a lot of our listeners are grappling with how to do that, how to become AI first, how to incorporate AI into their workflows. What have you seen that works? What are some of the best practices that your customers are using? Yeah. So too many customers are saying, help us become an AI-first company. In fact, there's a mandate to be an AI-first company, and they don't back it up with the proper resourcing, right?

16:13And we've got an incredible delivery team that works really closely, you know, no daylight between us with our customers in helping them, you know, hands-on keyboard really build out solutions that can go live in a matter of days, matter of weeks. But the real scale of the program is going to depend on, do you have agentic product managers that can rewire workflows that have been built up over decades? Do you have AI builders who, you know, we can now address a whole range of software capabilities, right? It's like vibe coding on steroids at Rider, right? I mean, you're literally, this is, you're writing a prompt of a process and we are unfurling the agentified version of that, right, that you just really need to modify.

16:59The code is there already, right, that you're able to go out and modify. And so, you know, when you think about what that means is, you know, we have made building tooling and software a thousand times easier in the enterprise. They can go straight to like build it and scale it because the guardrails are really powerful. But back to your question, this is a lot for organizations to think about. And, you know, we are seeing best practices be, yes, I've invested in, you know, full-time folks focused on generative AI. And I think about it as two in a box down every team, engineering and the business, right, together.

17:39Because you as a technology team aren't going to be able to build agents for the enterprise in an isolated way. you need them there for a highly iterative collaboration. And that's just a different set of skills that, you know, we really help mirror for our customers in the people that they have put up to say, yes, this is our agent PM. This is our agent builder. These are the people who are going to build on writer. So agent product manager and agent builder, these are new types of jobs. Can you talk about what these look like? Yeah, absolutely. Your product manager, right, your junior product manager is going to be somebody who can absolutely transition to be an agent PM.

18:22But they've got to start thinking about the agent development lifecycle, right, versus software development lifecycle. And remember what we said at the beginning, like agents are systems that are goal-oriented, that can create their own steps and workflows, that can rebuild those steps as they try to achieve a goal. And so the requirements gathering is no longer, oh, legal needs a chatbot. It's, all right, I'm going to build an agentic system that cuts down my contract review time, right? And that kind of outcome orientation and the specificity of the problem that we are solving much more than here's what the UI should look like or here is what the steps should be to solve this problem.

19:08You know, that's a change in mindset. And so, yes, it's a process design that's an input for things that are existing processes inside of a company, you know, how you pay an invoice or how you onboard a vendor in your supply chain or how you onboard a customer into your mortgage product. But it is the behavior design of the agent that you are building to address those problems that we really want the PM to own. What differentiates Rider and what's your competitive mode? Yeah, this is a space that moves so fast. Speed and iteration is the only moat. And, you know, I think you can boil that down even further.

19:51Culture is a moat in this space. And I think, you know, when we look at our business, we're very proud that our top 30-ish customers are at 217 % NRR. You know, this is customers that are tripling spend with us, you know, global 2000 customers in the like trailing 12 period. And it's only it's only going up. And a huge part of that is the multi-year investment in the architecture blocks in a highly disciplined way, maniacally focused on the enterprise. And it starts with our AI research team. You know, we got to synthetic data before anybody else. We got to graph-based retrieval before anybody else.

20:34We got to self-evolving models before anybody else because we've got this really exceptionally fast feedback loop between the customer and our research team. You know, the partners that we work with, you have agentic AI too. And I think there will be a lot of winners. You know, your traditional software companies are going to spend a lot of time building agentic interfaces into their own data and into their own products. And it's a really big market for the workflows that don't exist because it's too hard to create these interactions between data silos, right? And the unstructured data that sits in people's heads that really define how to do something.

21:17And so we play that role in a really unique way. Unlike a slew of other agentic AI companies, Ryder is not using off-the-shelf LLMs. They've built their own models, and we'll get to why they made this decision after a short break.

21:46Meet Nicole Nicholas, Capital One business customer and co-owner of Ansett Uncles, a plant-based restaurant and community space in Brooklyn, New York, that got its start from a need for unity. The inspiration, it was born from the desire to create a space that felt like home, where we can connect community culture, good food, and come together with family and friends. That's how we birthed aunts and uncles. Nicole and her husband, Mike, were fulfilling their dream of bringing people together out of their home kitchen. But they soon learned that the demand for community was greater than they knew.

22:17It became overwhelming and we were like, we need home, but not in our actual home. We realized that there was also a need in our community for something bigger in our neighborhood. So we had to find a place. Moving from a home operation into a storefront was a huge next step, but Nicole and Mike were able to take it on with the help of Capital One Business. It's not for the weak as a small business. Finding resources is super important because that's the way you'll be able to manage and scale. We would have never done that without having Capital One to be able to help us along the way. The cashback rewards are very helpful.

22:53You know, it just gave us that runway to be able to breathe a little bit. Then you get to focus on the cooking of the food and making the experience great. To learn more, go to CapitalOne.com slash business cards.

23:09So we love on this show to take our listeners behind the scenes and kind of unpack what's behind the products you're building. You mentioned a few kind of key features of your products that give you a competitive moat. So let's talk about those. So first of all, let's talk about your LLMs. I think it's really interesting that you decided to build your own as opposed to build on top of other foundation models. How did you make that decision and why? Yeah, we were born as a native transformer company. The story of writer is the story of the transformer. And our Palmyra X4 and X5 are the workhorses of agentic workloads.

23:50So token context windows and the millions of tokens, breakthrough speed, the ability to process thousands of pages in 20 seconds, the ability to tool call in a microsecond, adaptive hybrid reasoning. And, you know, when you think about the way that you daisy chain so many AI components together when you do agentic AI, incredibly important to have models that are fit for purpose. These are not small models. They're, you know, 500 build plus parameter models. But we have architected the transformers to be highly efficient. So the total cost of ownership is incredibly attractive to customers when they really understand that retrieval and agents and the platform are really all in one.

24:39And the reliability, the consistency. The last point is really around transparency. We give our customers more transparency into how we build our models, how we train our models than they get from open source. And that makes a huge difference in folks' abilities to get through model risk. We've got a lot of customers in financial services and their ability to really have models that they trust under the hood. You know, we get to help them with data that, you know, they consider highly confidential and really access it in a way that they understand is very safe. We don't use their data for training.

25:18Models are actually trained on completely synthetic data. But we give them, you know, the visibility they need into training data, weights, architecture, the guardrails that we use, the audits to limit bias and toxicity. You get to audit the model's thought process. There's explainability built in. That is very unusual, by the way. That's very differentiating because most products out there, you're using the LLM as a black box and you have no idea what's going into it and also what it's learning, right? Yeah, 100%. Now, we are a Gentic orchestrator, so you could use, you know, an agent to call out to any other model.

25:56We're multi-model in that regard. And folks use Gemini for image generation. Firefly is a big partner of ours. But 95 % of the workloads are on the Palmyra models. So the Palmyra X5, you announced, is able to get to state-of-the-art performance basically three times faster than the other models and four times cheaper. And one of the key components or secret sauces, I guess, to this is your use of synthetic data. Back in my Affectiva days, we would use synthetic data generators. For example, if we're training, I don't know, we're training a model to detect driver drowsiness, right? The traditional way is to look for examples and, in fact, like go out and collect data of people falling asleep at the wheel, which is kind of really expensive because you want diverse human beings and diverse situations and whatnot.

26:43So instead, we had a synthetic data generator that generated synthetic humans, and we could kind of change the parameters, change the way they're looking, change the degree of drowsiness. So walk us through, how are you generating the synthetic data? What's the input to the synthetic data generation models? Yeah, well, it's secret sauce. Secret sauce. I won't tell you too much, right? From a data pipeline perspective, you know, what I can share is think about our use of synthetic data as the creation of training data that is precision created, right, for specifically the training of models. And so the way that we structure the data, that's the input to a model that we fine tune to actually synthesize data that is fit for purpose for training the models that we want to have certain types of behaviors.