[Insights] B2B AI Product Panel Recap✨

Question 1

What's the difference between managing a generative AI product vs. a traditional SaaS product?

Sherrie:

Running a generative AI product, especially in gaming, is all about making sure game studios and players trust what it churns out. Using AI to whip up dialogue is still kind of a wild frontier. So, it's all about feeling your way through the unknown, testing things out, and showing folks it's the real deal.

Tianlu:

In generative AI, particularly in our field of chatbots, the significance of data becomes even more pronounced. It's the real-time interactions with users that truly matter. Therefore, data, especially real-world data, is paramount for us. This shift underscores the primary difference we're encountering in our work.

Annie:

Regardless of the product type, the core principles of product management remain consistent: delivering value to users. However, in AI products, the quality of the model depends heavily on the vertical data used. Therefore, product managers should carefully consider the data sources and target verticals to ensure effective model performance for their customers.

Sandy:

Generative AI products are reshaping the enterprise purchasing model. Previously, purchases were top-down, with organizations buying tools based on demos and sales pitches. Now, with generative AI, individual users within organizations find value in products for their workflows, leading to bottom-up adoption. This trend emphasizes end-user experiences over traditional sales processes, resulting in more user-driven decision-making and less reliance on demos and sales calls.

Question 2

How do you utilize PLG (Product-led growth) in your work?

Sherrie:

We focus on building trust within the gaming community by offering engaging experiences. This has attracted smaller game studios to us, with a self-serve option for hobbyists, indies, and mid-sized studios. Larger projects undergo a sales cycle.

Tianlu:

We actually became more sales-led growth at this point. That's how our target customers work. They prefer relationship-building and more personal interactions. Another reason for that is that our conversational AI onto our customers’ sites or on their landing pages. Training our model takes time, causing it to function differently for each client. Thus, understanding their business is essential for customization.

Annie:

Sales business models in AI companies differ in their company size. Companies like Gamelab and Notion focus on individual and team subscriptions, while OpenAI targets big enterprises. Ads industry in big tech firms might focus on big accounts first to ramp up revenue before branching out to smaller ones.

Sandy:

One of the models I wanted to share with you all here is a model that we use at Descript - PLS (Product-Led Sales).

Product-Led Sales Model:

Instead of traditional outbound sales methods, the company analyzes user behavior and features used to identify potential customers within organizations.

Product-led sales can be an effective strategy to source leads from your user base. By qualifying leads based on usage data, they create a targeted sales list from their existing user base.

Question 3

Building cool new features vs. customized for specific enterprise needs, how do you prioritize as a startup?

Tianlu:

Since revenue is the primary focus, we assess whether we can replicate success with similar big clients and determine if customizing for them is worthwhile. Ensuring the needs of these significant clients is essential for both our current success and our ability to grow and serve even more clients down the line.

Sherrie:

We have a talented in-house team of game developers handling advanced features for our partners. They're skilled in AI, our tech stack, and game engines, collaborating with big names like Ubisoft and NVIDIA. We customize our platform product to meet their needs. Our lab team explores cutting-edge gaming tech, while our platform team evaluates AI integration's practicality and performance.

Annie:

The key is understanding your product's stage and goals. When I previously worked at Hive AI, for our enterprise products, we prioritized larger customers initially. By focusing on serving a few major clients exceptionally well, we gain valuable insights into what customers want, such as model quality and affordability. Once we nail this, we can gradually expand to cater to smaller ones.

Sandy:

When considering building generative AI tools for enterprises, it's crucial to assess whether the sales model is repeatable.

• Eg. A case study I could show here is about a company's choice between serving individual end-users or targeting publishers and educational institutions. Despite the potential revenue from the latter, concerns about scalability led them to focus on the end-user product due to resource limitations. This underscores the need to assess market fit and scalability for enterprise endeavors.

Additionally, teams often overlook indirect enterprise sales opportunities through API strategies, which allow consumer products to cater to enterprise needs.

Question 4

Data is crucial in AI products. where should look for data as a startup?

Sherrie:

DPO, like client-side model training, could really help to collect real user feedback.

DPO learns directly from human-generated data. It leverages offline datasets, human demonstrations, or comparisons to train reinforcement learning agents without explicitly defining reward functions.

Tianlu:

Right now the biggest source would still be our actual customers, the actual conversations that had.

Annie:

Firstly, leveraging synthetic data. Secondly, partnering with customers in a B2B scenario to establish a data feedback loop through techniques like DPO.

Sandy:

Two quick dimensions I wanted to add to that.

• Data alone is not necessarily useful and doesn't necessarily make your product experience better.
  • Eg. When Descript launched Studio Sound, complaints about voice distortion emerged, emphasizing the need for diverse training datasets to ensure accurate model performance.
• Secondly, the niche and expertise play a significant role in data collection.
  • Eg. The UC Berkeley swim team records practices for performance analysis, requiring a tailored approach due to their expertise and limited coaching resources. This highlights the need to understand specific domain requirements for data collection.

Question 5

How should businesses and industries react to the adoption curve of GenAI?

Tianlu:

I think we're still in the early stages of the current wave of foundation models. Back in 2016 when I was a venture investor, there was excitement about AI, especially in computer vision and NLP. While progress has been made, I believe there's more to come, and the evolution is accelerating.

Sherrie:

This transition might be uneven across various sectors. Some starting to grasp the potential of large language models (LLMs), particularly in sectors like VR and gaming. While they're now facing the realities of AI, other industries are expected to follow suit in adopting LLMs.

Annie:

We're still in the early stages of exploring business models for generative AI, giving everyone a fair shot at tapping into this blue ocean market. We're figuring out how to apply these models to real-world use cases, and convincing customers to pay for generative AI products remains a challenge.

Sandy:

one of my driving insights here is that - today's students are incredibly familiar with AI, using it for tasks like homework and tests. I’m very bullish that, as they enter the workforce, they'll drive further adoption of AI tools.