Is the SaaSpocalypse Over?

The episode begins with a light-hearted remark that the SaaS apocalypse has passed and the industry has yet to face its end. Craig engages with Aaron, who built the AI tool DocsBot, about his current feelings on the viability of AI SaaS products. Aaron shares that despite slower growth—indicated by a leveling out of monthly recurring revenue (MRR)—his business remains stable. He notes that while acquisition has declined—seeing a significant drop in new customer sign-ups—his existing customers are paying more, and churn has decreased. This shift suggests a focus on better targeting, product refinement, and higher-value customers.

Aaron discusses how his team has shifted their strategy towards focusing on the right customer segments by narrowing their ideal customer profile (ICP). They have targeted their marketing efforts more precisely, restructured plans, and increased prices on their most popular plans to align with competitors. As a result, revenue per customer has grown by 20%, and churn has decreased. Aaron emphasizes the importance of talking to customers directly through surveys and calls to understand their needs, and applies insights from Jason Cohen's book 'Hidden Multipliers' to refine ICP segmentation. He cautions that very small customer bases—less than 500—struggle to yield reliable conclusions due to low signal density.

Aaron reports that traditional channels like organic search remain significant, although their effectiveness has declined slightly. Referrals and AI-related search contributions are also important. The team has invested heavily in in-person events and conferences such as WordCamp and industry-focused gatherings, building a partner program to enable agencies to resell their product. While this channel is promising, Aaron notes that establishing effective partnerships takes considerable effort, and an exponential growth from this approach has yet to be seen. Despite these challenges, they see events and in-person networking as valuable, especially as a way to sell directly through demos and building relationships.

The discussion shifts to enterprise customer behavior, emphasizing that larger companies tend to move slowly due to internal approval processes. This inertia makes their commitments stickier once adopted, as changing infrastructure entails significant effort. Aaron shares how AI has transformed customer support procurement, with customers initially using AI tools internally but now increasingly expecting AI-driven support solutions integrated into their workflows. He describes their own product’s success, noting a 30% ticket reduction after deploying their AI support agent, and highlights the importance of monitoring and updating documentation based on customer interactions to improve AI performance over time.

The episode explores the financial considerations of running AI SaaS, discussing how model token costs influence pricing strategies. Aaron explains how his team initially avoided paying for inference by letting customers supply their own OpenAI API keys, then transitioned to a credit-based pricing system that factors in token consumption. They balance model choice by considering speed, cost, and performance, avoiding always opting for the newest, most powerful models. Evaluating models internally is complicated by options like user customization and potential regressions, such as changes in classification behavior—like GPT-4 responding more literally. They emphasize the challenge of keeping pace with rapidly evolving AI models while maintaining stable, reliable product performance.

Aaron elaborates on the limitations of deploying large models locally, citing challenges around reliability and cost. Instead, his team uses specialized models for tasks like document extraction, applying smaller or domain-specific models such as Mistral for PDF parsing or Cohere for multilingual embedding. He envisions a future where highly optimized, small models tailored to specific domains could run on consumer hardware, like modest MacBooks, which would be game-changing for privacy and accessibility. He notes that open models are currently more suitable for regulated industries concerned with data security, but infrastructure and cost barriers remain. They also employ privacy measures, such as on-device models that redact PII in real time before data leaves the user’s browser.

Towards the end, Aaron discusses the potential of new AI paradigms, such as re-ranking search results with more cost-effective models. He mentions the recent release of models like Gemini and new approaches like using the Jev classifier for improving retrieval accuracy in complex queries. Aaron sees these innovations as shifts that could redefine how AI is integrated into SaaS products, making solutions more efficient, accurate, and adaptable. Despite his enthusiasm for cutting-edge models, he emphasizes the importance of practical considerations like speed, cost, and stability for real-world applications. The episode closes with Aaron expressing ongoing curiosity and excitement about future AI developments—especially the prospect of highly specialized, small models tailored for specific tasks.