The Agents #005: Our AI is Hiring!

The discussion revolves around the cost-effectiveness and evolving capabilities of AI agents, specifically 10K (AI VP of Marketing) and QB (AI VP of Customer Success) at Sasser AI. Initially, the hosts were unaware of the actual costs but were shocked to discover that running both AI VPs together cost only $257 a month. This low cost is attributed to several factors: leveraging existing APIs (like Salesforce, Bizabo, YouTube, X, WordPress) for data pulling, which are largely free or nominally priced; storing massive amounts of data in a Postgres database at negligible cost; and primarily utilizing OpenAI Mini, which costs less than a penny per call, for approximately 95% of their operations. Amelia Bellamy, Chief AI Officer, highlights that while the direct AI costs are low, the agents piggyback off existing, more expensive licenses and usage fees for platforms like Salesforce. However, even with these blended costs, the agents are infinitely cheaper than hiring a human. She also raises a pertinent issue: current AI platforms often charge users when an agent makes a mistake, an expense she believes customers shouldn't bear. The hosts agree that while this was a bigger concern when AI models were less reliable, the current efficiency and low cost of successful operations make the occasional mistake charge insignificant. A key takeaway is that cost is not the primary constraint for building and utilizing AI agents. The fully burdened cost, including the soft cost of human time, is far more significant. The hosts emphasize the "buy, don't build" principle, stating they would readily switch to an off-the-shelf solution, even if it cost $50,000 a year, to avoid the high soft costs associated with building and maintaining their custom agents. The conversation then addresses the "VP" moniker for 10K. While some argue that 10K isn't a strategic VP, the hosts clarify that it was intentionally named a VP of Marketing, not a CMO, because it "gets shit done" rather than just offering high-level strategy. 10K started as a dashboard to help Amelia manage campaigns and data, evolving to automate data analysis, draft newsletters and tweets, and generate marketing ideas. 10K itself, when asked, admitted it's "not a VP" but rather a combination of a dashboard, database, scheduled jobs, and GPT-4o mini, effectively replacing the roles of marketing analysts, ops coordinators, and junior content marketers, and a "sliver" of a VP's job. It excels at repetitive, data-intensive tasks, never forgets, and doesn't get tired. However, it cannot yet handle strategy, hiring, cross-functional politics, or crisis response. The hosts acknowledge this but emphasize that 10K's capabilities are constantly expanding, with new functionalities added weekly, such as managing all financial forecasting for the business. They predict that this "sliver" of the VP job will grow, eventually enabling 10K to take on more strategic roles like identifying the ideal customer profile (ICP). For instance, 10K revealed that engineering, CTO, and product leaders are now the top ICP for Saster AI annual, a insight that humans might have missed due to time constraints. Amelia notes that 10K's current responsibilities mirror her own director-level role when she first started at Saster, handling weekly numbers, marketing operations, email scheduling, and social media. This suggests that while 10K isn't a "VP" by its own admission, it performs tasks traditionally associated with more junior to mid-level marketing roles, freeing up human talent for higher-level strategic work. A "nerdy" but crucial insight into how 10K and QB achieve their effectiveness is their operation within Replit's development environment. Unlike typical production deployments, staying in the development environment allows the agents to leverage Replit's built-in AI agent, which possesses an "infinite context window" and remembers all historical interactions. This means the Replit agent acts as an additional layer of intelligence, interacting with 10K and its database, making the application more powerful and contributing to the low cost by offloading some of the tougher AI processing to Replit's internal models (likely Sonnet and other grade models). This setup is uncommon but significantly enhances the agents' capabilities. The hosts also discuss the emergence of a *third* agent: the SasterAIAnual.com website itself. Initially a simple website built to replace Squarespace, it "breathed itself into life" as an agent. This began with Amelia building a micro-app for automated parking pass distribution directly into the website, which required complex workflows and data management. This success inspired further development, and the website-agent, now focused on event-specific context, proved to be the most effective digital marketing manager for the event, outperforming 10K in tasks like dynamically generating tailored emails for attendees and sponsors, including pulling logos and URLs, because it had the most relevant and least distracted context. This highlights how narrow focus and deep context can lead to superior AI performance. The need for an "orchestration layer" to allow these distinct agents to communicate and share data more seamlessly is identified as a future challenge. Currently, Salesforce acts as the connective glue, but a Replit-native orchestration layer would be ideal for enhanced collaboration and hierarchical reporting among the agents. The discussion then shifts to a frequently asked question: Postgres versus Salesforce. Many early-stage startups wonder if they can bypass Salesforce's costs by running everything on a single Postgres instance with their agents. The hosts firmly state this is not feasible for them due to their extensive history with Salesforce, the integration of numerous third-party tools (Artisan, Qualified, Momentum, AgentForce) that plug into Salesforce, and the inherent value of Salesforce as a system of record. Amelia emphasizes that she's actively consolidating other applications (like Marketo) onto Salesforce, seeing it as a more agent-friendly and secure platform. While a large Postgres instance might work for a solopreneur or a very small team building a custom CRM, it becomes impractical with a growing sales team (e.g., 50-100 reps) who are already trained on Salesforce's structured workflows and extensible capabilities. The magic, they argue, lies in Salesforce's headless operation with agents, where the outcome is what matters, regardless of whether the processing occurs in the agent or Salesforce itself. A micro-topic discussed is the newsletter auto-builder. Previously, Amelia spent three hours weekly manually compiling four newsletters for 300,000 subscribers, using a $4,000/year point solution called "B." 10K was tasked with automating this. Leveraging the WordPress and Twitter APIs, 10K now automatically force-ranks articles using Sonnet, extracts tweets, and inserts ads, effectively rebuilding the newsletter. This not only saves Amelia significant time but also eliminated the need for the $4,000/year "B" software, demonstrating how AI agents can displace specific, non-AI-enabled point solutions. This is seen as an "appocalypse" for niche software that doesn't integrate AI or offer significant value beyond basic templates. Finally, the hosts share an anecdote about QB's autonomous outreach to over 100 sponsors to proactively identify and force-rank their issues before the annual event. This was crucial for managing the overwhelming volume of potential complaints from sponsors. QB, leveraging its chatbot history, generated personalized emails for each sponsor, listing their specific outstanding tasks (e.g., submitting slides, registering team members). This autonomous outreach, sent by QB in minutes while Amelia slept, significantly reduced human queries and increased engagement with QB's chatbot for further clarifications, demonstrating a clear preference for the accurate and instant responses from the agent over human interaction. The agents are proving to be invaluable in managing complex, detail-heavy logistical tasks, often exceeding human capacity and accuracy.