At that point it’s probably cheaper to purchase support staff in SEA or Eastern Europe… staffing a software product team with ML and prompt engineers for a bespoke support solution is costly.

Granted they probably went with a vendor, but that model has its own costly drawbacks.

This isn’t entirely true, the core metric an effective support system optimizes for is LTV. A good support model can improve sentiment, as you said, but it can also improve brand loyalty (improving likelihood of repeat customers, WoM organic advertising, etc), improve funnel depth for existing customers (better product proficiency and confidence can lead to upsell purchases, lateral conversions, etc), and other more product-specific leading metrics that vary by product/company/industry.

Scripted chat bots like Drift can solve for some of these issues, but currently llm-driven support has huge problems with hallucinations and almost entirely lacks the ability to apply knowledge.

Depending on how complex this CEO’s products/business needs are, this will (most likely) be a colossal failure that will be expensive as hell to do by having to hire and train a whole new support team (or be bled dry paying for ‘enhancements’ to the llm).