Your Company Doesn't Ask a Million Questions. It Asks a Countable One.
On June 3rd, Anthropic published one of the clearest explanations of why naive natural-language-to-SQL fails that anyone has written this year — and then, almost as an aside, told you the price of fixing it. Their analytics agents hit ~95% accuracy. Out of the box, before the scaffolding, that same setup scored 21%. The 74 points in between aren't a model upgrade — they're a permanent engineering organization. It's a genuinely great piece. It's also a blueprint maybe 5% of companies can build and even fewer can keep alive. This is the argument for the other road: the one where you don't generate SQL from scratch and hope it's right, you take a query that's already been tested and let AI reshape it at runtime. Deterministic by default. Accurate because it was validated once, not hopefully, every single time.
