George Fraser, Tristan Handy, and A16z’s Martin Casado on why companies are building their own AI analysts, why connecting agents straight to SaaS tools falls short, and why to keep your stack open.
Agents being great at the long tail of tedious spelunking is spot on. It's also why plugging one straight into SaaS APIs falls short. Without the modeled, governed layer underneath, it's just guessing faster. Keeping the stack open matters more now.
Tristan's 80% figure assumes a data team already exists. Most nonprofits I work with never built that layer, so the direct-to-MCP shortcut is the only route they can afford, dead end or not.
Ramp hitting 50x the questions in six months is the capacity number I'd plan for first. The semantic layer and row-level checks have to hold when every Slack channel can fire an analyst, not when a handful of humans queue tickets. I'd put those checks in the warehouse path before any agent MCP ever sees a SaaS tool.
the mcp point is the one that deserves more attention. the permissions problem is what kills the direct-wire fantasy in practice. every system of record has its own model of who can see what, and a protocol does not reconcile those models, it just moves the question.
Agents being great at the long tail of tedious spelunking is spot on. It's also why plugging one straight into SaaS APIs falls short. Without the modeled, governed layer underneath, it's just guessing faster. Keeping the stack open matters more now.
Tristan's 80% figure assumes a data team already exists. Most nonprofits I work with never built that layer, so the direct-to-MCP shortcut is the only route they can afford, dead end or not.
Ramp hitting 50x the questions in six months is the capacity number I'd plan for first. The semantic layer and row-level checks have to hold when every Slack channel can fire an analyst, not when a handful of humans queue tickets. I'd put those checks in the warehouse path before any agent MCP ever sees a SaaS tool.
the mcp point is the one that deserves more attention. the permissions problem is what kills the direct-wire fantasy in practice. every system of record has its own model of who can see what, and a protocol does not reconcile those models, it just moves the question.