AI is a fine research assistant. But it has never watched a formulation break on a hot-fill line, never negotiated a co-packer minimum, and doesn't know which supplier's “in stock” quietly means eight weeks. What we sell is the part that can't be scraped: judgment, relationships, and pattern recognition from decades inside the industry. AI can tell you what a co-packer is. We can tell you which three will actually perform.
To be clear, we use it daily — for research coverage, for first drafts, for speed. Used well, it lets a two-principal firm move like a ten-person one, and we'd rather our clients get that leverage than pay for it. So this isn't a lecture about the machines. It's a note about where the machines stop.
AI is trained on what's been published. This industry runs on what hasn't. The co-packer's real minimum — the one that moves when you know how to ask — versus the published one, which doesn't. The supplier whose paperwork is immaculate and whose lead times aren't. The ingredient broker who returns calls during a shortage, and the one who goes quiet. None of that is on the internet. Which means none of it is in the model.
There's a subtler problem: AI gives you the average answer. What usually happens, across every category, at every scale, to everyone. But product lines and deals don't fail on averages — they fail on specifics. This formulation, on that line, at this volume, under that contract. The average answer is where our work begins, not where it ends.
And the last thing: accountability. When a recommendation is wrong, AI doesn't take the call. We do. When the investment committee asks who stands behind the diligence, “a chatbot” is not an answer. A signed memo is — and accountability is the one deliverable no model will ever ship.