SIGN0015 — Agenticality Signals

74% of enterprise AI agent deployments are getting rolled back. The 26% that survive have one thing in common — and it isn't the model.

We watched this play out inside Agenticality this week. We swapped the generalist model under iFQ on a Tuesday afternoon and nobody noticed. The reconciliation tone stayed the same. The pattern-matching on shadow subscriptions stayed the same. The categorisation calls a household actually trusts stayed the same. The "veteran" — the thing that knows what a Sydney family with two currencies and a mortgage refinance actually means — sits in the loop, not in the model.

Here's the so what for any operator watching the rollback wave: the IP isn't the agent, it's the trace. Every judgement call you let your loop record becomes a hill the next quarter climbs. Skip that work and you are renting a personality from whoever owns the weights this month.

One foundation. Three products. Three loops compounding nightly. Swap the generalist tomorrow — keep the veteran.

This is also what the Parramatta bookkeeper running three small client books on one set of hands can finally afford: a loop that learns her clients, not a model that forgets them at every release.

Tomorrow: the one question every AI-buying CFO should be asking — and almost none of them are.

— Agenticality