The first ninety days of a company brain
A practical account of what deployment actually looks like: which systems to connect first, what to gate, and where the resistance comes from.
The failure mode of an ambitious platform deployment is trying to be everywhere at once. What follows is the sequence that has worked, roughly, across the deployments we have run.
Weeks one to three: perceive only
Connect read-only. Finance, the CRM, the ticketing system, and wherever documentation actually lives. Cortexia writes nothing in this period. The output is a picture: here is what your company is currently doing, drawn from your own systems, with the inconsistencies visible.
That picture is usually the first genuinely uncomfortable artefact of the project. Every organisation has processes that exist on paper and not in practice, and this is where they surface. Budget time for the conversations it starts.
Weeks four to seven: propose, do not act
Pick one process. Something high-volume, well-bounded, and irritating. Supplier onboarding and expense exceptions are common first choices. Cortexia proposes the decision, a human makes it, and the disagreement rate gets logged.
Two things come out of this. You learn where the system's judgment is weak, and more often you learn where your policy was never actually written down. Most of the work in this phase is specification, not tuning.
Weeks eight to twelve: release a class
Take the narrowest decision class where the disagreement rate is near zero and let it run autonomously, with gates still in place on everything adjacent. Watch it. Then take the next one.
Resistance, when it comes, is rarely about the technology. It is about accountability: who is responsible when an autonomous decision is wrong. The answer has to be a named human, the same as before, and the audit trail has to make their oversight genuinely practical rather than nominal. Teams that get this settled early move quickly. Teams that leave it vague stall at exactly this point.
What ninety days does not get you
It does not get you a company that runs itself, and anyone promising that is selling something. It gets you one or two processes genuinely running end to end, a memory graph with real precedent in it, and most valuably an organisation that has learned how to specify its own decisions. The second quarter is much faster than the first.