The default playbook sounds right. Give the model a lead record, let it research the company, generate a personalised message, repeat. The assumption underneath is that outbound needs a new argument for every prospect.
That gets strategy backwards. The job of strategy is to reduce the number of decisions your company has to make - not to hand AI hundreds more.
The sharper your ICP, the fewer things there are worth saying. Across a well-defined market you see the same buying situations, the same triggers, the same objections, the same conversations, repeating. That collapse is not a constraint. It is what a resolved ICP looks like from the outside. Those recurring conversations become your canonical messages.
Writing every message from scratch throws that away. You are asking a model to re-derive your positioning once per lead, from a fragment of context, and hoping it arrives at the same place twice.
And something else is lost, quietly. A message that exists once can be sent. It cannot be improved. Canonical messages can be ranked, strengthened and retired on evidence - your messaging compounds because the knowledge survives. Generated messages start from zero every time. That is a system producing activity and destroying the evidence in the same motion.
So we run it the other way. Take fifty to a hundred accounts and let AI do what it is genuinely good at: finding the buying patterns across them, in the context of your business model. Derive the talking points from the pattern, not from each lead. Build the sequence as a narrative. Then place decision points where the model selects which proven argument this account has earned.
AI is exceptionally good at recognising patterns. Strategy is deciding which patterns matter. Those are different jobs.
Most companies use AI to generate more messages. I would rather use it to make fewer, better decisions.