The technical part of moving from a manual calling team to an automated one takes a week. The part that decides whether it works takes longer and is not technical at all.
Start where nobody is defensive
Do not begin with the calls your best people make. Begin with the calls nobody wants: document chasing, appointment confirmation, the second and third follow-up attempts that get skipped when the day gets busy.
Two reasons. Those calls are genuinely well suited to automation, and automating them makes your team's week better rather than threatening it. The first project should visibly remove drudgery.
Take the script from the people who make the calls
The prompt should not be written by whoever owns the tool. Sit with the person who has made a thousand of these calls and ask what they actually say — including the part where the person interrupts with a question, and what they say to that.
You will get behaviour rules you would never have thought of, and the team gets a stake in the outcome rather than something imposed.
Run both for two weeks
Split the list. Half automated, half manual, same week. Compare connect rates, completion, and what the next stage thought of the candidates.
This is worth the cost. It answers the "is it as good?" question with evidence rather than assertion, and if the answer is no in some area, you find out at half scale.
Be explicit about what happens to the people
Everyone doing those calls is wondering. Ambiguity produces quiet non-cooperation that is very hard to diagnose — transcripts stop being read, callbacks stop being made, the pilot "doesn't work".
If the plan is redeploying the team to the conversations that need judgement, say so plainly and early. If it is not, say that too. The worst option is letting people infer it.
Keep humans on the calls that need them
Automate the first contact, not the closing conversation. A candidate deciding between offers, a customer who is upset, a parent choosing a college — these convert on the strength of the person on the line.
The programmes that work best do not have fewer human calls; they have humans making better ones, because the filtering happened before.
Expect the second month to be better than the first
The first month produces a working agent. The second produces a good one, because you have read a few hundred transcripts and know exactly where it is awkward. Budget for that, and do not judge the programme on week two.