After running both approaches across plenty of accounts, the framing that sticks is: manual gives you control over price, smart bidding gives you control over outcome, and you can't have both at once. The question is which one you're positioned to specify. If you don't know your real CPA yet, you can't specify an outcome. Manual isn't a downgrade in that situation, it's the honest option.
On those hundred dollar clicks, that's often an unreachable target rather than a misbehaving algorithm. Set a tCPA the account has never hit and Google won't refuse it. It bids up on the small number of queries its model thinks could convert at that number, because those are the only ones that could get there. An unrealistic target produces exactly what you're seeing. Worth checking the tCPA you set against the CPA you were actually achieving on manual before concluding smart bidding is the problem.
One note on the portfolio bid cap suggestion: it works, but it isn't free. Capping max CPC under a tCPA strategy puts you back in the manual tradeoff. You're excluded from auctions the model wanted to enter while it keeps optimizing as though it could enter them. You get price control back and pay for it in model quality. Good as a guardrail against outliers, worse as a permanent setting.
What CPA were you getting on manual, and how does that compare to the target you set?
