The whole point of automation is to save on labor, at the cost of simplicity. Machines are excellent at handling complexity that some experts can come up with and can execute it reasonably well. Except that this is precisely the opposite of what's desired in an election.
Voting is one situation where it is DESIREABLE to have many people manually involved in a simple process. Auditing voting machines requires a pretty in-depth knowledge of information systems and experience with a lot of edge cases.
Proprietary voting machines only make auditing worse, basically hiding all flaws behind a curtain of marketing and legal threats. Over 95% of the citizens have no way of auditing the integrity of the results and have to put blind faith in for-profit corporations who are (1) disincentivized from disclosing flaws and (2) do not suffer penalties when security flaws are found to be present by researchers.
In a paper scheme, it takes more human effort to tally an election, but every citizen can participate in that process, and the others can safely observe and easily understand what's going on. The only downside is that it may take a few hours instead of seconds. But latency is infinitely preferable over opacity and potential inaccuracy.
I used to think e-voting would be OK if there was a requirement to use open-source voting software but even that is problematic. As a software engineer, it would be my pride to be able to help, however that would still be depriving non-software people from the ability to participate in the process, and I feel that is wrong. Now that my focus at work is on building resilient systems, I believe paper is the way to go, especially considering one of the other requirements is secrecy.