You are thinking of complexity _per_ agent. That is not where the complexity comes from.
If you have to evacuate an area using 4-seater cars - do you really not see the difference in complexity:
A - a village with 200 people, each household has their own car. You need 50 cars, maybe need to requisition/rent a few, the rest people bring by themselves. You don't need a police escort or barriade for orderly movement of 50 cars. People self organize, and you are good to go.
B - a town with 10000 people, 20% households have their own car. Now the administrator needs to arrange for 1000s of vehicles, make sure fuel is available for all of them. You will need to appoint "marshalls" for each block cause you won't be able to handle everything yourself. You also can't expect anyone to self organize at that scale.
Lets say you appointed 50 marshalls. You probably can't even manage all those by yourself. So you appoint 5 "chiefs" - 10 to a marshall.
You need additional staff to manage and track each vehicle. You also need enough police on hand to provide escort. You need to define "assembly points" so that people know where to show up.
Do you seriously not see the complexity that arises when you have 20,000 counters vs a million?
The difference between running a startup with 20 employees - you can do that in your house. No need for HR, building services etc.
Versus a 1000-employee company?
And vote counters at least in India are not volunteers - that would be a disaster. THey are govt employees on deputation. Similarly after polling you don't simply take the ballot boxes to the corner and count. There are a handful of counting arenas - heavy security, representatives of all candidates present etc. Senior "notaries" have to remain present, inspect every box for tempering and certify, with agreement from the candidate reps. WHen you are thinking of a million counters, do you seriously think there can be a say 10,000 notaries present to certify, and a million representatives of candidates keeping their eye on each vote counter? Thats not how the real world scales man.
Stop thinking in terms of 800 votes/counter - theres a MUCH MUCH larger structure there, you keep missing that.
The reason EVMs help is they are smaller and faster. A set of notary, election officer, candidate reps can process a machine in about 10 minutes when counting. An average machine may have 400-500 votes on it. So that means that one set of people can now count about 3000 votes an hour or 25000 votes per day. To count 600 million votes, you need 25,000 sets country wide.
With paper ballots, counting 500 votes would take half a day. So one set would count 1000 votes a day. Now you either need 625,000 sets of counters or more days.
Theres also the issue of securely printing, transporting, storing, issuing 600 million ballots, vs 1.2 million machines.
In anycase - the biggest flaw in your argument is you are arguing theory - 800/counter, so no complexity - and simply ignoring the real world 25 year long experience.
Things simply don't scale like you think they do. 100 page views/ web server doesn't mean you can simply install a million servers and serve out google.com. Saying this added complexity is "not from web serving" makes no sense.