Starting from scratch can be a huge advantage.
The techier folks definitely have a different set of problems but the speed at which hings get done is night and day. Companies with old school work patterns (which, in my personal experience, means dusty old banks) are terminally entrenched in their ways.
Taking some hopelessly byzantine, spreadsheet-driven process and “automating” it by building a Rube Goldberg scripting framework around it is the kind of totally stupid automation that doesn’t work.
Actually getting down to surface level and understanding fundamentally what each of those humans is accomplishing via those spreadsheets, extracting that all the way back out to a domain model and process flow diagram, and then selectively replacing process steps, whole cloth, with tech designed to be an actual subservice with SLA targets, is the right way to do it.
Throwing the spreadsheets and/or humans out altogether and starting “from scratch” is so exceedingly and needlessly risky from an information loss and hubris point that, well, good luck, but you’re nearly certain to fail.
Yes, a poorly designed process sucks but it works at some level. That means the rough flow of it is figured out. Yes, there are exceptions and complications and all kinds of odd things but it's fundamentally different. It's not "from scratch" as you're taking an existing working-but-broken process where you know the input, know the output, and rethinking everything in between.
In an "inventing" scenario, you have what you think should be the input, a notion of what the output should be, and you're trying to build towards that notion.. without the validation that you're thinking of it correctly.
The first is a harder social problem (aka getting people to change) while the second is a harder technical problem.
But if you are trying to solve a novel problem, and the proposed solution involves "ML will magically predict the future", you'd better have a very good idea of exactly how the problems will be solved, or else you're probably better off starting with good old-fashioned human intelligence.
What often ends up happening is a large manual processes is automated bit by bit, and you end up with the situation you describe: a poorly designed manual process painstakingly replicated in code. Full automation is often never actually achieved here.
The absolute worst thing to do, though, is to begin automating the thing without fully understanding it. It's putting rocket boosters on your self-driving car without first understanding the rules of the road.
