The right question isn't how many errors, it's which ones
If you're weighing up an automation and you want to know how many errors it makes, the honest answer is that the number on its own is no use to you, because a person and a machine don't get things wrong in the same way, and until you understand how they differ you don't even know what you're comparing
A person gets things wrong at random, at six on a Friday more than at ten on a Tuesday, they get it wrong when the phone rings halfway through recording something, a figure on this invoice and a date on the next one, and every error is different from the one before
An automation never gets distracted, never has a bad day, never picks up from the wrong point after an interruption, and on a thousand identical cases it does exactly the same thing a thousand times
Which sounds good as long as that same thing is right, because that's exactly where the other side of it sits
When a machine gets something wrong, it gets it wrong across the board
If a rule inside the automation is written badly, or was understood back to front, that error doesn't turn up once in a while, it turns up on every case that goes through that rule, and it makes them one after another without ever stopping to think something might not add up
A person who copies an amount into the wrong field notices by the third time because it looks odd to them, the machine doesn't, the machine does it the thousandth time as calmly as the first
Say, and these are numbers made up here just to show the reasoning, that you have five hundred documents a month and one wrong rule about how the date is read, a person would have ruined a few of them, the system ruins all five hundred
That's why anyone who promises you their system never gets anything wrong is selling you a phrase, and anyone who only tells you it gets things wrong less often is telling you nothing useful, what counts is what you have put around it
This is where the checks come from, not from an excess of caution
The nets around it are born here, and this is why an automation built well would rather stop than guess
The checks are there to catch the rule error before it becomes a multiplied error, the totals have to add up, the data has to make sense, and when something doesn't add up the job stops and lands in a list a person looks at, instead of merrily carrying on
You also need a log of what the system did and when, because without it you find out about an error made across the board from an angry customer and not from your own screen
And you need a gradual rollout, with the system working alongside the person on the same cases for a while and the results compared, because that's where the rules that were misunderstood come out, while they still cost little
If you want the detail on those nets, it's all in the guide on how you stop an automation doing damage, and on how the rollout is handled there's the guide on the first day the automation really works
If a supplier talks to you about accuracy percentages and doesn't talk about what happens when a case comes in crooked, they're telling you the comfortable half