An automation consultant warns that automating workflows without understanding their original purpose can encode outdated rituals and lost context, telling a story of a meaningless 24-hour hold that survived a retired batch system.
In May a distribution company hired us to automate their order approval flow. The brief was something one we hear constantly: "make it do exactly what we do now." 6 steps are there and 5 of them explained themselves. The 6th held every order above a certain size for 24 hours before confirmation and when I asked what the hold was for, the room gave me 3 answers within a minute. Probably it was compliance. Daniel set it up before he left and it's always been like that. None of the 3 was a reason. Quick introduction since an opinion needs a source. I have spent 8 years building software and companies pay us to automate flows exactly like this one which is how I know "make it do what we do now" is the most dangerous sentence in the business. It sounds like the safest possible request and it hides the question of whether anyone still knows where the current process came from. We spent 2 days pulling on the thread before writing any code. The hold turned out to be a workaround for a credit check that used to run in an overnight batch, on a system the company retired 4 years ago. Real time checks replaced it, the batch died and the waiting survived because Daniel left and took the reason with him. A stranger discovery sat underneath that one. The 4 people running the flow each ran it a little differently with their own shortcuts and judgement calls, so automating meant picking one version to become official. In practice, whoever gets interviewed on mapping day decides. What you end up encoding is one employee's memory of the process not the process itself. What worried me most though, was the feedback we were about to switch off. The person doing the hold complained about it roughly weekly and complaints like that are information: a human stuck with an annoying step re-examines it on every single run. She had also developed a feel for orders that looked wrong during that pause and over the years she had caught two fraud attempts almost as a side effect. Our automation would have kept her delay and thrown away her noticing and neither would have appeared on any requirements document. Perfect execution creates its own problem here. A flawless system running a pointless ritual never has a bad day so nobody ever gets a reason to ask what the ritual is for. While humans ran this flow, somebody complained every week and every complaint was a fresh chance to ask why. Once a machine took over that chance would stop arriving. The step would harden into infrastructure and the dashboard would stay green while the world moved on around it. So we refused to automate the flow as it stood and ran a slower exercise first: every step had to be explained aloud in one sentence by a current employee and we recorded the answers before freezing anything. Two steps had no living owner of a reason and got dropped... The hold became a real time check that takes 40 secs. The automation shipped with a page listing the assumptions it depends on and a date next year when somebody has to re-verify them. Ask your team why each step of your oldest one exists and then count the answers that begin with "I think" or "we have always." That count is how much frozen memory you are operating on. Our answers fit in a 9 min recording and I would call that file the most valuable deliverable of the whole project
An automation designer reflects on how a workflow that automates most of their daily tasks has left them wondering what a productive workday should look like when machines handle the repetitive work.
A critique of poorly built automation systems created by so-called experts who ignore error handling, documentation, and governance, leaving clients with fragile workflows that fail in production.
A discussion post questioning whether AI automation actually saves time once you factor in oversight, error-checking, and maintenance. The author asks readers to share experiences where automating a workflow turned out to be less efficient than doing it manually.
The article argues that AI agents are revealing how unstructured and chaotic many corporate workflows actually are, suggesting that successful automation depends more on clean systems and documentation than on advanced models.
The author discusses a problem where scheduled AI workflows run successfully without completing the target task, and proposes a solution focusing on workflow targets with checkpointing and continuous progress until completion.