Business process automation is one of the best ideas business software ever had, and one of its most reliable disappointments. The concept has been right since before we called it BPA. The implementations kept killing it.
I have some history here. In 1981 I was a law clerk teaching a Pitney Bowes word processor to produce the repetitive half of my discovery work, and I have been automating knowledge work ever since. So I have watched every wave of process improvement arrive with the same correct diagnosis and leave much of the same wreckage. Reengineering in the 1990s: Michael Hammer's 1990 article told everyone "don't automate, obliterate," and a generation of consulting engagements obliterated a great deal of load-bearing work. BPM suites in the 2000s: process owners got modeling tools that produced beautiful diagrams, and binders that nobody opened once the engagement ended. RPA in the 2010s: screen-scraping bots that broke every time a vendor moved a button, until the bot farm needed its own maintenance department.
The diagnosis was never the problem. Most knowledge work really is stuffed with waiting, rework, redundant review, and copy-paste between tools. The lean movement named those waste categories decades ago and was right about all of them.
Where every wave actually died
Look closely at the failures and the anatomy repeats.
Mapping cost too much. Understanding a process meant workshops, interviews, and months of consultant time, so the map itself became the deliverable. By the time it was accurate, it was obsolete, and the budget was gone before anything changed.
The as-is got automated. When mapping is expensive, redesign gets skipped, and teams automate the process they have. That is paving the cow path. Automate a broken process, and you get the same chaos, faster.
The plumbing was brittle. Rules engines and screen-scrapers could only handle perfectly deterministic steps. Anything requiring judgment stayed manual, and everything automated shattered on the first UI change.
Nobody managed what got built. Automations were projects, not operations. They shipped, the team moved on, and they rotted silently until someone noticed the queue had been jammed for three weeks.
What AI actually changes
The tempting claim is "now the AI can run your process." That repeats the original mistake with better technology. Intelligence pointed at a broken process produces the broken process at machine speed.
What AI actually changes is the cost of the discipline that always worked.
Mapping is cheap now. A managed agent that already works inside your tools can draft a value-stream map in hours instead of months: read the artifacts, trace the handoffs, tally where humans touch the process, and propose the map to the process owner for correction. Discovery stops being the budget.
Stripping still comes before automating. This is the lean part, and it is where most of the win lives. Delete steps. Merge steps. Move approval to the last gate. Cut the human touches down to the irreducible set, the judgment calls only a human should make. The best automation is still a deleted step: zero cost, zero maintenance, zero failure surface.
Each surviving step gets the cheapest rail that fits it. A mail filter where a filter will do. A script where the logic is deterministic. An agent only where judgment is required, and an agent with a named human behind it wherever the act is irreversible: publishing, sending, paying. We run this as a standing rule, and it generalizes: never spend a token where a filter would do, and never let software own an irreversible act.
What gets built gets managed. Every scheduled automation we arm carries a verification step and a failure alarm, because a green status light tells you the job started, not that it worked. This is the piece every prior wave skipped, and it is the piece we think matters most. Supervision is not overhead on the automation; it is the product.
We are the first test case
We are running this method on ourselves before we sell it to anyone. The pilot is our own content pipeline, idea through publication through syndication, mapped step by step and routed rail by rail. The artifacts it produces are the proof I intend to show: the value-stream map, the routing plan with its gates, and the automations running behind them. Our managed agents already run parts of their own pipelines this way, from scheduled draft batches to headless publishing, each with a human gate at the step that cannot be undone.
I am not going to quote you a savings percentage. We have not measured one, and unmeasured numbers are how this industry earned its reputation. When the pilot produces artifacts worth showing, they will be shown, and you can judge the method by them.
Current view, subject to change
BPA kept failing on two things: implementation cost and absent management, and never on the concept. AI collapses the first. Managed agents supply the second. That is the bet.
What would change my mind: if our own value-stream maps go stale the way the BPM binders did, or if the maintenance burden that killed RPA reappears wearing an AI badge. Six months from now, if the pilot's map is a museum piece, this post was wrong, and I will say so here.
Regards,
Charles Stack