Introduction
You bought automation to get time back. A few weeks in you are checking its work, fixing what it got wrong, apologising for the one that slipped through, and rebuilding it every time something changes upstream.
That pattern is real and worth naming honestly. It is not evidence that AI does not work. It happens when the work that comes with a tool is never counted as work.
What is the second job made of?
Five parts. None of them appear on an invoice.
- Preparation. Feeding it. Writing the request properly, pasting in the context, finding the file it needs before it can do anything.
- Checking. Reading everything it produced, because you do not yet trust it to be right without you.
- Correcting. Fixing the ones that were wrong, and repairing whatever the wrong one already touched on its way out.
- Escalating. Catching the cases it should never have taken, then handling them yourself, usually late and usually while apologising.
- Maintaining. Keeping it alive. A price changed, a form changed, a person left, a connection expired, a template moved.
The task got faster. The job around the task got bigger. Nobody measured the second one, so the business felt busier without anybody being able to explain why.
Why does none of this show up in the demo?
Because the demo is one task, once, watched by everybody.
The second job is created by volume and by time. Checking one draft is nothing. Checking a queue of them every week is a role. Fixing something the day it launches is easy. Fixing it much later, when the person who set it up has moved on and nobody wrote down why the rule exists, is a different exercise.
Point tools are sold on the single task. The work they create sits outside the frame.
How do you size the second job honestly?
Measure it for one week, with marks on a page rather than a stopwatch.
Take the five parts above and put a mark next to one every time you do it. Preparing an input. Checking an output. Correcting something. Rescuing a case it should not have touched. Fixing the thing itself.
By Friday you will know which of the five is eating the time, whether the corrections are all the same correction repeating, and whether the person doing the checking is the person whose time you were trying to protect.
Do the same count for the work as it was before the tool arrived, using the same definition of the task. That comparison is the honest way to tell whether anything improved, and it is worth more than any calculator that multiplies a guess by an hourly rate.
Why does adding more tools make it worse?
Every tool you add is another thing to feed, check and keep alive, so the second job grows with the number of them.
It also creates the problem owners describe most often. The work moved but the handoffs did not. One tool answers the enquiry and drops it somewhere nobody watches. Another drafts the quote without knowing the job was already booked. Somebody copies between them, which is the manual work you were removing.
Five tools that each do one task well can leave a business worse off than one connected path, because the joins are where the checking lives.
What actually removes the second job?
Two things, and neither is a better tool.
The first is boundaries. Checking exists because the system is allowed to do things it should not be trusted to do. Narrow what it can decide and the checking narrows with it. A draft you approve needs a glance. An unbounded system needs supervision, which is a job.
The second is visibility. Most of the checking people do is looking for problems one record at a time, because they have no other way to know. When every item carries a state, an owner and a next action, and the exceptions gather in one place, you stop checking everything to find the few that need you.
Judge a build by how little of your attention it needs in a normal week.
When should an automation be retired?
Sometimes the honest answer is that a workflow is not worth keeping, and that is a decision worth making on purpose rather than by neglect.
Retire it when the checking costs more than the task ever did, when the same correction keeps coming back after the rule has already been corrected, when nobody in the business owns it, when it was built for a way of working you have since changed, or when it only ever handled the easy cases and a person handles everything else anyway.
Turning something off is a legitimate result. It is also the part of a review most providers never raise, because it reduces what they can bill.
FAQs
Is the extra work just the setup period?
Some of it is, and it should visibly fall over the first few weeks as the rules get corrected. What should not happen is checking staying flat month after month. If it does, the boundary is wrong or the rules never got fixed, and more time will not solve either.
Who should be doing the checking?
Not the owner, past the first few weeks. Approvals that need judgement about price, scope or a customer relationship stay with the person who owns that decision. Everything else should be visible enough that nobody has to go looking. If the owner is still the safety net after a couple of months, that is the finding.
What do we do if this has already happened to us?
Start with the count, then the boundaries. Most businesses in this position do not need another tool. They need to know which of the five parts is taking the time, and to decide what the system is allowed to do without asking. Bring a workflow you are currently babysitting to an Operations Review and we will go through it with you.
SOURCES
CONTENTS