Here is a story we could not verify. A business owner reportedly posted a six-month post-mortem. The setup it describes sounds sensible. Automated content. An AI bot wired into the CRM. A chatbot for customer service. Dashboards.
Six months later, one thing was still running. ChatGPT, for one specific task, once a week.
The CRM bot got switched off after clients complained. The support chatbot gave wrong answers. Nobody opened the dashboards. The content drafts always needed rewriting. The summary in the post:
“I was so focused on building the AI version of my business that I forgot to just run the business.”
We didn’t collect that thread ourselves, and we could not find the original post. We read it in a round-up on theusefuldaily.com, which is where the quote comes from too. That site describes its articles as produced by an AI-assisted editorial system and doesn’t name the poster. Treat the story as unverified. We use it to illustrate a pattern, not to prove one.
If any of it sounds like your year, start with this. Switching off a tool that cost you more than it saved was the right call. You made it correctly.
It isn’t an AI problem. It’s an arithmetic problem.
Owners who quit these tools are usually not against AI. One headline on the same site put it well: they’re anti-babysitting, not anti-AI.
That distinction matters, because it tells you where the failure actually was.
A tool that saves you two hours a week and costs you two hours of checking has sold you nothing. It just moved the work. You went from doing a job to supervising a job, and supervising is the part nobody wanted.
So the question was never “is AI good.” The question is whether a specific task comes out ahead after you subtract the supervision.
Count the supervision tax before you buy, not after
Here’s the only honest measure of an AI task:
Time the tool saves, minus the time you spend checking it, correcting it, re-prompting it, and cleaning up what it got wrong.
Most people only do the first half of that sum. The second half is where the tools die.
An example. A support chatbot handles 40 customer questions a week, and gets 36 right. That sounds good. It also means four customers got a wrong answer from something with your name on it. Now you’re apologizing to four people, and reading transcripts to find out what else it has been telling people. The tool handled 40 questions and created a new job: watching it.
Compare that to a tool that transcribes your meetings. It saves you twenty minutes. It costs you a two-minute skim. The sum works, which is why that one is still on your computer.
This is also why training matters more than people expect. Capsule CRM’s roundup of small business AI statistics reports that 12% of SMEs invest in AI-related training, while 29% cite lack of training as their biggest obstacle (Capsule CRM). A tool nobody was taught to use properly generates supervision work by default.
The one question worth asking: does this task end?
Look at the setups that last. They tend to have one thing in common. The task has a finish line.
A transcribed meeting is done. A drafted reply is done. A summarized document is done. You look at the result once, and it’s over.
A chatbot answering live customers never ends. It runs all week, generating conversations you didn’t see, some of which you’ll have to fix. A dashboard never ends either. It just sits there producing numbers that ask for your attention.
So before you set anything up, ask one question:
Does this task end, or does it start producing work for me?
| Tasks that end | Tasks that keep going |
|---|---|
| Transcribing and summarizing a meeting | A chatbot talking to customers unsupervised |
| Drafting a reply you read and send | A dashboard you have to check |
| Turning a photo and a few notes into a social post | An automated content pipeline you have to edit |
| Pulling the figures off a supplier quote | Anything that posts, sends, or promises something without you seeing it |
| One automation that fires when a job is marked done | A stack of automations you now have to maintain |
The right column isn’t impossible. It’s just a different kind of commitment, and it needs real setup work and clear rules about what it can and can’t say. Nobody mentions that when they sell it to you as a weekend project.
One thing, properly, before a second thing
A common failure pattern is a stack. Four or five tools, set up in the same month, because that’s what the advice said to do.
The setups that last are often one or two tools doing specific recurring jobs.
That’s not a coincidence. Each tool you add has its own setup, its own quirks, and its own share of the supervision tax. Add five at once and you can’t tell which one is costing you the afternoon. Add one, run it for a month, and you’ll know exactly what it’s worth.
Sequence beats breadth. Every time.
Sometimes the right answer is don’t automate it
This part rarely gets said out loud, so here it is.
Automation needs repetition to pay for itself. If you do something twice a week, the setup time, the maintenance, and the checking will probably cost you more than the task ever did. A note on a whiteboard will beat a workflow.
The advice to automate everything was written for businesses with volume. Below a certain number of repetitions a week, it’s wrong, and following it is how owners end up maintaining machinery for work they could have done by hand.
If you tried to automate something low-volume and it felt like more trouble than it was worth, that’s because it was.
The test: does this task end, and does it still save time after you subtract the checking? If either answer is no, don’t set it up. If something you already set up fails that test, switching it off is a result, not a defeat.
How we pick what to set up
This is most of the job, and it happens before anything gets built.
We start with the tasks that eat your week, and we run each one through the subtraction. What does it save, what will it cost you to supervise, and does it have a finish line. Then we pick one. Not five. We set that one up properly, with the checking built into it rather than left for you, and we get it earning its keep before anyone talks about a second thing.
And when a task comes out negative, we say so. “Don’t automate this” is a real answer, and it’s worth more to you than another tool recommendation.
Where this leaves you
If you tried AI and it turned into another job, you learned something useful. You learned that capability isn’t the same as completion. A tool that can do a task is not the same as a tool that finishes it.
You don’t have to start over, and you don’t have to rebuild a stack. Pick one task with a finish line, run the subtraction honestly, and keep the thing that passes.
That’s the whole method. It’s less exciting than what you were sold the first time. It’s also what’s still running six months later.