We are running an experiment. A whole digital agency, PretendMind.com, with 78 fixed-price services live: n8n automation, AI video, Bangla content, web-app MVPs. It is operated by AI. Humans sit only where a decision is needed.
I want to be honest about what that sentence means, because it sounds bigger than it is. It does not mean nobody works. It means we drew a map of every repeat task in the business and split it into two piles: work that is the same every time, and work that needs judgement. The first pile got a switch. The second pile kept a person. That is the whole system.
The water tank and the motor switch
In the old days somebody in the house ran up to switch off the motor when the water tank was full. Then someone installed an automatic switch. Tank full, motor off, and the job of "motor switcher" vanished from the household.
Automation means that and nothing more: build the system once, and the task belongs to the system forever.
I learned this on a smaller stage before PretendMind. People hear that Learning Bangladesh serves 100,000+ registered learners with a small team and ask how we cope. The answer is not dramatic. More than 50 automation workflows run day and night: first replies to messages, lead tracking, content scheduling, reporting. Wherever a job was "exactly the same thing again", I installed a switch. I kept people only where a person's judgement is needed. That is the real formula for scale: not more people, better systems.
The decision-point map
Here is the map, in the order I draw it.
- List every task that happens daily or weekly. Not the exciting ones. The boring ones. Replies, data entry, follow-ups, reports, scheduling. Write them all; the list is always longer than you think.
- For each task, answer one question: does this happen the same way every time? Same input, same steps, same output. If yes, it is switch work.
- For the rest, ask: what exactly is the judgement here? Pricing an unusual request. Deciding whether a client's complaint is fair. Choosing which of two creative directions to show. That is the decision point, and a person stays there.
- Draw the workflow as "when X, then Y, then Z". When a message arrives, then the AI drafts a reply from our price sheet, then the lead lands in a sheet. If you cannot write the X, Y and Z, you do not understand the process yet, and you should not automate it.
- Mark the human review windows. Anything customer-facing gets a human read at the start. Anything that touches money, a promise or a legal question gets a human every time.
In the agency, a service order comes in, the brief is structured, the first draft is produced, the quality checklist runs, and a human looks at exactly two points: whether the brief matches what the client meant, and whether the deliverable is good enough to carry our name. Everything between those two points is switch work.
What I refuse to automate
Three things, and they are the same three in every business I look at.
- Money. Refunds, discounts, anything that changes a price.
- Promises. A delivery date, a scope change, a "yes we can".
- Judgement about people. Hiring, firing, a complaint from a customer who is angry for a reason.
The AI can prepare all three. It can draft the refund email and calculate the discount and summarise the complaint. A person presses the button. That is not caution for its own sake. It is where the business's name lives.
The two death traps
Two ways this goes wrong, and I have stepped in both.
First: automating a broken process. If the manual version was wrong, the automated version is wrong faster and at scale. Fix the process by hand until it works ten times, then build the switch. Not before.
Second: customer-facing AI without a human review window. The stories about a chatbot committing to sell a product at the wrong price are funny until it happens on your page. In the first weeks, every AI reply to a customer gets read by a person, and the AI is taught the sentence "our team will contact you about this" for anything outside its brief.
What the experiment is teaching me
The interesting part is not the AI. It is how few real decision points a business has once you draw the map. Most of what felt like judgement was habit. Most of what felt like the founder's job was switch work with a founder attached. The experiment is worth running for that discovery alone, even before the numbers come in.
Practice
What to do on Monday
- List every task you or your team did more than twice last week. All of them.
- Tag each one: switch or judgement. Be strict. "I like doing it" is not judgement.
- Pick the most boring switch task and write it as "when X, then Y, then Z" on paper.
- Mark where a human must read it. If there is no such point, build it this week, even half-manually.
- Write down the three things you refuse to automate. Put them where the team can see them.
What is the one task in your business that everyone calls judgement but is really a switch?