DEV Community

Elarigent
Elarigent

Posted on AI-assisted

Don't Automate That Yet: The Jobs That Still Need a Human

It is a Saturday morning, and the phone rings before the shop has properly opened. A regular customer is reading out an email she received overnight. It apologises for a mistake she never reported, offers her a refund she never asked for, and promises a replacement by Monday that nobody can keep. Nobody in the business wrote that email. A system did, working through the inbox on its own, trying to be helpful. The owner spends the rest of the morning unpicking a promise that should never have been made, and the customer spends it wondering whether anyone is actually in charge.

This article was drafted with AI assistance. It is written in plain English for people who run businesses, not for people who build software, and it makes no claim about any particular product. It is about drawing one line clearly: the line between work a system can safely take a first pass at, and work that still needs a person.

The wrong question

The usual question is: can a system do this job? Most routine jobs can be done, after a fashion, by something. Sorting messages, copying numbers from one place to another, drafting a standard reply — all of it can be done. That is not the question that matters.

The question that matters is: what happens when it goes wrong, and who finds out first?

A mistake that stays inside the business is a mistake you can fix. A draft with the wrong number in it, sitting in front of you waiting for a look, costs a minute. A mistake that reaches a real customer before you have seen it costs something else entirely: trust, money, or a promise you now have to keep. The line between those two kinds of mistake is the whole job of deciding what to automate.

That is the relief worth picturing. Not a business with nobody in it, but a Saturday morning where the routine work has been gathered, sorted, and drafted, and nothing has left the building without a human look. The inbox is in order. The drafts are waiting. The owner reads them over a cup of tea, changes one, deletes another, and sends the rest. The customer who needed an apology gets one written by a person who means it. The work still exists. The difference is who does the first pass, and who has the last word.

Three questions before you hand a job over

There are three plain questions that draw the line well enough for most small businesses. Ask them about any job, in this order.

First, can it be undone? Reading a document, sorting a pile, and drafting a reply can all be undone. Sending money, sending a message, changing a record, or making a promise cannot be taken back so easily. If a job cannot be undone, a person should do it, or at the very least a person should check it before it happens. No system should send the un-undoable on its own.

Second, does it need judgement about a person? A customer who is upset is not a category. A supplier who is late for the third time is not just a date on a list. A job application, a complaint, a request for more time to pay — each of these carries context that is not written down anywhere. A system can gather the facts. Deciding what the facts mean, for this person, on this day, is a human job. If the right answer depends on knowing the person, keep the person in it.

Third, would a mistake reach the outside world before you saw it? This is the test that catches most trouble. A system that prepares work for you to check will show you its mistakes. A system that acts first and reports later will show you its mistakes after the customer has already seen them. If the answer is yes, the job is not ready to be handed over. Draft first, act second, every time.

A job that passes all three — easy to undo, no judgement about a person needed, and nothing leaves without a check — is a good first job to hand over. A job that fails any one of them is not a bad job. It is just still yours.

What stays human, in practice

Some examples make the line concrete. Sorting the inbox into piles, pulling the numbers out of an invoice, and drafting replies to the usual questions are good first jobs. They are repetitive, they are checkable in a minute, and a mistake in a draft costs nothing because the draft has not gone anywhere yet.

Answering an angry customer is not a good first job, even though drafting the first version of the reply can be. The draft is fine. The sending is the human part. The same goes for refunds, discounts, and anything that promises a customer something in writing. Let a system prepare the facts and the draft. Let a person decide.

Negotiating is human work. So is hiring, firing, and anything that decides how a person is treated. So is advice where getting it wrong hurts someone — health, money, or legal trouble. A system can collect the information a person needs to give that advice. It should not give the advice as if it were the person.

And a job you have never done yourself is a bad first job to hand over. You cannot check work you do not understand. Do it yourself for a while. Learn what right looks like, and what wrong looks like. Then you will know what to check, and you will spot the mistake in the draft in seconds. That is not wasted time. It is the training that makes the checking quick.

Drawing the line, and keeping it drawn

Drawing the line once is not enough, because jobs drift. A system that drafts replies this month may be offered the chance to send them next month. A tool that reads invoices may ask for the ability to pay them. Each small step feels reasonable on its own. Together, they move the line without anyone deciding to move it.

So keep a short, written list. For each job, write down who does the first pass and who has the last word. Review it when anything changes — a new tool, a new kind of customer message, a bad day that shows up a weakness. The list does not need to be long. It needs to be honest, and it needs to be the list you actually use.

Start with one job, not ten. Pick the job that eats the most evenings and passes the three questions. Run it with a human check for a few weeks. Watch what goes wrong, and fix the checking, not just the job. When checking a draft takes longer than doing the job yourself did, something is wrong with the setup, not with you. When checking takes a minute and catches the odd mistake, the line is in the right place. Only then think about a second job.

The short version

Do not ask whether a system can do a job. Ask what happens when it gets the job wrong, and who finds out first. If a mistake can be undone, needs no judgement about a person, and will not reach a customer before you have seen it, it is a good job to hand over — with a human check. If a mistake would be hard to undo, needs judgement, or would reach the outside world first, keep the job, or keep the last word. The goal is not a business with nobody in it. It is a business where the routine work is gathered and drafted, and the decisions that matter are still made by the person whose name is on the door.

Ready-made AI agents for small businesses, built in public.

Top comments (0)