Automate the Task Nobody Complains About

Somewhere above you a decision was made. The company is paying for AI now, so integrate it, build some automations, do something with it. The budget is approved, the seats go out, and the ask is deliberately open-ended.

That decision is usually right. Being early beats being late, and the person approving it is not going to pretend they know which workflow to point it at.

Then very little happens. On my team the subscription went into writing LinkedIn messages and email copy. That is a reasonable use of it and it is not an automation. We had a system that can operate software and work through a queue for an hour without stopping, and we used it as a better word processor.

Nobody was being lazy. Writing is what it looked like it was for, and nothing about the way it arrived suggested otherwise.

A mandate says where you want to be. A specification says what to build, and it is a separate document that usually nobody has written.

It also cannot be written from the place the mandate is issued, and that is structural rather than careless. The detail that makes a workflow worth automating never travels upward as information. It travels as a slightly annoying afternoon.

Three from my own week. A filter in our CRM that resets every time the page reloads, so a long job cannot be left and come back to. A block of text that pastes correctly on macOS and breaks on Windows, which is what most of the company runs, so every message gets its line breaks put back by hand. A rate limit on outbound email that appears in no documentation anywhere, found out because somebody once sent too many too quickly and got the domain blocked.

None of that will ever be escalated. Nobody files a ticket about a slightly annoying afternoon. So the person holding the budget picks a direction without the one piece of information that would make the choice a good one, and the person holding that information does not know they have it.

The spec is a task, not a technology. Five things make a task the right one to take first.

It repeats. The same shape, many times, on a rhythm you can predict. Repetition is the whole reason automation pays, because a task somebody has already done fifty times has had its edge cases found by a person rather than discovered in production.

It is laborious rather than difficult. That distinction matters more than anything else here. A difficult task needs judgment and will fight you the whole way. A laborious task needs nothing except attention, and attention is the one thing a machine has in unlimited supply.

It requires no taste. If a step means deciding what to say to a customer, or which deal is worth chasing, leave it with the person. If a step means opening a record and setting a field to a value somebody has already decided, take it.

It has an end you can define. You can say what finished means without anybody arguing about it, which is also how you find out whether it worked.

It is boring enough that people have stopped complaining about it. This is the counterintuitive one and it is where the good candidates hide. Loud problems already have owners. The task nobody mentions has been absorbed into what everybody assumes the job simply is, and that silence is the signal rather than the absence of one.

Run that across your own week and the answers come out embarrassingly small. That is the correct result. The size of a task has nothing to do with the size of the return, because the return comes from how often it happens.

The inverse test is worth as much. Avoid the impressive thing. A workflow that runs once a quarter, or where the hard part is a judgment call, or where a mistake cannot be taken back, makes a poor first project however well it demos.

Reassigning CRM records. It repeated constantly, it needed no judgment whatsoever, and one campaign meant opening a couple of hundred records one at a time to change a single field. Nobody had ever complained about it, because it was understood to be the job.

So I built a small agent for it, working through the CRM's API rather than driving the browser, and able to be taught a workflow once and recall it later. It is not a large system, and that is rather the point. Five people ended up using it daily.

Buy the subscriptions. That part was right.

Then go and find the spec, which already exists in your building, unwritten. Ask your team what took longest last week, then ask which part of it needed no judgment at all. Avoid asking what they would like to automate, because that question returns the interesting answer rather than the correct one.

Take one item off that list and ship it all the way to the end, alongside the person who has the problem. One workflow a team uses on Monday morning is worth more than five pilots that demo well.

Then measure the right thing. Seats issued is a procurement number. The question is whether anybody's week changed.

The useful thing about AI right now is not that it can do something remarkable. It is that it will do something tedious, correctly, for the four hundredth time, with nobody sitting in front of it. Find the dull repetitive thing everyone stopped noticing years ago, and take it away from them.