What's actually worth automating with an AI agent

Every other post I scroll past right now is selling the same dream. Point a swarm of AI agents at your business and go to the beach while they run it. Twelve agents that replace a whole team. Automate your startup end to end. I build software on my own, so of course I clicked. Then I went and built exactly one agent, for one job, and left almost everything else alone. That gap between the pitch and what I actually did is the whole of what I want to talk about, because the real question turned out to be a small one: what is actually worth automating with an AI agent in the first place.

The short version, before the detail. One task, run by an agent, that I understood completely before I handed it over. It has saved me a real slice of time every week since. Most of the other jobs I considered never cleared the bar, and saying no to those was the part that actually made it work.

What an AI agent is actually good at

Quick footing first, because agent is a slippery word. A normal script does the exact steps you wrote, in order, then stops. An agent gets a goal, a few tools and some room to decide, and it works out the next step itself based on what it finds. That decision-making is what makes it feel clever, and it is also what makes it a bit of a liability. It will happily improvise, which is wonderful when the task is fuzzy and alarming when the task touches anything you can't take back.

So the sweet spot is narrow. An agent earns its keep on work that repeats, has a clear shape and can survive the odd wrong call without anything breaking. The moment a task needs real judgement, or one mistake is expensive, you are back to doing it yourself or watching very closely. I wrote a whole separate piece on the guardrails I put around mine before I trusted it unattended. This one is the step before that: deciding whether the task belongs to an agent at all.

What makes a task worth automating with an AI agent

Here is the test I actually use now, worked out mostly by getting it wrong. Deciding what to automate with an AI agent comes down to four things for me. Before I hand anything over, the task has to clear all of them, and if it misses even one I leave it as a manual job and stop fighting it.

• I do it over and over. A one-off is almost never worth automating, however satisfying the idea is. The payback only exists if the task comes back every week.
• I can describe it completely. If I can't write down every step and every edge case in plain words, the agent can't either, and it will fill the gaps with confident nonsense.
• The inputs are stable. If the format of what goes in changes every time, I spend more time patching the agent than I ever saved.
• A small mistake is cheap. I only hand over work where a wrong result costs me a few minutes to spot and fix, not a customer or a bill.

The task I landed on ticked all four. It was a recurring studio job I had done by hand dozens of times, so I knew its shape cold, the inputs looked the same every time, and if it got something slightly wrong I would catch it in seconds. That is a boring description of a task, and boring is exactly what you want here.

Do the payback maths before you build anything

The bit the hype skips is that automating a task is itself a task, and usually a bigger one than the thing you're automating. My rough rule is to guess the build time honestly, then double it, because I always underestimate. Then I work out how much time the job actually costs me in a normal week. If the build won't pay itself back within a couple of months of saved time, it goes back on the shelf.

And the maths doesn't stop at launch. An agent is not something you build once and forget. Providers change, inputs drift, a tool it leans on updates and quietly breaks. There is a standing maintenance cost, small but real, and it has to come out of the time you think you're saving. I have killed more than one idea that looked brilliant until I added the upkeep and realised I was automating myself a second job.

Automating a task is itself a task, and usually a bigger one than the thing you are automating.

How I set mine up

When a task does clear the test, the build is less exotic than it sounds. I keep it self-hosted, on infrastructure I control, so the logs and the keys stay mine rather than living on someone else's platform. Tools like n8n exist for exactly this, wiring a model up to a few actions without writing everything from scratch, and you can run them on your own box. The agent gets a model as its brain, a short written brief of exactly what to do, a small set of tools and just enough memory to hold context between steps.

The rule I never skip is to run the whole thing by hand, on real inputs, until it's dull. I watch it decide, check every output against what I would have done, and only widen its latitude once it has stopped surprising me. Testing by hand a dozen times is tedious and it is the cheapest insurance you will ever buy. An agent you haven't watched closely is just a fast way to make the same mistake over and over.

The tasks I left alone

I want to be straight about how much I did not automate, because that list is longer than the one I did. Anything with money past a small threshold. Anything that sends something to a real person in my name. Anything where a confident wrong answer would cost me a relationship rather than an hour. Talking to customers, in particular, I keep by hand on purpose, because the whole point of staying small is that the person emailing you gets a person back.

None of that is technically hard to automate. I just couldn't answer yes to "is a small mistake cheap here", so it stayed manual. The agent drafts and suggests for some of it, and I am the one who presses send. That is the same arrangement I would have with someone I had just hired and didn't fully know yet.

If you take one thing from this, resist the roster of twelve agents and go looking for the single most repetitive, best-understood, lowest-stakes job in your week. Automate just that one, properly, and let it earn your trust before you build a second. One workflow running quietly beats twenty you half set up and can't rely on. Mine has been running for a while now, it saves me an afternoon most weeks, and honestly the best part is how little I think about it. Which, once you strip away the beach and the swarm of agents, is all I actually wanted.

Start smaller than feels satisfying. The narrow, dull automation is the one still running in six months.

Benjamin

Founder & sole developer, Novaire Digital