Practitioner Lessons

What Ten Days Off the Grid Taught Me About Running Always-On AI Systems

What Ten Days Off the Grid Taught Me About Running Always-On AI Systems

An unplug is a system test, not a holiday. When you remove the operator from a set of automated workflows for ten days, the work that keeps running tells you what you built well, and the work that stops tells you what still needs a human. I switched off every feed, every notification, and every content engine for ten days. Most of it ran without me. The part that did not is the part worth keeping.

I had been treating my time off as a reward. Wrong frame. The mountains were the reward. The disconnection was the experiment. I told myself I would not open a single social app, would not check a single dashboard, and would not approve a single draft. No exceptions, no quick peeks. Ten days. Mountains first, then a safari. Two sons, my wife, clean food, the daily walk I had let slide for months, and air that did not smell like a server room. I went in expecting anxiety. The thing nobody warns you about is that the anxiety never came.

Here is what actually happened while I was gone.

My content pipelines kept producing. Drafts got written, queued, and held for review on schedule. The morning routines that aggregate my reading and surface the day's signals ran every day I was away. The blog you are reading runs on a cadence that does not need me sitting at a keyboard on a Sunday night. When I came back and opened the queues, the work was there. Some of it was good. Some of it needed a second pass. None of it had collapsed because I stopped watching.

That sounds like a small thing. It is not. For two years I had carried a quiet belief that the machine needed me hovering over it. That if I looked away for a week, the whole thing would drift, the quality would slide, and I would return to a mess. I had built always-on systems and then refused to let them be on without me. The unplug killed that belief in about 72 hours. By the third morning, when I deliberately did not open the laptop and nothing caught fire, I understood that I had been the bottleneck I kept complaining about.

I want to be precise about the texture of those three days, because the shift was not philosophical. It was physical. The first morning, my hand reached for the phone before my eyes were fully open. That reflex is two years deep. I had to physically put the phone in a drawer in another room to break it. The second morning the reflex was weaker but the unease was louder. A low hum of "something is happening and you are missing it" sat behind everything, even on a trail with my younger son pointing at deer tracks. By the third morning the hum was gone. Not managed, gone. And in its place was a clarity I had not felt in a long time. I started noticing that the worry had never been about the systems at all. The systems were fine. The worry was a habit wearing the costume of responsibility.

But not everything ran.

The conversations did not run. The relationship I have with a Chief Information Officer who texts me his hardest questions did not run. The decision I owed a partner on a new project did not get made. The reflection that turns a week of client calls into a point of view nobody else has did not happen, because nobody can do that for me. The judgment work waited. The meaning work waited. And when I returned, those were the things sitting at the top of the pile, untouched, because they were never automatable in the first place.

One example made the distinction concrete. While I was away, a Chief Operating Officer at a large consumer brand sent me a question about whether to consolidate three overlapping AI tools his teams had each adopted on their own. My systems could have drafted him a comparison of the three. They did not, because I had not asked them to, and that is the point. The comparison was never the hard part. The hard part was the read on his organisation, the politics of taking a tool away from a team that loves it, and the judgment call on which fight was worth having this quarter. That answer sat in my pocket for ten days and lost nothing by waiting. If I had tried to automate it, I would have shipped him a tidy table that answered the easy question and ignored the real one. Some work is supposed to wait for the person who can see the whole board.

That contrast is the whole lesson, and it gave me a framework I have not been able to stop thinking about since.

There are three kinds of work in any operation that runs on automation. Most people blur them together and then feel guilty about all three when they step away. Separating them is what lets you leave without the place falling apart, and it is what tells you where your actual value sits.

The first kind is maintenance work. This is the recurring, rules-based output that keeps the lights on. Drafting, scheduling, summarising, formatting, monitoring. It has a clear input and a clear output and a repeatable shape. This is the work AI handles best, and it is the work most builders are most reluctant to release, because doing it feels like progress. It is not progress. It is motion. When I unplugged, every piece of maintenance work ran exactly as designed. If your maintenance work stops the moment you leave, you have not automated it. You have just scheduled yourself.

The second kind is judgment work. This is the call only a person with context can make. Should we take this project. Is this draft saying the true thing or the convenient thing. Does this client need a framework or a frank conversation. AI can prepare the inputs for judgment work. It can lay out the options, surface the risks, and draft the first version. It cannot own the call, because owning the call means owning the consequence. Judgment work waited for me, correctly. The mistake is pretending it can be queued like maintenance work. It cannot. It pools.

The third kind is meaning work. This is the rarest and the most human. It is the reflection that converts experience into a point of view. It is the original thought that did not exist before you sat with the discomfort of not having an answer. On the mountain, with no feed to fill the silence, this kind of work actually got better, not worse. The ideas I am writing down now arrived because I had ten days of nothing pulling at my attention. Meaning work does not just survive the unplug. It depends on it.

A diagram makes the separation clear.

              THE WORK YOU LEAVE BEHIND

  Maintenance work   ->  runs without you      (automate, release)
        |
  Judgment work      ->  waits for you         (cannot queue, must return)
        |
  Meaning work       ->  needs you absent      (silence is the input)

  The unplug sorts the three. What ran was maintenance.
  What waited was judgment. What got better was meaning.

Read the three together and the design goal changes. For two years I had measured my systems by output volume. More drafts, more posts, more cadence. The unplug reframed the target. The real measure of a well-built operation is not how much it produces while you watch. It is how cleanly it sorts the three kinds of work so that the maintenance runs without you, the judgment waits patiently for you, and the meaning has room to arrive. A system that needs you for all three is not a system. It is a job with extra steps.

There is a personal edge to this that I did not expect to write down. The reason I had never fully unplugged was not that the systems needed me. It was that I needed them. The always-on feed was doing something for me that had nothing to do with the work. It was filling a silence I had grown uncomfortable sitting in. Ten days in the mountains, with my sons and my wife and a walk every morning, gave the silence back. And the silence turned out to be where the best thinking lives. The health I had let slide, the family time I kept postponing, the clean eating, the question of what actually matters most. None of that needed a dashboard. All of it needed me to put the dashboard down.

So if you build and run automated systems, and you have been afraid to look away, do three things this week.

  1. Pick one full day this week and remove yourself from every recurring system you operate. No feeds, no dashboards, no approvals. Write down on paper which outputs still appeared the next morning and which did not. The list of what ran without you is your real maintenance layer. Trust it more than you do now.
  2. Take the work that waited and sort it into judgment or meaning. For each item, ask whether it stalled because it needed your decision or because it needed your reflection. Judgment work gets a calendar slot when you return. Meaning work gets protected silence, not another meeting.
  3. Schedule one ninety-minute block this week with no inputs at all. No phone, no notes, no music. Sit with one open question about your work and let the silence do its job. Most people have not given an idea room to arrive in years. You will be surprised what shows up when nothing else is competing for the space.

These three actions cost you nothing but the discomfort of stepping back. That discomfort is the point. The people who build systems they can leave are the people who learn what their systems are actually for.

One last thought before you close this tab. The promise of automation was never more output. More output is the consolation prize the industry sells while it figures out the real one. The real promise is absence. The freedom to be on the mountain with your kids while the maintenance runs itself, the judgment waits without rotting, and the meaning has the silence it needs to show up. I built always-on systems for two years and only understood what they were for when I finally turned myself off. The machine was never the point. The machine was the thing that bought me the ten days. Spend yours on the work no machine will ever cover, and the rest will keep running while you do.

While you are here, the back catalogue has more on this.

FAQ

Common Questions

What are the three kinds of work in an automated operation?

The three kinds of work are maintenance work, judgment work, and meaning work. Maintenance work is recurring rules-based output like drafting, scheduling, and monitoring, and it runs without you once automated. Judgment work is the decision only a person with context can own, because owning the call means owning the consequence. Meaning work is the reflection that turns experience into an original point of view. The test of a well-built operation is whether it sorts the three cleanly, so maintenance runs in your absence, judgment waits for your return, and meaning has the silence it needs to arrive.

How does maintenance work differ from judgment work in an AI workflow?

Maintenance work has a clear input, a repeatable shape, and a defined output, which is why AI handles it and why it keeps running when you step away. Judgment work has no fixed shape because it depends on context and consequence, so AI can prepare the inputs but cannot own the call. The practical difference shows up the moment you take time off. Maintenance work appears on schedule without you. Judgment work pools and waits, because queuing a decision is not the same as making one.

Why does stepping away from your own systems feel risky even when they run fine?

It feels risky because most operators have never tested the assumption that their systems need them. The belief that quality will slide without constant supervision is rarely checked, so it hardens into a habit of hovering. There is also a personal pull. An always-on feed often fills a silence the operator has grown uncomfortable sitting in, which makes stepping away feel like loss rather than relief. The fear usually breaks within the first few days of an actual unplug, once nothing collapses and the silence turns useful.

When should you deliberately unplug from the systems you operate?

Unplug deliberately when you suspect you have become the bottleneck in your own operation, when your time off keeps getting interrupted by quick checks, or when your best thinking has gone quiet under a constant stream of inputs. A full unplug works as a diagnostic. It sorts the work that runs without you from the work that waits for you and the work that needs your absence. The wrong moment is never. An operation you cannot leave for ten days is not a system you own. It is a job that owns you.

What is the first step to building systems you can step away from?

The first step is to remove yourself from every recurring system for one full day and write down which outputs still appeared the next morning. That list is your real maintenance layer, the work that genuinely runs without you. Most operators trust it far less than they should. Once you can see what runs in your absence, you can stop supervising it, redirect that attention to the judgment work that waits and the meaning work that needs silence, and design the next iteration of the system around being absent rather than present.