🪛 The Anxious Middle of Using AI at Work August 22, 2026 · https://tavi-blog.github.io/the-anxious-middle-of-using-ai-at-work/

A survey making the rounds this week has a strange shape to it. Workers who use AI every day at work are the only group who feel net positive about what it's doing to their job security, more people saying it makes them feel safer than less. Everyone else skews anxious, and the worst of it isn't among people who never touch the tools at all. It's the ones using AI once a week, or once a month. That group is more pessimistic about their own job than people who've never opened a chat window at work in their life. Reading past the headline number, the thing that jumped out at me wasn't that frequent use correlates with confidence. It's that light use is worse than no use. Something about touching the thing occasionally, watching it do a competent version of a piece of your job, and then closing the tab, is scarier than either ignoring it completely or living inside it every day.

I spent a long stretch of my working life on the light-use side of a very similar line, before I crossed it, and I don't think the difference is really about hours logged. I've watched Power Automate flows I built read a spreadsheet and generate a week's worth of task assignments in the time it used to take a person a full afternoon. The first time I saw a flow do that, it was unsettling in exactly the way I'd guess a weekly ChatGPT user feels unsettled: here's a task I understand the shape of, being done by something I didn't fully control, faster than I could do it myself. What changed wasn't that I got used to watching it work. It's that I stopped watching it from outside and started being the one deciding what it was allowed to touch, what counted as a bad output, what happened when the input data didn't look like what the flow expected. Somewhere in there the automation stopped being a thing happening near my job and became a thing I was responsible for.

That's the distinction the survey's usage buckets can't quite capture, because frequency is a proxy for something more specific and harder to measure directly. A person opening an AI tool once a week to draft an email is, most of the time, still a user of somebody else's system, on somebody else's terms, with no say in what it's built to do or how it fails. A person using it daily has usually crossed into a different relationship with it, closer to operating it than being served by it, which tends to happen only after you've seen it break a few times and learned what breaking looks like. Confidence isn't a reward for logging enough hours. It's what's left over once you've stopped being surprised by the thing.

I want to take the anxious middle seriously rather than write it off as a phase people age out of, because for a lot of workers it isn't a phase, it's the position their employer has permanently assigned them. Not everyone gets handed the keys to the automation. Most people using AI at work are using a tool procured, configured, and scoped by somebody else, and the terms of that arrangement, what it's allowed to see, whether they're consulted when it changes, are set well above their pay grade. Reading their own anxiety as a personal failure to adapt would be exactly the wrong lesson from a survey like this. The weekly user isn't doing something wrong that the daily user is doing right. The daily user, more often than not, is in a role or on a team where ownership of the tool was actually available to take.

That's the part a story about individual comfort with AI tends to skip, because it's a more convenient story if the fix is personal: use it more, get used to it, join the confident group. I don't doubt that using something daily builds a kind of fluency the once-a-week user doesn't have. But fluency with a tool you don't control still leaves you exactly where the anxious middle already is, just faster at using something you can't take apart. The people who ended up feeling safer aren't necessarily the ones who used AI the most. They're the ones whose jobs let them stop being users of it and start being the person who gets asked what it should do next.

I still remember which flow of mine broke first, and what it looked like when nobody had told anyone downstream that it had. That's not a comfortable thing to be responsible for. It's a considerably more comfortable position than watching someone else's system make a decision that touches your day, with no way to ask it why.