Here's a number worth sitting with.
A new study found that 73% of small businesses that integrated AI tools in 2024 and 2025 are now scaling back their usage. Not abandoning AI. Scaling back. Pulling the stack down. Dropping subscriptions. Returning to some things they used to do by hand.
That's not a fringe result. That's almost three out of four businesses that were early enough to actually go all-in.
And when you ask them why, the answer is not what you'd expect. It's not "the AI didn't work." It's not "it was too expensive." The most consistent answer is some version of: "I feel more exhausted than I did before."
That answer points to something that almost nobody talks about when they sell you AI tools. Something that got added to your plate the moment you subscribed to the first one.
The Job Nobody Put on the Job Description
When AI adoption is discussed, the language is all subtraction. AI will take tasks off your list. Automate the repetitive stuff. Free up your attention for higher-value work. You'll get hours back. You'll need fewer people. Your overhead will drop.
What nobody mentioned: you would also get a new job.
The job is called AI supervision. It's not listed anywhere as a line item. It doesn't have a title. But it's real, and if you've been using AI tools seriously for any length of time, you've been doing it — probably without calling it that.
AI supervision is everything required to keep AI doing what you need it to do:
- Writing, testing, and refining prompts — because AI outputs at a quality level proportional to how precisely you told it what you wanted
- Reviewing outputs before they go anywhere — because AI is confidently wrong at a rate you cannot fully predict
- Catching hallucinations and errors before they reach customers, spreadsheets, emails, or financial records
- Rebuilding automations when integrations break — and they break, and they don't tell you
- Context-switching between five different tools that don't share data and require five different mental models to operate
- Monitoring what's running in the background to verify it's still running
That's not the list of a tool you use. That's the job description of someone whose job is managing a junior employee who works fast, makes plausible-sounding mistakes, and needs constant direction.
The "junior employee" framing is worth taking seriously. An AI tool that isn't supervised regularly makes a bad hire look competent. At least a bad hire will eventually tell you they don't know something.
Why Supervision Work Is Worse Than the Original Work
Here's what makes AI supervision particularly draining, in a way that's hard to articulate until you've felt it:
The work you replaced with AI was visible work. Writing an email takes 20 minutes. Scheduling a meeting takes 10 minutes. Summarizing a document takes 30 minutes. You know when you're doing it. You know when you're done. The finished product exists.
AI supervision is invisible work. Reviewing AI output doesn't feel like doing the task, but it requires enough attention that you can't really do anything else. Re-prompting to fix a mediocre draft isn't a discrete task on your calendar — it's scattered throughout the day in 4-minute increments. Monitoring a workflow automation doesn't feel like working — until it breaks, and then you lose an afternoon.
The research team that coined the term "AI brain fry" noticed something specific: the most cognitively depleted workers weren't the ones using AI most. They were the ones using AI in a certain way — bouncing between multiple tools, reviewing outputs without fully trusting them, and staying in a low-grade supervisory state most of the day.
That state is mentally expensive and productivity-invisible. You can be in it for four hours and produce almost nothing while feeling like you did a lot.
The Three-Tool Threshold
There's a number here that's worth knowing.
The original research on AI and productivity found a clean curve: going from zero AI tools to one tool produced a significant, measurable productivity boost. Going from one to two — smaller but still positive. Going from two to three — the gains start tapering. Beyond three tools, in many cases, productivity starts declining.
The most common small business AI stack in 2026 has somewhere between five and eight tools.
Those businesses are, on average, past the peak of the productivity curve and into the part where every additional tool is creating more management overhead than output value. The stack grew faster than the ability to manage it — and now the stack is the problem.
The 73% scaling back number makes sense in this light. It's not disillusionment with AI. It's people arriving, empirically, at the conclusion the research already predicted.
What the Scale-Back Actually Looks Like
The businesses that have made this work — that reduced their AI supervision burden without giving up the benefits — tend to arrive at the same approach from different directions.
They stopped treating AI as a productivity layer and started treating it as a process replacement. The difference is important. A productivity layer sits on top of what you were already doing and makes it faster. A process replacement removes the original process entirely and runs a different one. Supervision overhead is highest when AI is a productivity layer — you're still doing the original work, plus reviewing AI's contribution to it. It's lowest when AI fully handles something you've fully removed from your own plate.
They cut to a smaller stack and went deeper. The businesses reporting the lowest AI fatigue tend to use one to three tools heavily rather than five to eight tools lightly. The context-switching cost alone from managing multiple AI products — different interfaces, different prompt conventions, different integration quirks — adds up to a meaningful daily tax.
They let some automations fail rather than monitor everything. This sounds reckless, but there's a real principle underneath it: the cost of an automation silently failing for 48 hours is often lower than the cost of monitoring every automation every day. Not all automations warrant daily oversight. Deciding which ones do — and only monitoring those — is a real management decision that reduces supervision load significantly.
They stopped prompting and started templating. Every business using AI for recurring tasks should have a small library of tested prompts for those tasks. Treating prompts as something to write from scratch each time is like writing your email signature by hand on every email. The supervision work drops significantly when the inputs are standardized.
The Part That's Actually Encouraging
Here's the thing about 73% of businesses scaling back: it's not a failure story.
The businesses scaling back are the same businesses that were early enough to try in the first place. They went in, they learned something real, and they're adjusting. That's not retreat. That's the productive phase of any technology adoption — after the hype, you figure out what actually works for your specific situation.
The businesses that never tried, or that adopted AI in name only without actually using it, have none of that operational knowledge. They can't make the call to cut their stack because they don't know what their stack can do.
The frustration behind the 73% number is the frustration of people who are now informed. They know which AI tasks are worth supervising and which aren't. They know what their tools can and can't do. They have the specific knowledge to make better decisions going forward.
That's not the same as failing. That's what learning looks like when the thing you're learning is genuinely complicated.
The Practical Question
If you're in that 73% — or you suspect you're close to it — the useful question is: what am I actually supervising, and what would break if I stopped?
Not everything you're monitoring needs monitoring. Not every prompt you're writing needs to be written fresh. Not every tool in your stack is paying for itself.
The useful audit isn't "should I use AI?" That question is settled. The useful audit is: which specific tasks am I supervising that AI could handle reliably with the right setup — and which ones have I been supervising because I never fully trusted the tool enough to stop?
The answer to that question is different for every business. But almost everyone who's asked it honestly has found something they were supervising out of habit rather than necessity.
That's where the time is buried.
The Useful Daily covers practical AI for small business owners. If this was useful, consider sharing it with someone who's been feeling the same kind of tired.