Issue No. 14 - September 15, 2026

I've been thinking about QuickBooks. Not because anything is broken, but because I keep choosing to open it myself instead of asking AI to pull the information for me.

There's a lot of content right now making the case that AI-automated bookkeeping is functionally equivalent to the traditional version. The same result, less friction. And maybe for some businesses that's true. But I keep opening it manually anyway, and I finally figured out why.

The act of reviewing the data IS the review of the business. When I scroll through the expense categories, reconcile a line item that doesn't look right, notice that a particular vendor cost is creeping up. That's not just bookkeeping. That's me staying current on what's actually happening. If I hand that to AI, I'm not just automating a task. I'm outsourcing the awareness that comes with doing it myself.

That realization led me somewhere bigger.

I've been thinking about what makes me go into a specific app versus asking AI to access it on my behalf. And I think it comes down to trust. Not trust in the abstract. Two very specific kinds of trust.

The first is whether I believe AI knows where to find what I'm actually looking for. Not the general category, the specific thing. In QuickBooks, I'm not looking for "expenses." I'm looking for a particular pattern across a particular time period, contextualized against what I know was happening in the business during that window. That requires knowing what to look for before you can find it. I'm not always sure AI does.

The second is whether I believe AI can translate raw data into actual information. There's a difference between those two things. Data is the number. Information is what it means given everything else I know. A vendor cost that's up 30% might be a problem, or it might be exactly where it should be because we scaled a program. The number doesn't tell you which one. The context does. And I'm the one holding most of the context.

Both of these trust questions matter. And they're also, I've realized, connected to something I've been wrestling with on the instruction side.

When I try to automate tasks by telling AI to do them "the way I would," I'm setting my current standard as the ceiling. The AI learns my process, but it also inherits my blind spots, my shortcuts, the places where I've gotten comfortable with good enough. I'm not improving the work. I'm just replicating it with fewer keystrokes.

The version of automation that actually helps me get better isn't replacing what I already do. It's AI surfacing findings and forming a hypothesis for me to evaluate. The final mile is still mine. Not because I don't trust the tool, but because that's where the judgment lives. That's the part that requires knowing the business, not just the data.

I don't think this means avoiding automation. It means being intentional about which tasks I automate and why. Some things genuinely should be handed off. Other things, the doing is the knowing.

QuickBooks stays manual for now.

Where this lands in the three buckets

Immediately actionable: Before automating any recurring task, ask one question: is the act of doing this task itself how you stay informed? If yes, automation removes the work and the awareness at the same time. That's worth knowing before you set it up.

Start thinking about: The two trust questions. Does AI know where to find what you actually need, not the general category but the specific thing? And can it translate data into information given the context only you hold? Your answers to those two questions are a better automation decision framework than any generic guide.

Next phase: Build toward a human-in-the-loop model for work that requires judgment. Let AI surface findings and form a hypothesis. You draw the conclusion. That architecture is more durable than full automation, and it keeps you in the loop without keeping you in the weeds.

Meghan Brenner is COO at JB Sales and founder of The Operator's Notebook. The Muddy Middle publishes every Tuesday.