Sense with Cents

There is a thing happening with AI that reminds me of something I watched happen in a very different trade, a long time ago.

Back when I was turning wrenches for a living, there were two kinds of mechanics in the shop. The old hands could walk up to a running engine, put a hand on the valve cover, listen for a second, maybe sniff the exhaust, and tell you within about thirty seconds what was wrong. They were usually right. Then there was the new breed coming up behind them — trained on code readers and diagnostic computers. Plug it in, read the code, remove and replace the part the machine points at. Run it again. If the light goes out, the car is “fixed.”

Both groups got cars back on the road. But only one of them actually understood what was happening under the hood. And when the code reader gave a misleading answer — which it did, often — only one of them could figure out why.

I think a lot of us who have been doing payroll, or accounting, or really any craft for a few decades are about to be that old wrench again. Not because AI is going to replace us. Because AI is going to make us more valuable, and quietly create a generation behind us that never gets the chance to learn what we know.

What AI Actually Does for a Payroll Pro

Let me get concrete, because abstract talk about AI gets tiresome fast.

Payroll audit work is a good test case. If you train an AI with the right context — the state or states you operate in, the special items you need to watch for, the fields you care most about, and a chunk of prior payroll history to compare against — it will catch things. It will summarize well. You can feed it tax deposit records and have it monitor them on a regular schedule. None of that is a replacement for a payroll person. All of it makes a payroll person noticeably faster and harder to fool.

If you work in somebody else's payroll department, this is worth thinking about. Bringing AI into your own workflow — thoughtfully, with the context and history you already carry in your head — makes you a more valuable member of the team. You catch more. You finish faster. You hand your manager cleaner work. That is a good thing for you and a good thing for your employer.

Just keep in mind that what you build on the clock generally belongs to the company you built it for. The prompts, the training context, the saved memories that reflect how your employer runs payroll — those live on the employer's side of the line. Your judgment and experience walk out the door with you at the end of the day, the way they always have. The AI setup you built on company time may not. Worth a conversation with your employer about where those assets should live, and who has access, before you get too far down the road.

A Small Example from My Own Desk

This week I gave an AI a data structure, a file spec, and asked it to write me the code to produce a fixed-width output file. It worked — with a few tweaks that, honestly, only an old coder would have known to look for. Edge cases. Field-boundary issues. The kind of thing that separates “compiles” from “correct.”

Then I flipped it around. I took the generated data file, uploaded it, and had the AI compare it against the original spec. It did in seconds what would have taken me an afternoon with a highlighter and a hex editor.

That is the loop. Expert knows what the right answer looks like. AI does the tedious middle. Expert verifies. The person without the expertise at either end — the one who just asks the AI to “make me a payroll file” and ships whatever comes back — is the one who gets burned.

From a Tool for Experts to a Tool That Is the Expert

Medlin Software has been quietly making this same trip since the 1980s.

The first versions ran on DOS and were written for accountants and bookkeepers — people who already knew payroll cold and just needed something faster than a pencil and a thirteen-column pad. The software assumed expertise at the keyboard. If you did not know what federal withholding was, the program was not going to explain it to you.

Over the years, as the customer base shifted from payroll pros to small-business owners doing their own books, the software shifted with it. Tax tables built in. State-by-state rules handled in the background. Forms that fill themselves out. Error checks that catch the mistakes a first-time user tends to make. Each release closed a little more of the gap between what the user brought to the table and what the job actually required.

Forty-plus years in, the software carries a lot of the expertise that used to have to live in the user's head. Not all of it — payroll still has judgment calls, and always will — but enough that a small-business owner with no formal payroll training can run a clean payroll, on time, every time.

AI is the next step on that same road. The same direction we have been walking since the floppy-disk days. A tool for experts, becoming a tool that serves as the expert, so the actual expert can spend their time on the things that still need a human brain.

Where This Gets Interesting for Medlin Customers

I may or may not be 😁 working on custom output from my software for customers to feed directly into an AI for audit and review. Over time, I could share training data and prompts to help folks get started. Since my own payroll happens to be in California, coming up with a solid California starter kit would be straightforward — I know the format, I know what to check, I know where mistakes tend to hide.

A small-business owner in California who combines Medlin Payroll, their own data, and a properly set-up AI assistant would have something close to a junior payroll analyst sitting at their desk. Not a replacement for judgment. A very capable assistant that never gets tired of looking at the same report one more time.

How You Actually Learn This Stuff

Worth pausing here to say how I learned any of what I know, because I think it matters.

I learned to code from books. Actual paper books, the kind you set next to the keyboard with a heavy object on the spine to hold it open. I learned by writing programs, breaking them, and figuring out why. And I learned by reading other people's code — sometimes code that worked, sometimes code that did not, always code written by somebody who had been at it longer than I had. You pick up tricks. You pick up discipline. You pick up a feel for when something is going to cause trouble six months from now even though it runs fine today.

I learned to be a mechanic the same way. From the old dogs in the shop who had already seen every failure mode twice. Guys who spent their days under somebody else's car and their evenings and weekends racing their own cars — a quarter mile of clay in the summer, or in the winter, a sixth-mile indoor concrete track coated in soda syrup for traction. The shop paid the bills. The track was where the real education happened. Parts came from the local wrecking yard, because that was the budget, and because a junkyard teaches you things a parts counter never will. Nobody handed them a diagnostic printout. They learned by taking things apart, putting them back together, and occasionally getting it wrong in a way they would never get wrong again.

Both paths have the same shape. Study the thing. Do the thing. Break the thing. Watch somebody better than you do the thing. Ask questions. Repeat for about a decade. At the end of it, you have judgment — which is just pattern recognition built up from enough real encounters with real problems that you can feel an answer before you can explain it.

That process does not shortcut well. You cannot read about it and have it. You cannot watch a video and have it. And — this is the part I keep coming back to — you cannot have an AI do it for you and end up with it, either. The AI ends up with it. You end up with the output.

The Part That Worries Me

Here is where I turn back into a grumpy dinosaur for a paragraph.

The tool does not make the expert. Knowing what to feed the tool, what to ask of the tool, and how to manage the tool — that is the expertise. And that expertise comes from years of doing the work the hard way first. Reading the regulations. Getting the penalty notice. Walking a client through an audit. Finding the error on page 47 of the register because something just looked off.

The person entering the field today can use AI to produce output that looks exactly like the output an experienced payroll person would produce. Same format. Same confident tone. Sometimes the same answer. The problem is they will not know when it is wrong, because they never built the pattern recognition that tells you something is wrong before you can explain why.

We are going to end up with a gap. Experts on one side. People wanting to learn on the other. And not much of a bridge, because AI will be holding the “experience” for them, and experience you do not hold yourself is not really experience at all.

Those of us who put in the years before the tools got this good are in a strange, lucky spot. We can use AI as a force multiplier without losing the judgment that makes the multiplication worth anything. The folks coming up behind us are going to have to work a little harder, and a little more deliberately, to get there — and some of them will. The ones who do will be worth their weight.

The rest will be replacing parts until the machine says it is fixed.

See also: How Many of Me Are Left?, To AI or Not to AI, There Is NO Question, and Learning or Declining: There Are Only Two States of Knowledge.