I’ve spent twenty-six years building systems that take work off people’s desks. I believe in the tools. Search “title examiner training AI” and you’ll get pages promising the ramp gets shorter — new hire, smart software, productive in months instead of years. It’s a good story. I think it’s exactly backwards. Automation doesn’t flatten the learning curve. It deletes the bottom of it.

Nobody Learns Judgment From a Manual

Ask a good examiner how they knew something was wrong on a file and you’ll get an unsatisfying answer. It looked off. Something about the way that release was worded. That isn’t mysticism, and it isn’t talent. It’s three hundred clean files — three hundred times the chain ran straight and the release was recorded where it should have been. The pattern library gets built out of the boring ones. Then file 301 comes across with a satisfaction that references the wrong book and page, and it registers in the back of your neck a full beat before you can say why.

You can’t get that from a training binder, and you can’t get it from shadowing either — watching someone else clear a file teaches you their conclusion, not the sensation of normal. You have to do the reps yourself, and most of them have to be unremarkable — unremarkable is the baseline everything else gets measured against. Expertise is compressed experience, and most of it is dull.

The Trap, Stated Plainly

Here’s the part that keeps me up. Automation is most valuable exactly where it is most destructive to apprenticeship. That isn’t bad sequencing; it’s the same property. Repetitive and predictable is what makes a task worth automating. Repetitive and predictable is also what makes it good practice material. High-volume, low-complexity work — the clean residential refi, the file where nothing is wrong — is simultaneously the best ROI on the automation spend and the entire on-ramp for the next generation of examiners. You can’t remove one without removing the other, because they’re not two things.

So the economics push, hard and rationally, toward stripping out precisely the material that builds the people you’ll need in a decade. Every individual decision is correct. The aggregate is a shop where the only files a junior ever touches are the ones nobody knows how to teach. That’s a steeper curve, not a flatter one. Same distance to competence, no bottom rungs.

“The AI Will Handle the Hard Files Eventually” Is Not an Answer

I hear this from people on my side of the table, and I don’t buy it. Set aside whether it’s true. Someone in your building still has to look at the output and know whether it’s right. A shop with automation and no independent judgment isn’t running the software. It’s trusting it. Those are different risk positions and they get priced differently the day something goes wrong. The moment nobody on staff can evaluate a hard file on their own, your vendor’s error rate quietly becomes your error rate — and you have no way to detect it, because detection was the thing you stopped staffing for. I’m a vendor telling you not to be in that position with respect to vendors. That’s the whole point. I’d rather sell to a shop that can catch me being wrong.

The Rung, Not the Ladder

The July 28 post argued that AI raises the floor on the minimum competence required to hold a job, across a lot of industries. I still think that works out over a hundred years and is genuinely painful over twenty — and the people it happens to don’t much care about the hundred-year version. Title is one instance. The entry-level seat isn’t going away because the machine is doing senior work. It’s going away because the machine is doing junior work — and the junior work was how anyone ever got to be senior.

Manufacture the Reps

The answer is not to refuse to automate. Declining while your competitors don’t is unilateral disarmament, and it protects nobody’s career — you lose the shop and the training program with it. The answer is to deliberately manufacture the reps that used to happen by accident. Have your junior examine files the software already cleared, then compare. The places where they disagree with the machine are your curriculum, and they’re better teaching material than raw files ever were, because a disagreement isolates the judgment call. No hunting for the interesting part. The interesting part is circled.

Rotate experienced people back through the automated steps on some cadence, even when it’s inefficient; knowledge atrophies fastest in the people who already have it, and they lose it quietly. And change what you think the output is for. Most shops treat automated work product as a finished thing that goes in the file. It’s also a record of how a hard call got made, sitting right there. That’s the reframe worth taking from this: automation can generate better training material than the manual process ever did. It just won’t do it by itself. You have to build that on purpose, and almost nobody is.

The Part With No ROI Story

I’m not going to pretend this is free. You are paying a person to redo work a machine already did. There’s no clean line on the P&L where that pays for itself, and if you ask me for a payback period I don’t have one. It’s deliberate inefficiency, funded out of margin you’d rather keep. I think it’s cheaper than the alternative. The alternative is finding out in six years that nobody in the building can evaluate a hard file, and by then you can’t buy your way out — the people who could do it learned somewhere that no longer exists.

Where the Software Earns Its Keep

This is a design question we argue about internally, and I’d rather be honest about the position than sell around it. Scribe shows its reasoning, not just its conclusion — which document it’s reading, what it thinks the issue is, why. Partly that’s for the examiner clearing the file. Partly it’s so the output can function as a teaching artifact instead of a verdict. Same for what Autopilot does upstream: work you can inspect is work someone can learn from. I’m not claiming we’ve solved this. I’m claiming it’s the right thing to optimize for. If your vendor’s output is a black box that hands you an answer with no path to it, ask them why. The answer matters more than the accuracy number.

The shops that come out of this decade well won’t be the ones that automated first. They’ll be the ones that automated and still had someone who could tell when it was wrong.

Ask your vendor how their AI reaches a conclusion →