Company Discovers You Should Know Your Job Before Asking AI to Do It

There is finally an AI policy every employee can understand:

First, learn how to do your job.

Then you can ask AI to do your job.

Valon, an AI-driven mortgage servicing software company, recently turned off AI access for new non-engineering hires. Employees get access back once their manager believes they understand the work well enough to question the model and verify its output.

Which is a surprisingly old-fashioned policy for a company operating at the frontier of AI adoption.

It is also increasingly hard to argue with.

The problem wasn’t that AI wasn’t working.

It was that it was working before some employees knew enough to know when it wasn’t.

The Grunt Work Was the Training Program

Most people didn’t learn their jobs by reading the employee handbook.

They learned because somebody gave them a messy spreadsheet.

Or made them reconcile the numbers themselves.

Or asked them to sit through a customer call.

Or handed them a deck that needed to be rebuilt by tomorrow morning.

You did enough of the annoying work that eventually you understood why the work was annoying.

Then you got faster.

Then, ideally, you got good.

AI threatens to reorder that sequence.

A new employee can now produce something that looks like the work of an experienced employee long before they have the judgment of an experienced employee.

That’s useful right up until something is wrong.

Then the person reviewing the work has a new question:

Did you make a mistake?

Or did you paste in a mistake you don’t understand?

Terrible AI output is easy to reject.

An answer that is almost completely convincing requires someone who understands the subject well enough to find the part that isn’t.

Which means the scarce skill may not be prompting.

It may be knowing enough to disagree with the answer.

“You did enough of the annoying work that eventually you understood why the work was annoying.”

We May Have Automated Part of the Career Ladder

There’s a bigger management problem hiding here.

Companies have spent years trying to remove low-value work from expensive employees.

That makes perfect sense.

The senior analyst shouldn’t spend three hours rebuilding a report if AI can help finish it in 20 minutes.

But that three-hour report may also have been how the junior analyst learned why the numbers moved.

The experienced employee remembers doing the work manually.

The new employee enters a workplace where the manual version barely exists.

So how does the new employee become the experienced employee?

That’s a considerably harder problem than buying everyone an AI license.

For most of modern office history, apprenticeship wasn’t always formally called apprenticeship.

It was the work.

You reviewed somebody else’s spreadsheet.

You sat next to the experienced person.

You handled the easier version of the problem.

You made mistakes while the stakes were still relatively low.

Eventually somebody trusted you with the harder version.

If AI eliminates enough of those intermediate steps, companies may need to become much more deliberate about replacing the learning that disappeared with them.

Otherwise we risk creating an odd corporate hierarchy:

Experienced employees using AI because they already understand the work.

New employees using AI because they don’t yet understand the work.

And managers trying to figure out which is which.

And Yes, Someone Eventually Opened the Spreadsheet

There is also, naturally, a budget angle.

Business Insider reports that Valon expects the changes around AI usage to reduce annualized token spending from roughly $15 million–$20 million to roughly $4 million–$5 million.

So we have reached an important milestone in enterprise AI adoption.

Companies are now concerned employees are using too much AI.

Somewhere, a manager is preparing a presentation titled:

RESPONSIBLE TOKEN UTILIZATION FRAMEWORK

But the cost issue is almost secondary.

The more interesting lesson is that one of the companies pushing aggressively into an AI-enabled future has run into a very human limitation.

AI can give you an answer before you have earned the knowledge required to evaluate it.

That isn’t necessarily a reason to stop using AI.

It may be a reason to rethink what training looks like when the shortcut is available on Day One.

Because the goal was never supposed to be having employees who could produce work that looked right.

The goal was having employees who knew when it wasn’t.

And apparently we needed AI to remind us of that.