Why Are New Employees at AI Startups Often Banned from Using AI on Day One?

If on your first day at an AI company,

the most surprising thing might not be:

The company expects you to use AI daily.

But rather, that your manager tells you:

"Don't use it yet."

This recently happened at New York fintech company Valon.

Valon is not a traditional company that rejects AI.

On the contrary.

They are building a modern software platform for mortgage services, featuring AI Agents, workflow automation, and complex financial tasks at its core.

Yet the company recently made a seemingly contradictory decision:

All new non-engineering employees have their AI access initially disabled.

It's not a permanent ban.

Nor is it because they think AI makes people dumber.

Rather, the company discovered a new problem:

AI completes work so quickly that newcomers don’t truly learn the job.

Experienced staff finish tasks in 20 minutes, potentially cutting months off newcomers' learning curves

Andrew Wang, CEO of Valon, explained that experienced employees using AI

can complete tasks

which used to take hours,

or even days,

in as little as:

20 minutes.

This sounds like an unqualified good thing.

If a company invests in AI,

isn't the goal to be:

faster,

cheaper,

and more efficient?

The problem is,

those tedious tasks eliminated by AI used to serve another important purpose:

training newcomers.

How did new employees learn mortgage service work?

Not just by reading a manual.

But by encountering:

reconciliation errors,

fund transfer anomalies,

data mismatches,

system exceptions,

or process errors,

and then tracing back to find out:

where exactly was the mistake?

Which data came in first?

What changes were made in which system?

What counts as a normal exception?

When is there a real problem?

These tasks were slow and frustrating.

But over time, newcomers gradually built a:

mental map of how the job actually works.

The biggest issue with AI isn’t wrong answers, but newcomers’ inability to detect those errors

Valon points out a crucial situation.

New employees can quickly get a

roughly 95% accurate answer from AI.

95% sounds quite good.

But the real danger lies in the remaining 5%.

Someone with years of experience might immediately recognize:

This is incorrect.

The money shouldn't move this way.

A rule exception is missing.

This customer’s case requires special handling.

Newcomers seeing the same answer may just think:

It’s well-written.

The logic flows well.

It should be fine to use.

So the question isn’t:

Does AI make mistakes?

But rather:

Does the user have the ability to recognize AI’s mistakes?

Mistakes used to be part of the learning process

This is an often overlooked issue when integrating AI into work.

We typically equate efficiency with:

fewer steps,

less time,

and the ideal of pressing a button

to get an immediate answer.

But if you are learning something,

those skipped steps

can actually be the learning itself.

A junior accountant reconciling accounts for the first time

starts to understand where each figure originates.

A customer service trainee who misjudges a customer question

learns which exceptions can’t be answered by standard replies.

An engineer debugging for the first time

gradually builds system interrelationships.

An entry-level marketer analyzing data themselves

discovers that correlation is not causation.

If AI delivers

answers,

summaries,

judgments,

and solutions immediately,

newcomers may quickly seem

competent.

But when an unprecedented exception arises,

they might find themselves completely lost on what to do next.

So Valon isn’t permanently banning AI, but establishing "unlock conditions" first

This is an important distinction.

Valon hasn’t said:

New hires can’t use AI for a year.

Nor have they declared:

AI is harmful to employees.

Andrew Wang calls the temporary AI access suspension a:

rough, provisional measure.

The current principle is:

Once a manager confirms the newcomer understands enough to

question models,

validate outputs,

and detect errors,

AI access is granted.

In other words, the condition isn’t about:

"how long before you can use AI,"

but

"do you have the ability to supervise AI?"

This distinction is huge.

Interestingly, the issue now affects experienced employees too

If newcomers produce large amounts of output that looks thorough but contains errors,

who fixes them? Usually, the senior staff do.

According to Business Insider, some experienced Valon employees support this policy

because they’ve had to spend time cleaning up AI outputs that newcomers didn’t adequately check.

This creates a strange illusion of efficiency.

On the surface:

New hires finish a task in five minutes.

In reality:

Senior employees then spend 40 minutes reviewing and correcting.

Company dashboards may show:

AI usage increasing,

higher individual output,

but the overall workflow may not actually be faster.

More AI usage doesn’t mean a more AI-native company

Wang also discovered another issue:

Sometimes employees use high-end, costly AI models for even simple tasks.

This is a common pitfall when adopting AI in businesses:

Confusing using AI with genuine progress.

The real questions should be:

Does this task really require AI?

Does it require such a powerful model?

Who reviews the AI output?

If it’s wrong, what is the cost?

What exactly does the company save?

Valon estimates that after adjusting their AI usage,

annual token expenses might drop from roughly $15–20 million to $4–5 million.

This is their internal estimate, and not a guaranteed 70% savings for all companies.

But it proves one thing:

More AI usage is different from better AI usage.

The real impact hits the disappearance of the apprenticeship model

In many professional roles, newcomers traditionally go through an unpleasant phase:

organizing data,

drafting initial versions,

researching documentation,

reconciling numbers,

running tests,

and handling minor issues.

Senior staff handle key decisions,

newcomers take on foundational work.

This system has flaws.

Sometimes it just dumps boring tasks on juniors.

But it’s also the foundation of many industries’

apprenticeship training model.

You start with small tasks,

make small mistakes,

get corrected,

gradually understand the entire system,

and eventually take responsibility for major decisions.

Generative AI is most capable of

drafting,

organizing,

searching,

comparing,

synthesizing,

and basic analysis.

So companies face a new problem never seen before:

If AI takes over all entry-level tasks, where will future senior employees come from?

This isn’t about telling newcomers "don’t learn AI"

Saying newcomers shouldn’t use AI is also an extreme stance.

In the future, career skills will inherently include:

the ability to use AI,

know how to assign AI tasks,

verify AI outputs,

and recognize when not to use AI.

Valon’s approach isn’t noteworthy because they initially ban AI usage.

It’s noteworthy because they’ve broken down employee capability development into stages:

Stage one:

Understand the work.

Stage two:

Then amplify your abilities with AI.

The order cannot be fully reversed.

Without stage one,

stage two risks becoming merely:

AI does the work,

and people only believe it.

Ultimately, newcomer training needs redesigning

In the past, companies asked:

What materials should new hires study in the first week?

Who should they shadow?

How long until they can work independently?

Now, another question may emerge:

When can they start using AI?

The answer shouldn’t be:

Full AI access from day one,

or

AI completely forbidden for three months.

A more reasonable approach might be:

Which tasks can newcomers start using AI on?

Which tasks must they at least do manually once?

What errors must they be able to independently recognize?

What competencies qualify them to delegate tasks to AI?

AI can even become a teacher,

not just an answer generator.

For example, require newcomers to first judge a case themselves,

then let AI offer alternative perspectives.

Have them identify errors before asking AI to help.

Have them draft workflows before AI searches for gaps.

That way, AI doesn’t remove the learning process,

but rather

accelerates feedback.

In the AI era, judgment may be rarer and more valuable than answers

Information was once hard to find.

So people who knew answers were valuable.

Now answers are increasingly cheap.

In seconds you can generate:

emails,

reports,

programs,

analyses,

presentations,

and recommendations.

What becomes precious is:

how you know whether an answer is trustworthy.

This kind of judgment is hard to get from a simple prompt.

It usually comes from:

doing work,

making mistakes,

tracking problems,

observing exceptions,

and finally building experience.

So Valon’s seemingly anti-AI decision

actually reflects a more radical AI mindset.

It’s not about:

"How do we get everyone using AI as fast as possible?"

But about:

"How do we ensure AI magnifies human capability rather than human ignorance?"

This may be the true next challenge companies face.

When AI allows newcomers to skip all the tedious steps,

what companies need to preserve

may not be the tedious tasks themselves,

but what those tasks originally taught.

Today, grow together with AI.

Learn one AI skill daily.

Save a little time every day.

Improve a little more every day.

SasaDaily, growing with you.

Recommended Reading

Cisco Rolls Out AI Agents to 90,000 Employees Yet The Real Challenge Is Redesigning Work

Why Only 16% of Japanese Companies Fully Adopt Generative AI Despite 86% Having Used It