Skip to Content

THE NIGHT WE HANDED AI THE CAR KEYS (AND WHY WE DIDN’T LET IT DRIVE FAR)

March 12, 2026
Martin Kastenbaum

We were reviewing a team’s attempt to use AI to modernize a legacy codebase — the kind of system that’s been quietly running the business for twenty years, held together by tribal knowledge, duct tape, and a few “don’t touch that” comments buried in the code.

As we walked through the AI‑generated refactors, something clicked.

It felt a bit like handing the keys of a very expensive car to an enthusiastic child.

Not because the child is malicious. Not because the car is unsafe. But because enthusiasm and horsepower don’t automatically add up to good judgment.

And that’s exactly the dynamic we’re seeing with AI‑assisted modernization.

The Enthusiastic Child Behind the Wheel

AI is eager.  

AI is fast.  

AI is wildly confident.

It will happily tell you it can drive — and it will say it with the kind of certainty only a child can muster.

And to be fair, it can do some impressive things:

– Spot patterns humans overlook  

– Suggest refactors in seconds  

– Surface hidden dependencies  

– Explain unfamiliar code with surprising clarity  

If you’ve ever spent a weekend untangling a legacy module, AI feels like a miracle.

But then you look closer.

Where Things Start to Drift Toward the Guardrail

Legacy systems aren’t just code.  

They’re history.  

They’re politics.  

They’re a decade of business logic layered on top of decisions no one remembers making.

AI doesn’t see any of that.

So it does what enthusiastic children do:

– Fixes things that weren’t broken  

– “Simplifies” logic that was intentionally weird  

– Removes edge cases that matter  

– Produces code that compiles but quietly breaks the system’s soul  

It’s not trying to crash the car.  

It just doesn’t know what the brake pedal does yet.

The Hidden Question: Is This Actually Saving Us Time?

This is the part teams whisper about after the demo.

AI can rewrite a file in seconds — but validating that rewrite can take hours.  

The paradox is simple:

AI accelerates creation.  

Humans still own correctness.

So the real question isn’t “How fast is the AI?”  

It’s “How much human time does it actually remove from the process?”

And the answer depends on what you measure.

Where AI really helps

– Faster understanding of unfamiliar modules  

– Instant first drafts of refactors  

– Quick pattern detection  

– Reduced cognitive load on repetitive tasks  

Where humans still spend time

– Correcting hallucinations  

– Re‑running prompts  

– Reinstating business logic AI flattened  

– Checking for side effects AI can’t see  

– Ensuring the system still behaves like the system  

Sometimes it feels like babysitting a very fast intern:  

helpful, but not hands‑off.

The ROI grows when teams measure the right things:

– Time‑to‑understanding  

– Time‑to‑first‑draft  

– Review‑to‑edit ratio  

– Error discovery speed  

– Reduction in cognitive load  

Not lines of code.  

Not “percentage of code rewritten.”  

Not the illusion of speed.

Legacy Code Isn’t a Playground

This is the part people underestimate.

Legacy systems encode:

– Implicit contracts  

– Undocumented behaviors  

– Tribal knowledge  

– Business rules that only exist in someone’s head  

– Constraints that were never written down  

AI can read syntax.  

It can’t read context.

And modernization without context is just rearranging parts of a machine you don’t fully understand.

So What’s the Right Role for AI?

AI belongs in the passenger seat — not the driver’s seat.

Use it for:

– Drafting refactors  

– Surfacing patterns  

– Explaining unfamiliar modules  

– Proposing modernization paths  

But keep humans responsible for:

– Validation  

– Architecture  

– Business logic  

– Risk assessment  

– Knowing when “simpler” is actually dangerous  

AI accelerates the work.  

Humans ensure the work is correct.

That’s the partnership.

The Architect’s Job Isn’t Going Anywhere

If anything, AI makes the architect’s role more important.

We define the destination.  

We set the guardrails.  

We decide what “good” looks like.  

We understand the history, the politics, the trade‑offs.

AI can help us get there faster — but only if we keep a hand on the wheel.

The Future of Modernization

AI isn’t replacing engineers.  

It’s not rewriting your legacy system while you sleep.  

And it’s definitely not ready to take the car out on the open road alone.

But with the right supervision, it becomes an incredible accelerator — the enthusiastic helper that makes the work lighter, faster, and more fun.

The key is simple:  

Let AI drive the enthusiasm.  

Let humans drive the car.

About the author

Applications & Cloud Technology | United States
As a Senior Manager in the Applications & Cloud Technology practice and a Fellow at Sogeti Labs, Martin specializes in Kubernetes, DevOps, and AI-assisted development, helping shape technology strategy and mentor teams across the organization.

Leave a Reply

Your email address will not be published. Required fields are marked *

Slide to submit