Architecture Before AI

The dominant conversation about AI is the wrong conversation.

Which model. Which tool. Which prompt. Which agent.

The engine, in other words.

The thing under the hood. The horsepower number on the spec sheet. The benchmark score.

That is the cheap half of the work.

The expensive half is engineering the car the engine sits in.

Almost nobody is doing it.

In 1978, Chrysler launched a front-wheel-drive economy car called the Dodge Omni. The engine was a 1.7-liter four-cylinder sourced from Volkswagen. It produced 75 horsepower. The car weighed 2,145 pounds. It got 39 miles per gallon. It was not designed to be quick. It was designed to be cheap, efficient, and forgettable.

In 1984, Carroll Shelby got hold of it.

He started layering on power. First, a high-output 2.2-liter. Then a turbocharger. By 1986, the final version, the Shelby GLHS, produced 175 horsepower. More than twice the original engine. They built exactly 500 of them.

The platform did not change.

Same wheelbase. Same suspension geometry. Same front-wheel-drive layout designed for fuel economy. Same chassis, same bushings, same steering, same tires.

On paper, they had a giant-killer. Faster than a Mustang GT. Faster than a Camaro IROC-Z. For a third of the price.

In practice, when the boost hit, the steering wheel tried to rip itself out of your hands.

The car would change lanes on its own.

Hagerty, the magazine of record for collector cars, said it plainly: nobody figured out how to manage the torque steer until electronic stability control arrived a decade later.

The engineers had not built a faster car.

They had built a car that physically altered its own direction under power.

This is the picture I want you to hold.

The danger of the GLHS was not power. The danger was the uncontrolled delivery of power through a platform that was never engineered for it.

Now do the translation.

Your AI is the engine. Your organization is the car.

The model is impressive. The benchmark is real. The horsepower under the hood is genuinely greater than it was eighteen months ago.

The platform around it (your decision rights, your judgment loops, your context architecture, your agreements, your strategic coherence, the actual structure through which work moves in your company) was engineered for a different purpose.

It was engineered for slower, predictable cognitive load.

The kind of work where a person drafted a thing, sat with it, asked clarifying questions, and pushed back before it went anywhere.

You are now pushing 175 horsepower through a chassis built for 75.

It is not failing the way you think it is.

It is not “the AI made a mistake.”

It is “we are no longer recognizing the direction we are heading, and we cannot smoothly throttle back.”

There is a clean distinction worth naming here.

Power installed is not the same as power controlled.

Installing AI is the cheap half. You can buy seats. Deploy tools. Run pilots. Hire a head of AI. Announce it on LinkedIn. None of that engineers the car.

Controlled power is the expensive half. It is the work nobody buys because it does not benchmark and it does not demo well.

It looks like rewriting decision rights.

Rebuilding review loops.

Naming what your company actually does and does not do.

Locking strategic coherence so volume amplifies the right signal.

Building real measurement systems instead of repurposed dashboards.

Designing the agreements between people and AI output the way an engineer designs the connections between an engine and a chassis.

This is the work the GLHS engineers skipped.

It is the work most founders are skipping right now.

I want to name the specific signal.

When your platform is not built for the power you are running through it, you will not feel “we are going too fast.”

You will feel something subtler.

The steering wheel starts to feel a little loose.

You will notice you are producing more output and recognizing less of it.

You will notice you authorized a direction and the company arrived somewhere adjacent.

You will notice that throttling back on AI usage feels jerky and disproportionate. Lift the gas and the back end gets light.

You will notice the smart, experienced people in your company are quietly working harder to verify, sanitize, and re-direct AI output downstream. And you cannot find that labor on any P&L line.

Those are the symptoms.

Those are the chassis flexing under load.

The good news.

The fix is not slowing down on AI. The engine is not the problem.

The fix is engineering the platform that the engine sits in.

What plays the role of torsional stiffness in your company? Whatever holds strategic direction rigid when volume amplifies it. Most companies do not have this. They have a vision deck.

What plays the role of suspension geometry? Whatever keeps the right processes loaded under different conditions. Most companies have a process designed for one load condition. AI changes the load condition.

What plays the role of traction control? Whatever stops a single team from spinning its wheels and dragging the rest of the company sideways. Most companies do not have this either. They have heroics.

What plays the role of real tires? Whatever actually grips the new ground AI puts you on. Measurement systems. Ground-truth verification. Feedback loops that close fast.

The list is long.

It is also boring.

None of it shows up in a keynote.

It is the work.