It fails months earlier, in a design review nobody wanted to sit through.
I started my career building plants. If you've never done it, the part that surprises people is how little of the work is construction. Long before anyone pours a foundation, you live inside the design. You develop the P&IDs line by line, arguing over what each stream does and where it goes. You build the specifications: what every pump, valve, and vessel has to tolerate, and why. You sit through safety reviews that pick apart your assumptions one failure mode at a time, asking what happens when this reads high and that runs dry. It is slow, it is unglamorous, and it feels like a tax on getting started.
It is the cheapest insurance you will ever buy.
I learned that on the back end, where the bill comes due. I worked a startup that went south, and I can tell you it is not pretty when a plant that looked finished refuses to run as designed. I once spent 6 months commissioning a unit that was scheduled for 2. Not because the steel was wrong. Because every loop has to be proven before it carries load, and the gaps you waved past in design all come back to be settled at the worst possible time, with the whole project waiting on you.
I'm watching executives implement AI the way a reckless team builds a plant. Straight to construction. It demos beautifully and then buckles the moment real work hits it.
A serious AI architecture has more in common with a capital project than a chat window. It lives or dies on the same two disciplines a plant does.
The first is front-end design. The time you spend up front on specification and review is what buys success on the back end. At any real scale, an AI architecture has to be deliberately structured, because the failure modes are structural. Feed a system everything and it rots: context bloats, instructions blur, and quality decays across the very use cases you built it for. Designing what each part sees, what it does, and what good output means is the AI equivalent of a P&ID. Skip it and the system still answers. It just answers with junk.
The second is commissioning. In a plant, nothing carries load until it has been validated against the conditions it will actually meet. AI deserves the same loop testing, and almost never gets it. An unvalidated build works right up until an unexpected use-case walks in the door, and then it fails, quietly and confidently. Each failure is rework, each rebuild is cost, and you pay it in production instead of in design, which is the most expensive place to pay.
None of this is an argument against experimenting. Building something quickly in a chat window is the right first move. It is the fastest, cheapest way to find out what AI can actually do for your business, and the use cases it surfaces are real signal. That is exactly what it's for. The mistake is treating the experiment as the design, and scaling a thing that was never engineered to carry weight.
If you want to know whether your AI investment is built on steel or on hope, three characteristics tell you.
Separate discovery from deployment. Let the chat experiments do what they're good at, surfacing use cases, and stop there. A great prototype is permission to design, not a system to scale.
Do the front-end design before you scale. Specify the use cases, the context each component should hold, and the standard for acceptable output before you build for the whole organization. Write the spec you'd be willing to defend in a review.
Commission before you commit. Validate each piece against the edge cases it will really face, not the happy path it was demoed on. Find the failures while they're cheap.
Here is what makes this practical rather than a counsel of perfection: the same AI is on both sides of the work. It can help you draft the spec, pressure-test the design, and find the edge cases you would have missed. Then it builds from that spec once you've set it. The up-front discipline isn't a slower way to work. It's the part AI is best positioned to carry, and the reason a designed system runs while an improvised one stalls.
A plant built without a design review still looks like a plant. Right up until you ask it to run.
Until next week,
Kendall -
Find me on LinkedIn or Book a 1:1 call
Not a subscriber yet? Subscribe here