The work moved

The part that looks like thinking isn't. It's probability, arranged on purpose.

When you engineer a context well, you're writing a program — just not the kind made of rules. Traditional code is if-then: you state the rule, the machine follows it, the same input gives the same output every time. A prompt doesn't work like that. You assemble words, examples, and instructions that make the output you want more likely than the ones you don't. You're not specifying the answer. You're shifting the odds toward it.

Call it a clever parrot if you like, but the label undersells what's useful here. With rules, someone has to know the rule in advance and write it down. With probability, you can mine a pile of text for what good output looks like and make that probable — even when nobody could have stated the rule cleanly. That's real. It does things the if-then world couldn't.

The cost sits in the same sentence as the benefit. It will never be deterministic. The same context can produce a good answer today and a worse one tomorrow, and you can't fully close that gap, because closing it would mean turning probability back into rules — the thing you gave up to get here.

Here's the part most people skip. The pitch is that AI lets you avoid the hard analysis. Stop extracting requirements, stop writing the logic — just describe what you want and let the model figure it out. It looks like the effort disappeared.

It didn't. It moved. Instead of extracting rules, you're now building a context in which the answer you need has a chance of showing up. That's still analysis. You still have to know what good looks like, what to include, what to leave out, and how to tell when the output is quietly wrong. The work changed shape. It didn't shrink.

And it doesn't stay done. A rule, once written, holds. A context doesn't, for two reasons that compound. The model isn't yours — a new version, a different temperature, a vendor's quiet retune, and the context you tuned last month reads differently. And the context is standing in for logic, so when the facts it leans on go stale — prices, names, policies — it doesn't throw an error the way broken code would. It just produces worse answers, and nothing tells you. Both pull the same way: you're back, adjusting, re-checking, re-grounding. That ongoing maintenance is most of what "managing agents" actually means, and it's the line nobody put in the budget.

This is the barbell seen from the inside, and the flip side of knowledge that isn't there: the engine is probabilistic, so the judgement doesn't leave. It relocates — to building and maintaining the context the engine runs on.

If your AI plan assumed the analysis went away, it's worth finding where it actually went. Reach out — happy to think through what that costs you, and where.