AI made the task ‘cheap’
The quotes do the work. Producing something that looks like the deliverable got cheap. What makes it right doesn't come with the tools.
I watched a drafting engine burn through three million tokens on a single deck.
The code was clean. The rendering was beautiful. The output was unusable. Not rough. Not “needs a pass.” Unusable.
You’ve seen this everywhere. AI site builders. Document generators. Code reviews that sound thorough and catch nothing. Research briefs that hold up until someone checks one citation. Law firms sanctioned for citing cases that don’t exist. A major consultancy refunded a government over the same thing.
Everyone keeps saying AI made the task cheap. Put quotes around cheap. What got cheap is producing something that looks like the deliverable. The distance between looks-done and is-right didn’t shrink. It got harder to see.
Here’s the deal. Most people think AI means a chatbot. A rung up, it’s calling APIs. Past that you hit the serious crowd: context engineering, system design, agents, evals. Real skills, and we run all of it, every day.
Still not enough.
The last ingredients don’t come with the stack. Problem-solving, and domain depth, vertical and horizontal. You can lay every tool out on the counter and still be missing what makes it rise. A cake without yeast is not a cake. It just looks like one.
With every ingredient in, the same models behind all that slop turned 220 documents into a case package a senior litigator called the best he’d seen in forty years.
And we still miss. That deck engine I opened with was ours. We tore it down and rebuilt it. Anyone who tells you they have this fully figured out is selling something.
It moves too fast for recipes anyway. New models every month, new tools every week, new methods every day. Whatever stack you bet on is dated by next quarter.
What persists is the approach.