We Automated the Easy Half
Every technology wave gets remembered by what it made cheap. Steam made power cheap. The shipping container made distance cheap. The internet made copies cheap, then distribution cheap, and we spent two decades rearranging the economy around that single fact. The current wave gets summarized the same way: artificial intelligence makes intelligence cheap.
That framing is close enough to be useful and wrong in the way that matters. What has collapsed in cost is production — the generation of plausible output. What has not moved is the cost of deciding whether that output is right.
Almost every piece of knowledge work has two halves. There is the making, and there is the checking. Draft the contract, then verify it says what you meant. Write the function, then confirm it does what you claimed. Assemble the research, then establish that the sources exist and actually support the conclusion. For the entire history of professional work these halves stayed roughly proportional, and the making was expensive enough that it rationed the checking automatically. Nobody had to design a process for reviewing ten thousand contracts a week, because nobody could write ten thousand contracts a week.
We have just removed that constraint from one half and left the other exactly where it was.
The bottleneck moved, and we cheered for the wrong step
Anyone who has run a factory floor knows what happens next. Speed up a station that was never the constraint and you do not get a faster line. You get a larger pile in front of the station that is. The pile is the real output of the improvement, and it grows in proportion to how proud you are of the change you made.
That pile has a name inside most companies right now, and the name is usually a person. It is the senior engineer whose queue of code review has quietly tripled while the amount of code being written went up by an order of magnitude. It is the general counsel reading drafts that arrive faster than any human can absorb them. It is the analyst who now spends her week confirming whether the summary she was handed is a summary of anything real. Every one of them will tell you that the work got faster. Almost none of them will tell you the organization got faster, because the organization moves at the speed of the slowest step, and we did not touch the slowest step.
There is a subtler version of the same problem that shows up in individual work rather than in queues. Checking is not merely slow — it is a different cognitive act, and it is the harder one. Producing a plausible answer requires knowing a domain well enough to sound right in it. Verifying an answer requires knowing the domain well enough to catch a confident thing that is subtly wrong. The first is undergraduate work. The second is expertise. We have automated the undergraduate half and left the expert half to a shrinking number of people whose expertise, uncomfortably, was built by doing the undergraduate half for years.
The temptation is not bad output. It is skipping the check.
I want to be precise about the actual risk, because the popular version of this worry is the wrong one. The danger is not that these systems produce garbage. Mostly they do not. Quality is high and climbing, and I use them constantly and would not give them back.
The danger is economic. When one half of a process becomes free and the other stays expensive, pressure builds against the expensive half — not through any decision anyone announces, but through a thousand small moments where checking looks like the thing slowing everybody down. Nobody circulates a memo saying we no longer verify. What happens is that a deadline arrives, the output looked fine, the last four looked fine, and shipping without a careful read becomes the cheapest available path. Do that enough times and you have not adopted a tool. You have lowered a standard, and you have done it without ever choosing to.
Software has spent decades building an answer to precisely this problem, which is why software is a useful place to look. Tests, type systems, continuous integration, staged rollouts and code review exist because we learned the expensive way that writing code is easy and knowing it works is not. That whole apparatus is machinery for making verification cheap and repeatable, so that the expensive half stops being a human bottleneck without stopping being a check. Every other field is about to need its own version, and most do not have one yet. There is no unit test for a legal memo, no continuous integration for a market analysis, no staged rollout for a diagnosis.
The next decade belongs to whoever makes checking cheap
If you want a rough map of where durable value gets built from here, that is it. Not faster generation — generation is heading toward commodity, and anyone can rent it by the token. The scarce and defensible thing is trustworthy verification: provenance you can follow back to a source, evaluation harnesses that catch the specific ways a system fails in your domain, tools that make a claim auditable in seconds rather than hours, formats that carry their own evidence. Boring infrastructure, mostly. It rarely demos well. It is where the constraint actually lives.
The organizations that will pull ahead are the ones that treat review capacity as the scarce resource and invest in it deliberately, instead of treating it as friction to be minimized. That means measuring throughput at the checking step rather than the making step, and it means being honest that a team producing five times the output while reviewing the same amount has not gotten five times more productive. It has gotten five times more exposed.
And where verification genuinely cannot be automated — where the judgment is irreducibly human because the consequences are — the answer is not a better tool. It is a named person who is answerable for the result. Not a process, not a policy, a person, because accountability that belongs to everyone belongs to no one. Automating the easy half is the achievement of the decade. Refusing to let it quietly erase the hard half is the discipline that decides what the achievement is worth.