RESEARCH · HUMAN IN THE LOOP
Human tokens
are gold.
Model output has become close to free, and it keeps getting cheaper. The input that stays scarce is human judgment: the decisions that say what good looks like, the review on the calls that carry risk, the correction that should only be made once. Build automation around that scarcity, and almost everything about how you build it changes.
THE SCARCE INPUT
The expensive part isn't the work.
A capable agent can produce more in an afternoon than a person could type in a month. When the machine's output is that cheap, it stops being the bottleneck. The bottleneck moves to the human: what to build, whether it is right, and what to change.
Andrew Ng framed it for people building software with agents: AI tokens are cheap; human tokens are gold. The same economics govern automating a business. A model will draft, extract, summarize, and route all day for almost nothing. Your team's attention is the resource that does not scale, so the real design question is where to spend it.
WHERE THEY GO
Three places human judgment earns its cost.
Most of a workflow can run on cheap tokens. A few points in it cannot, because they need a person to decide what is true, what is acceptable, or what is worth the risk. Those are the points worth a human.
DEFINE GOOD
Say what correct means.
The highest-value thing a person can do is name the target: what a good outcome looks like for this workflow, in cases specific enough to test. Everything downstream is cheap once that is clear.
REVIEW THE CALL
Approve what carries risk.
The send, the payment, the exception, the step that cannot be undone. The agent prepares it and a person approves it, so judgment sits where the cost of a mistake is highest.
CORRECT ONCE
Never say it twice.
When a person catches a mistake, the fix should not be a one-off note. It should become part of the spec, so the same error cannot return and the correction is spent only once.
Where human judgment earns its cost
THE WASTE
The costly kind of forgetting.
An agent that forgets something it worked out on its own is a small loss. It burns a few cheap tokens to figure it out again. An agent that forgets something a person told it is a real loss, because now the person has to spend a golden token saying it a second time.
So the corrections a human makes are the tokens most worth protecting, and the way to protect them is to write them where the system reads them: a spec, a test, a stopping rule. A correction that becomes a check is spent once. A correction that lives only in someone's memory is spent every time the work runs.
HOW MILLIE MEASURES THIS
We treat your corrections as the benchmark. Each time a person overrides the system, that case joins the test set for your workflow, so the same mistake cannot ship again.
That is what measured, and better over time, means in practice. The system's stopping criteria hold your judgment, and an update only goes live if those checks still pass. The human tokens go in once, and the system keeps honoring them. See what we measure.
The phrase "human tokens are gold" is Andrew Ng's, from The Batch (DeepLearning.AI), July 2026. The reading above is Hello Millie's.