A technically more accurate AI process can still be a worse system if it wastes disproportionate human time, money, attention, or momentum.
By Pawan Nayar · from the LLOS AI Learning record
A technically more accurate AI process can still be a worse system if it wastes disproportionate human time, money, attention, or momentum.
Optimize the total cost of achieving a trustworthy result. Accuracy matters, but it does not exist alone. Suppose one workflow produces 98% confidence in two minutes and another produces 98.2% confidence after forty minutes, hundreds of tool calls, and substantial human frustration. The second workflow is not automatically better. A useful AI system considers several costs together: * consequence of an incorrect answer, * cost of additional verification, * reversibility of the decision, * urgency, * computational expense, * human waiting time, * cognitive interruption, * and the value of additional confidence. This is especially important for agentic AI because agents can consume resources autonomously. An intelligent agent must learn not merely how to continue. It must learn when continuing is no longer worth the cost. **The axis this adds.** Every cost rule in this repo prices money paid to an API — the 5¢ spend gate, the estimate-is-the-budget rule, the thinking-budget rule. None of them priced Pawan's waiting time or his attention, so a workflow could stay perfectly inside its dollar budget while spending an hour of his day and returning almost nothing. Waiting time and interruption are real costs and they belong in the same arithmetic as the API bill. And where a human review stage already exists downstream, extra machine certainty ahead of it is usually the most expensive confidence available.
AI is asked to produce an internal prototype that will be reviewed by a human before publication. Instead of recognizing that the work is reversible and already has a human review stage, the agent attempts production-grade validation of every possible state before presenting anything. The additional verification delays the human who was supposed to provide the next meaningful judgment. The AI has optimized its own local certainty while making the overall human-AI process worse.
Resolved from this lesson's own record — real definitions, not invented for this page.
More from the same record, pulled by shared topic — not written new for this page.
An AI model should execute all approved steps consecutively. The AI model should not pause to ask for permission for each next step.
Enough to reduce meaningful risk, but not so much that verification becomes more expensive than the uncertainty it is trying to remove. Every additional check consumes something: * time, * computation, * money, * tool calls, * human attention, * and opportunity to make progress elsewhere.
Starting with existing code saves time. Existing code has already been tested.
A developer should include a version number and date. The developer must define the scope of the work.
A technically more accurate AI process can still be a worse system if it wastes disproportionate human time, money, attention, or momentum.