Back to course overview
Module 5Evaluating output 10 min

Human-in-the-loop

Designing checkpoints so review happens where it's effective — and staying sharp against automation complacency as trust grows.

'A human reviews it' is the safety story behind every workplace AI deployment. It's only true if the review is designed — placed where a human can actually catch what matters, not stapled on at the end where errors are polished and invisible.

Checkpoint placement beats checkpoint count

Recall Sam's workflow: extract facts → human reviews the facts → format the summary → human skims and sends. The review sits at the facts stage, where errors are raw bullets that are easy to spot — not at the polished-prose stage, where a wrong number wears a suit. The principle: review upstream, where errors are naked; approve downstream, where they're dressed.

  • Review the inputs when garbage-in is the risk (did the paste truncate? right week's data?).
  • Review the extraction when facts matter (bullets, tables, claim lists — fast to scan, hard for errors to hide).
  • Review the judgment always — any recommendation, ranking, or decision the output implies is yours to make, not confirm.
  • Skim the final for tone and constraint compliance only. If you're catching factual errors at the final stage, a checkpoint upstream is missing.

Automation complacency: the trap that opens after success

The dangerous phase isn't week one — you check everything in week one. It's month three, after fifty flawless runs, when your 'review' has quietly become a scroll-and-approve. Complacency isn't a character flaw; it's what human attention does around reliable automation. You counter it with structure, not willpower:

  1. 1Make review active, not passive. The reviewer must produce something — tick a 3-item checklist, correct one field, initial the verified number. Rubber stamps atrophy; tasks don't.
  2. 2Spot-audit on a schedule. Every tenth run (calendar it), verify one output fully as if it were week one. Log what you find.
  3. 3Track the catch rate. If your audits find zero issues for months, either the workflow earned lighter review (fine — decide that explicitly) or the audits went soft. Decide which on evidence, not vibes.
The intern standard, revisited

You'd never let even a brilliant intern send client emails unreviewed in week one — but you also wouldn't still be re-checking every comma in month six. Graduated trust with periodic audits is how you treat people; it's exactly how to treat the model. What you never delegate to either: your judgment and your name.