Skip to main content

Human approval and accountability

Automation that acts without accountability is a liability. The goal is not to remove humans — it is to remove toil while keeping a person responsible for anything that matters.

Decide what needs approval​

Not everything does. Route by stakes and confidence:

  • Auto-proceed when the action is low-stakes and confidence is high (e.g. sending a routine reminder).
  • Approve first when the action commits money, touches a customer relationship, or the model is uncertain (e.g. confirming a booking, sending a quote).
  • Escalate when rules conflict or an exception is detected.

Make approval a real decision, not a rubber stamp​

An approval step is only worth something if the human can actually judge it in seconds. Show them:

  • what the system proposes;
  • the evidence it used;
  • what it was unsure about, highlighted;
  • what will happen if they approve.

If your approval screen just says "Approve? Y/N" with no context, people will click yes blindly and the accountability is fictional.

Keep an audit trail​

Record who approved what, when, on what evidence, and what changed as a result. This is not bureaucracy — it is what makes the system defensible when something goes wrong, and it is often a selling point in itself.

Design for correction and rollback​

Humans will approve mistakes. Make them reversible where possible: a clear "undo", a correction path, and a record that the correction happened. A system that cannot be corrected will not be trusted with anything important.

Accountability stays human​

The model does not carry responsibility; a named person does. Say so explicitly in the product, especially in regulated contexts.