9. AI in People Work

AI-fy, augment, keep human

The decision framework

Use when

Every 'should AI do this' question in a people process

AI-fy, augment, keep human, shown as a grid of 3 parts: AI-fy, Augment, Keep human.

AI-fy

Low judgment, low empathy load, low error cost, decent data: automate the admin end-to-end, human audits

Augment

Valuable but judgment-bearing: AI drafts, assembles, themes; a named human decides and owns

Keep human

High empathy, high stakes, or consultation-sensitive: humans only, stated publicly

Assess each workflow on judgment, empathy, error cost, data readiness and sensitivity.

The decision path

  1. 01Is the task legally sensitive, consultation-sensitive, high-empathy or high-consequence? Keep the decision human. AI may assist only within explicit boundaries.
  2. 02Is the work rules-based, reversible, low-sensitivity and supported by reliable data? AI-fy it, with monitoring and a named owner.
  3. 03Everything between those points belongs in Augment. AI can draft, assemble or identify patterns. A named human decides and owns the outcome.

Move through it with these

Judgment
How much judgment does the task require?
Empathy
Is empathy part of the task, or only part of how the outcome is told?
Error cost
Who carries the consequence if it is wrong, and can it be reversed?
Data readiness
Is the data accurate, representative and permitted for this use?
Sensitivity
Does it touch pay, jobs, rights, privacy or a consultation duty?

How to use it

The public boundary is the trust move. Name the decisions AI will not make: discipline, ratings, hiring, grievances and accommodation. AI may assist with administration or preparation only where that use is lawful, governed and clearly bounded. A named human decides and owns the outcome.

Questions to ask

  • How much judgment does the task require?
  • Who experiences the consequence if it is wrong?
  • Is empathy part of the task or only part of how the outcome is communicated?
  • Is the data accurate, representative and permitted for this use?
  • Does the task affect someone's job, pay, opportunity, reputation, privacy or rights?
  • What decision must a named human continue to own?
  • How will the use be monitored and reviewed?

The output

A one-page decision record naming the chosen path, the permitted AI role, the accountable human, the controls, the public boundary and the review date.

Where it breaks

  • Treating technical capability as permission
  • Calling a person “human in the loop” when they have no real authority
  • Using weak or unrepresentative data
  • Automating the sensitive decision while keeping only the communication human
  • Failing to tell employees where AI is used
  • Treating the first classification as permanent

Card details

Model type
Grid
Origin status
Synthesis
Content last reviewed
28 July 2026

Origin

Synthesis

Automate, augment, keep human is common vocabulary in AI practice. What is specific here is the five-way scoring (judgment, empathy, error cost, data readiness, sensitivity) and the public never-AI list as a trust move. Assembled by Harsimran Kaur Kapoor, 2026.

Neighbouring models

NextGovernance gates for AI touching employees

Was this useful?