Human-centered AI delivery — field workbook

Role and task:

Recent concrete example:

Trigger → information → decision → action → receiving owner:

Exception path and workaround:

Vocabulary (term / business meaning / authoritative source / common misunderstanding):

Hypothesis: for [role] doing [task], [change] should improve [outcome] while preserving [boundary]. Compare it with a simpler process/interface change and with no change.

Draw six states: proposed, awaiting review, authorized, executing, confirmed, unresolved. For each, state what has happened, what has not happened, what evidence the person sees and their next permitted action.

Include missing information, unavailable tools and a lost acknowledgement. Approval must refer to an exact proposed action; the application revalidates current authority when executing. Ask a peer to explain each state without your narration. Record their actual interpretation before revising the design.

Task and tested environment:

• Keyboard: reachable controls, logical order, visible focus, no trap.
• Meaning: text labels, accessible names, status beyond color, equivalent diagram explanation.
• Layout: narrow viewport, readable text, no loss of required controls.
• Failure: validation errors, unavailable dependency, unknown outcome, safe preservation of work.
• Assistive technology: record what was actually exercised and what remains untested.
• Motion and time: reduced-motion behavior and deliberate time-limit policy.

For every barrier: reproduction steps → effect on task → correction → affected recheck. An automated scan is useful evidence but not a complete accessibility assessment or conformance claim.

Accepted useful outcome:

Comparable baseline and measurement window:

Case mix and staffing differences:

All effort: creation + review + correction + exceptions + support.

All relevant costs: model/tool attempts + infrastructure + human work + allocated fixed costs.

Allowed scope, receiving owner and support route:

Stop, redesign and expansion criteria:

What evidence would make us reject our original hypothesis?

Use the evaluation workbook for arithmetic. Synthetic arithmetic does not establish customer savings.

Decision needed and accountable owner:

Context and constraints:

Recommendation and rationale:

Actual authorization state (not yet decided / approved scope / rejected):

Observable revisit trigger:

Carry one bounded project through the six curriculum stages. Include:

1. Workflow and vocabulary grounded in observation or clearly synthetic practice.
2. Runnable implementation, meaningful checks and reviewable source change.
3. Data authority, retrieval and integration contracts.
4. Bounded model/agent behavior and versioned evaluation.
5. Security controls, selected infrastructure and full-cost assumptions.
6. User-centered reliability, recovery and accessible human interaction.
7. Domain-specific interpretation and explicit prohibited decisions.
8. A receiving owner who can explain and operate the demonstrated scope.

Choose one cloud deeply; compare responsibility boundaries in the others. Choose healthcare administrative operations or financial-services operations for the supplied industry practice. This does not confer clinical, financial, compliance or vendor certification authority.

Changed-condition review

The reviewer selects a safe change: revoked access, duplicate event, stale source, missing permission, incorrect model answer, incompatible data or overloaded queue.

Use this with an available peer or mentor. The workbook does not schedule a mentor, issue a qualification or establish field competence by itself. A real engagement adds authorized data, environment, domain supervision and accountable acceptance.

References: OpenAI FDE work, Nielsen Norman Group mental models, WCAG 2.2 quick reference, FinOps unit economics. Checked 2026-10-23.