How I think
Four ideas behind my work
Each one answers a question leaders ask before they put AI into real work.
1 · Governance
The two questions
What am I giving the tool, and what happens if the answer is wrong? Together they decide whether AI touches a task, and who has to approve it.
| If the answer is wrong ↑ | |||
|---|---|---|---|
| A decision about people or organizations | RedStop for formal approval | RedStop for formal approval | RedStop for formal approval |
| Costly to fix | YellowApproved tool, named owner | YellowApproved tool, named owner | RedStop for formal approval |
| Easy to catch | GreenUse it, check it | YellowApproved tool, named owner | RedStop for formal approval |
| Public | Internal | Personal or confidential | |
| What you give the tool → | |||
Green
Use it, and check the result.
For example: summarize a published public report.
Yellow
Approved tool, and a named person owns the draft.
For example: draft an update from an internal plan.
Red
Stop for formal approval.
For example: rank partners for investment from their public websites. The information is public, but it feeds a decision about organizations.
Formal approval starts at the blue line.
Work short of the line clears quickly, so the safe path is the fast path.
Over-caution counts as a failure too. The target: no more than 10% of safe tasks blocked.
See it at work:Assess your own taskThe AI curriculum and use guideAI Use Coach
2 · Expertise
Keep the expert in the loop
How does AI get better at the work without replacing the people who know it best? It drafts the standard; the expert on the ground changes it; the system keeps what they changed.
Step 1
AI drafts the standard version
The generic draft, from what it was given and nothing else.
Step 2
The expert changes it
The person who knows the place keeps, edits or rejects the draft.
Step 3
The system keeps what they changed
Their edits become guidance for the next draft.
Step 4
The next draft starts closer
So the expert has less to change each time, and the loop runs again.
The expert stays in charge at every round.
Where I’ve built it
Course production, United Way Worldwide
Agents drafted each course from the subject expert’s meetings. Every review went into the agent’s guidance, and over time the experts made fewer edits to each first draft.
Alf Studio
Alf drafts a unit; the teacher edits it in place. Alf notices the changes she keeps making, and none of it applies until she turns on “Start my drafts this way.”
Where else it fits: anywhere a standard plan meets a real place. Dispatch routes, inspection checklists, outreach plans, any operating procedure.
See it at work:AI course productionAlf StudioCourse Draft Agent
3 · Agents
Five levels of working with AI
How far should you let AI run on its own? Each level up, it does more of the work. Choose the level by how often the task repeats and whether the result leaves the team, then raise it one permission at a time, with test evidence.
Level 1
Prompting
One question, one answer.
Level 2
Reusable assistant
Standing instructions and sources, used again and again.
Level 3
Workflow
Repeatable steps, chained together.
Levels 4 and 5: where my course agents work
Level 4
Agent
Works through a task with some independence: reads inputs, edits drafts, fills gaps.
Level 5
Agent with a feedback loop
AI reviewers first, then people with a veto. Each round improves the next.
People keep the veto
- Any organization can buy the same models. The value is in the context they work from, the review built around them, and the feedback loop.
- Assistants and agents need a named owner, test evidence and a maintenance plan.
- Draw the line in the system, not in a prompt: agents get limits the system enforces, not just prompt wording.
See it at work:Where the course agents sitThe Agent Lab
4 · Workflow
Fix the process before adding AI
Where should AI go in a team’s work? Not into a broken step. Map the work, remove what needs no AI, then place AI where it fits and a person where mistakes can’t be undone.
No AI yet
Step 1
Map it
Map each step as it actually runs, including where work waits.
You getA map of today, with the waits marked.
Step 2
Fix it
Remove the waits and the steps that need no AI. Where rules or existing software already work, use them.
You getFixes that need no AI.
AI where it fits
Step 3
Rate it
Rate each step with the two questions. Put a person where mistakes leave the building or can’t be undone.
You getA redesigned map that a person approves.
Step 4
Pilot it
Measure a baseline first, then run it small, with conditions that stop it.
You getA pilot plan you can stop.
Drafting time saved isn’t organizational savings on its own. Measure the whole process: if work spends most of its time waiting, faster drafting may do little to shorten it.
See it at work:Workflow Redesign Agent