Kyle Branchesi Email

Work / AI course production

Course production where AI agents draft and people decide what ships

At United Way Worldwide, expert review was the slowest part of building a course. Faster drafting alone would only have moved the pile-up to review, which is why I redesigned both sides. The AI behind these agents is the same one anyone can buy. What I designed is everything around it. Each program got its own agent, working from the organization’s frameworks and the expert’s own words. Before the expert sees a draft, AI checks review it the way that expert would. Then every expert review goes back into the agent’s instructions, so the next draft starts closer.

AI course production. United Way Worldwide, 2026. I set how the agents operate and who can say no to what they draft.

Review

A checkpoint only counts if the reviewer can say no

Two people can stop any course

Saying no takes the subject knowledge, the source, enough time and the authority to reject. I gave that authority to the subject expert and to the learning team’s quality check. Nothing publishes without both.

The experts’ side got redesigned too

I gave it the same design attention as the agents: the quality rubrics and review guidelines the experts work from, and the authority to send work back.

Why not a generic chatbot

Context

A chatbot knows nothing about the organization. Each agent worked from its frameworks, the learner personas and the expert’s own words, one agent per program.

Voice

Generic output sounds generic. These drafts sounded like the expert, so over time skeptical experts made fewer edits to each first draft.

Review inside the process

First AI checks, set up to review the way the expert would. Then the expert, then the learning team. Nothing shipped without people approving.

Each round makes the next one easier

A chatbot starts from zero every time. Here, each round of expert feedback went back into the agent’s instructions, so the next draft needed less of their time.

Five levels

Five levels of working with AI, and where this work sits

Many organizations use AI only at the first two levels: a question to a chatbot, or an assistant with standing instructions. The gains start higher up, when AI works through a task, checks its own output, and gets better with each round. The person stays in charge of what ships.

Five levels of working with AI

How much should AI do on its own? Each level up, it does more of the work before a person steps in. Pick the level by how often the task comes up and whether the result leaves your team. Then give it more room one permission at a time, and only once testing shows it’s ready.

  1. Level 1

    Prompting

    One question, one answer.

  2. Level 2

    Reusable assistant

    Standing instructions and sources, used again and again.

  3. Level 3

    Workflow

    Repeatable steps, chained together.

  4. 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.

  5. Level 5

    Agent with a feedback loop

    AI checks first, then people who can say no. Each round improves the next.

    People can always say no

  • Anyone can buy the same AI. The difference is what you give it to work from, who reviews its work, and whether each review makes the next draft better.
  • Every assistant or agent needs a named owner, proof from testing that it works, and someone to keep it up to date.
  • Don’t rely on telling an agent what not to do. Set up the software so it can’t.
One of four questions to answer before you put AI to work.

The loop

How a course moves through the loop

Amber is the agent’s work. Blue marks where a person decides. The loop is why the expert’s time kept falling.

  1. The expert talks, not types. After each meeting, the agent takes the transcript or AI notes and updates the draft to match what the expert said.
  2. The agent fills the gaps. Sections the expert hasn’t written yet, including their own specialty, get a first draft, so they react instead of starting from a blank page.
  3. AI checks go first. They review the draft the way that expert would, before a person spends a minute on it.
  4. The expert decides. They can reject or send back any draft. Every review goes back into the agent’s instructions and examples, so the next draft is closer to their voice. The AI itself never changed; the instructions it works from did.
  5. Quality check, then publish. The learning team checks each course against a rubric and acceptance criteria. Experts started skeptical, and their edits to first drafts fell over time.

These agents ran as a chain: one finished, the next started. I later built the orchestrated version, where a lead agent reads the results and decides what happens next. See the Test Class.

Try it

Review an AI draft

You’re the subject expert. An agent drafted this page of a course on sharing data with partners, and it doesn’t ship until you approve it.

Based onThis is the expert review step in the course workflow. Every draft goes to someone like you, who can reject it or send it back.

Agent draft

Module 2. Sharing data with partners

I wrote this example page and its source for the site.

Results

67% less time per course

Each pair is drawn to scale, across 11 courses built in 2026.

Related

Back to all work