Kyle Branchesi Email

AI curriculum

Deciding when AI belongs in the work

Staff had AI tools before there was an AI policy, in a network where every organization sets its own rules. I designed a use guide for deciding when AI belongs in a task, and a ten-course curriculum that teaches it, from literacy to building agents.

AI Foundations and AI for Impact. United Way Worldwide. Informed by the NIST AI Risk Management Framework and OWASP guidance on AI agent security.

Try it

The use guide

Every task gets two questions: what am I giving the tool, and what happens if the answer is wrong? Sort eight everyday tasks the way the guide would, then check one of your own.

Based onThe Green, Yellow, Red use guide at the center of AI Foundations, the course I designed for United Way Worldwide staff.

How it works

  1. Read the task.
  2. Choose Green, Yellow or Red.
  3. Read where the guide puts it and why, then click Next task. Blocking a safe task counts as a failure too.

Task 1 of 8

Summarize a published industry report for your team

Choose where the guide would put this task.

Right 0 of 0. Safe tasks over-restricted 0.

Fig. 1 Where each task lands, by what you give the tool and what a wrong answer costs. Example tasks, written for this site.

Try it

Would you deploy this agent?

An agent acts on your behalf, so what matters is what it is allowed to do. This one drafts the monthly update to 200 partner organizations and sends it.

Based onThe Workflows and agents module of AI Foundations, where learners practice changing one agent permission at a time and rejecting a hidden instruction inside a document. The agent here is an example, written for this site.

How it works

  1. Turn off anything the job does not need, and check both boxes under Before it launches.
  2. Click Run the test. It checks two things: does a hidden instruction in one of six documents get out, and does a person see the update before 200 partners do?
  3. If it fails, change one setting and test again. When it passes, click Approve for launch.
Fig. 3 What the agent touches. Grey lines are permissions it does not have. Blue is a person deciding.

What the agent is allowed to do

  • The job needs this
  • The job needs this

Before it launches

Test evidence: not tested yet.

What happened

    Run the test to see what this agent does with these permissions.

    Agents need an owner, test evidence and a maintenance plan, and the system enforces their limits.

    Course

    AI Foundations

    The first course in the curriculum, designed for non-technical staff, and the one most staff would take. It is built to teach when AI belongs in a task, including the limits for agents that act on someone’s behalf.

    7 minStart with a real task
    12 minWhen a result earns trust
    12 minGreen, Yellow, Red
    9 minUse it, check it, own it
    9 minWorkflows and agents
    9 minFinal practice
    30 minOptional live clinic

    Fig. 4 58 minutes, self-paced, in six modules, plus an optional live clinic.

    Six design decisions

    1. It opens with a piece of real work, not a definition or a product tour.
    2. Every decision comes back to two questions: what am I giving the tool, and what happens if the answer is wrong.
    3. Agents get limits the system enforces, not just the wording of a prompt. Learners practice changing one agent permission at a time, and rejecting a hidden instruction inside a document.
    4. Six skeptic perspectives, including privacy, equity and job security, shaped the activities.
    5. No AI account is required. A learner can show good judgment by deciding not to use AI.
    6. One core course, with an adaptation kit each independent organization can choose to use.

    Curriculum

    AI for Impact

    Ten connected courses in four stages, from understanding AI to building assistants and agents with set limits. Each course starts from a real piece of work and has its own learning sequence, workflow, work product and assessment. Each course is designed to end with three things: something useful the learner made, a workflow they can repeat, and the judgment to know when AI should not be used.

    Not everyone is meant to take every course. Most staff would take AI Foundations and the studio for their own work, managers add the leadership course, and workflow owners and AI champions go further.

    Fig. 5 Ten courses in four stages. Each stage goes deeper and is meant for fewer people.

    1

    Understand

    Most staff

    One short course most staff would take: when AI belongs in a task, and when it does not.

    1AI FoundationsEnds with a real task placed with the use guide
    Starts from
    A task from the learner’s own week, not a definition of AI.
    The workflow
    Ask two questions of every task: what am I giving the tool, and what happens if the answer is wrong? Then place it Green, Yellow or Red, and check a result before trusting it.
    Format
    58 minutes, self-paced, in six modules, with an optional 30-minute live clinic.
    Where a person decides
    Deciding not to use AI counts as good judgment. No AI account is needed.

    Try the use guide

    2

    Apply to real work

    Each function; managers add the leadership course

    Five studios, one for each kind of work, and a course for managers. Each starts from a real piece of that work and ends with a reviewed work product.

    2AI Learning Design StudioEnds with a reviewed course outline
    Starts from
    A subject expert’s notes for a short course.
    The workflow
    Set what learners must be able to do. AI drafts the outline and practice questions from the notes. The designer checks every claim against the notes, and the expert signs off at set checkpoints.
    Work product
    A reviewed outline, one practice activity, and the prompt that drafted them.
    Where a person decides
    Nothing reaches learners until the expert signs off.
    3AI Communications StudioEnds with one story in three formats
    Starts from
    An approved quarterly report.
    The workflow
    AI drafts a newsletter story, a social post and a short talk track in the house voice. A person checks every fact, figure and quote against the report and edits for tone.
    Work product
    A set ready to publish, and a voice check the team can reuse.
    Where a person decides
    Someone’s own words or story are used only with their consent, and nothing is published unreviewed.
    4AI Research and Impact StudioEnds with a sourced brief
    Starts from
    Several public reports on the same question.
    The workflow
    AI summarizes and compares the sources, citing the page for each point. A person checks every figure against its source and writes the conclusion.
    Work product
    A two-page brief where every figure traces to a source.
    Where a person decides
    Raw survey responses with names stay out of general tools, and the conclusion is a person’s call.
    5AI Fundraising StudioEnds with a reviewed update and proposal section
    Starts from
    Approved materials about the organization’s programs.
    The workflow
    AI drafts an update and a proposal section from the approved materials. A person tailors them to the reader and checks every figure and promise.
    Work product
    An update and a proposal section ready to send, and a prompt the team can reuse.
    Where a person decides
    Personal records never go into a general tool, and every promise is checked, because the organization has to keep it.
    6AI Operations StudioEnds with a tested procedure and its owner
    Starts from
    A recurring process, such as a monthly report or an intake request.
    The workflow
    AI turns rough notes into a written procedure and a checklist. A person runs the procedure once, exactly as written, to find the step the draft missed.
    Work product
    A procedure anyone on the team can run, with a named owner.
    Where a person decides
    Employee and financial records stay out, and AI never approves a payment or a decision about a person.
    7Leading Human-Centered AIEnds with a team AI agreement
    Starts from
    The manager’s own team and its work.
    The workflow
    Map where AI helps the team and where it does not, agree what stays with people, and plan how to hear concerns about workload and jobs.
    Work product
    A one-page team agreement and a date to revisit it.
    Where a person decides
    AI output is never used to judge a person’s performance.

    3

    Redesign the process

    Workflow owners

    For the people who own a workflow: change the work, not just one task.

    8Systems Thinking with AIEnds with a before-and-after workflow map
    Starts from
    One whole workflow, start to finish, with the people who do it.
    The workflow
    Map every hand-off and where work waits. Decide, step by step, whether AI, a person or a simple rule fits best. Measure the current time before changing anything.
    Work product
    A redesigned workflow with an owner for each step and a baseline to measure against.
    Where a person decides
    The people who do the work review the redesign before it changes.

    4

    Build

    AI champions

    For AI champions: build assistants and agents with limits the system enforces.

    9Build a Useful AI AssistantEnds with a tested assistant and its owner
    Starts from
    A question the team answers again and again.
    The workflow
    Write the assistant’s instructions, limit it to approved sources, then test it with real questions and with questions meant to trip it up.
    Work product
    An assistant with a named owner, its test results and a maintenance plan.
    Where a person decides
    Finishing the course does not launch the assistant. It goes live only after formal approval.
    10Responsible AI AgentsEnds with an agent design, its permissions and its tests
    Starts from
    A multi-step task, such as gathering updates and drafting a summary.
    The workflow
    Grant permissions one at a time, enforced by the system rather than requested in a prompt. Test whether a hidden instruction inside a document gets through.
    Work product
    An agent design with its permissions, a test log and an approval step.
    Where a person decides
    Nothing is sent out without a person’s approval.

    Try the AI Foundations version of this test

    Fig. 6 Every course in the curriculum follows the same six steps, and ends with a reviewed piece of real work.

    Try it

    Map the work

    Before AI touches a piece of work, the work gets mapped: which parts AI can draft, which stay with a person, and which have to change first, because the information cannot go into the tool or the step should not exist at all.

    Based onThe second of the six steps every AI for Impact course follows (Fig. 6): mapping the work before AI touches it. The examples are written for this site, one for each of five courses above.

    How it works

    1. Pick a tab. Each one is a course from the curriculum above.
    2. For each of the five parts, choose AI drafts, A person, or Change it if the part has to change before it can happen.
    3. Click Check against the course to see how the course handles each part, and why.

    The work

      Dashed square: AI drafts it, and it stays a proposal until a person checks it. Blue square: a person does it. Crossed square: that part has to change first.

      0 of 5 parts decided

      Rules the curriculum holds to

      1. Completing a course demonstrates learning. It does not grant permission to use a tool, upload information, or launch an assistant or agent.
      2. Assistants and agents need a named owner, test evidence and a maintenance plan before anyone relies on them.
      3. In the agents course, permissions are enforced by the system, not just requested in a prompt, and nothing is sent out without human approval.
      4. Targets are set after measuring a baseline. Time saved on drafting is not reported as organizational savings on its own.

      Next

      Leadership programs for executives and new CEOs