How I think / AI enablement plan
Taking about 12,000 people from having AI tools to using them well
Staff across United Way Worldwide and its network had AI tools before they had guidance. I designed a plan in seven parts. It starts with 90 days of finding out what works, before anything goes wider.
Designed in 2026. IT and the Office of General Counsel approved the use guide, and the AI Foundations course was built. The other parts are designed and not yet rolled out. Examples marked as samples are made up to show how a part works.
The plan
The first 90 days
Rules come first, so people know what’s allowed. Then people practice on their own work, while the team tests what’s worth keeping. In week 13, an executive sponsor weighs the evidence and decides what happens next.
The first 90 days
How do you get a whole organization from having AI tools to using them well? Rules come first, then practice on real work, then a test of what’s worth keeping, then a decision based on the evidence.
Weeks 1 to 4
Set the rules
Confirm the approved tools, publish the Green, Yellow, Red guide, brief managers and measure where things stand today.
You getRules people can follow, and a starting point to measure from.
Weeks 5 to 8
Build skill on real work
Pilot the AI Foundations course, run workflow labs with two teams, and ask local United Ways what they already do with AI.
You getStaff practicing on their own tasks, and pilot plans.
Weeks 9 to 12
Test and learn
Test the course coach with a set of real questions, check at 30 and 60 days whether people work differently, and share the first write-ups of what’s already working in the network.
You getEvidence, good and bad.
Week 13
Decide
An executive sponsor weighs the evidence and decides, part by part: scale it, keep testing, change it or stop it.
You getA decision for each part, and the next 90 days.
People liking a part isn’t enough to roll it out. It goes wider only when four things are true: learning has been measured, behavior has changed, there have been no serious incidents, and safe work isn’t being blocked more than the target allows.
The parts
Seven parts, each for a different problem
People skip AI tools, or misuse them, for different reasons. Each part of the plan deals with one of them. Open any part to see a short sample from its design.
1
Green, Yellow, Red use guide
ApprovedOwner: learning team, with IT and legal
Answers “Can I use AI for this?” in under a minute, with two questions: what am I giving the tool, and what happens if the answer is wrong?
See a sample: six tasks, sorted
- GreenDraft social posts for a public event, from the published event page. Check dates, links and tone before posting.
- GreenExplain a spreadsheet formula using made-up sample data. Test it on the real data yourself.
- YellowDraft a grant narrative from an internal program plan. Approved tool only. Verify every number and check the funder’s AI policy.
- YellowSummarize a public regulation for an internal briefing. Public, but costly if wrong, so legal reviews it before anyone relies on it.
- RedPaste a client’s case notes to draft a referral letter. Never in a general tool. Follow client-confidentiality rules.
- RedRank donors by giving potential. A decision about people, using confidential data. It needs formal approval and a bias review.
Not sure? Pause, ask your manager the same day, then the AI help channel. Made a mistake? Report it within a day. Reporting is protected; hiding it is not.
2
AI Foundations course
Designed and builtOwner: learning team
A short course for staff with no technical background: use AI on their own tasks, check what it produces, and know when to stop.
See a sample: the checking card
Before anything AI drafted leaves your hands, check:
- Values and numbersagainst the source
- Entitiesnames, titles and organizations
- Referencesopen every link and citation
- Instructionsdid it do what you asked?
- Fittone, reading level and language for the audience
- You own itwould you sign it?
The final practice: pick one real task, sort it with the two questions, write the request, check the result with this card, and write two sentences on what you would and would not use AI for in that task.
3
Manager conversation guide
DesignedOwner: learning team, with HR
Managers set the tone for whether people feel safe talking about how they use AI. The guide helps them set expectations, review AI-assisted work fairly and talk through mistakes without blame.
See a sample: three questions and three scripts
When reviewing AI-assisted work, ask: What tier was this? What did you check? What are you unsure about?
- Opening the conversation“I’d rather hear ‘I tried AI and it got this wrong’ than have anyone hide what they’re doing.”
- A worried employee“It’s fair to worry about what this means for your role. Here’s what I know, what I don’t, and how I’ll keep you informed. Your judgment is the part AI can’t do.”
- Being too careful“This is a Green task, so go ahead. Being careful is good, but we also lose time when safe work waits.”
4
Workflow labs
DesignedOwner: each team lead, with the learning team facilitating
A team maps one real process, fixes what needs no AI, places human checkpoints and leaves with one small pilot, not five.
See a sample: a newsletter team’s lab
Sample only. A communications team’s monthly newsletter, from collecting updates to sending:
- YellowEdit submissions into the house style, and suggest headlines. The editor checks every fact and link.
- YellowFirst draft of the Spanish edition. A fluent reviewer approves it before it goes out.
- No AIChasing late updates. A submission form with a deadline fixes it.
The pilot: three newsletter cycles, measured against today’s editing time. It stops if a factual error reaches a sent newsletter twice.
5
Course coach
SpecifiedOwner: learning team and IT
An AI coach inside a course, kept to a narrow job. It answers only from the course’s approved materials, gives feedback on practice work and hands off to a person when it should.
See a sample: what it will and won’t do
It will
- Answer from the course materials, and name the module
- Comment on a draft against the course rubric
- Play a practice partner in a scripted scenario
- Answer in Spanish from reviewed Spanish materials
It won’t
- Search the web or chat about anything else
- Grade work that affects anyone’s job
- Accept names, case details or donor information
- Take actions in email, files or other systems
When the materials don’t cover it, it says “This course doesn’t cover that” and names a person to ask.
6
Shared library of what works
DesignedOwner: learning team, with local United Ways
In a network of independent organizations, what works is usually already happening somewhere. The library collects it, checks it and shares it so others can adapt it safely.
See a sample: how an entry earns trust
- ReportedDescribed by the local organization in a short interview.
- CheckedConfirmed by a second person who has seen the work.
- MeasuredA before-and-after number, with how it was counted.
- Safe to sharePassed a privacy and Green, Yellow, Red review.
Interview questions include: What problem were you solving? Where does a person check or decide? What changed, and how do you know? What would another organization need to copy it?
7
90-day learning agenda
DesignedOwner: the executive sponsor, with the learning team
Instead of a big launch, 90 days of structured learning, with the questions to answer, the evidence to collect and the rules for what to keep. It is the timeline at the top of this page.
See a sample: the decision rules
- ScaleLearning is measured, behavior improves, there are no serious incidents and safe work isn’t being blocked more than the target allows.
- ContinuePromising but mixed. Run another cycle with specific fixes.
- ChangeA clear gap, such as few managers taking part or confusion about tiers. Redesign that part.
- StopA serious data incident, no change in behavior after two cycles, or costs that outweigh the gains.