Warren J

Myth: a small team needs a big AI strategy before it can start. Fact: it needs one safe task, one person checking the work, and a clear stop rule. Start smaller than your anxiety.

A lot of teams are trying AI. Fewer have agreed rules for what belongs in a tool, who checks the answer, and what gets documented. This week I’m starting with one simple workflow card: task, data boundary, reviewer, quality check. What would you put on yours?

Weekend read: New York City’s AI policy is worth reading past the headline. It pairs a student-facing pause with literacy lessons, limited pilots, and a coalition that will study what happens next. That is a more useful debate than “AI: yes or no.” https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat

A five-minute check can prevent a very public mistake: read every word in an AI-made graphic out loud, verify the date, look at the contrast on a phone, and ask whether the image actually reflects the people you serve.

I tried to frame Google Pics as a draft partner, not a replacement designer. The useful part is making a few visual starting points inside Workspace. The non-negotiable part is checking the names, dates, text, and people shown before anything goes public.

Quick question for nonprofit and school leaders: if you had one page for your AI rules, what would be on it? My short list is approved uses, off-limits data, the human check, and who owns the final call.

A school can slow down on student-facing AI without pretending the technology will disappear. What would your school or youth program want to teach first: how AI works, how bias shows up, or how to check an answer? NYC’s new policy gives us a real case to discuss: https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat

Weekend read: AWS’s guide to guardrails at the tool boundary. You do not have to be a builder to use the idea. Ask: what can our AI read, change, send, or call, and where would we notice a wrong move? Read it here: https://aws.amazon.com/blogs/security/extend-amazon-bedrock-guardrails-to-tool-interactions-using-the-strands-agents-sdk/

Before connecting an AI tool to email, files, or finances, write down the actions it can take. Then remove the actions it does not need. Fewer permissions make mistakes easier to contain.

A safe way to test an AI finance feature: start with one low-risk task, keep permissions narrow, and make a person approve every change. If the tool cannot show why it made a match, pause the experiment.