Hey everybody,
Last Thursday I spoke at the Greater St. Louis Safety & Health Conference about using AI safely at work. I asked who had used AI that week. Plenty of hands went up. Then I asked whose organizations had clearly explained which tools people could use, how to use them, and why. The hands started coming down.
That little moment has stayed with me. We are getting remarkably good at giving people AI tools. We are less consistent about giving them the clarity and confidence to use those tools well.
And I get it. Most of us already have full-time jobs. Nobody needs a second one called “monitor the AI industry.”
When the work matters, don’t just ask AI for an answer. Design the assignment, set the boundaries, and keep the decision where it belongs.
The conversation has moved from “Can AI do this?” to “What should AI be allowed to do?”
A qualified human still belongs in the loop
On October 8, Anthropic published an update to its usage policy. Two points stand out. When AI affects someone’s health, legal rights, finances, livelihood, or access to essential services, a qualified human still needs to be in the loop, and the affected person needs to be told AI was used. A new requirement adds that AI connected to physical equipment that could cause injury needs a qualified operator who can observe it and stop it. The updated policy takes effect November 12. Read Anthropic’s October 8 policy update.
Instructions matter. Permissions, supervision, and testing matter too.
Then, on October 9, Anthropic published examples of unintended model actions observed during testing and internal use. In one case, a model asked to create sample tasks from web pages submitted a police tip form about an unsolved homicide. The tip was flagged as spam and never forwarded. This is a particularly useful reminder that instructions are important, but system permissions, supervision, and actual testing matter, too. Read Anthropic’s report on unintended model actions.
Dots: always-on agents working in the background
Meanwhile, OpenAI’s newly introduced Dots are designed as always-on agents that keep working on projects in the background, with their own cloud computers and app connections. For now, they are available on ChatGPT Pro and Business Premium plans. This is an exciting direction, and it raises a very ordinary question: Before an assistant works on my behalf, do I know what it can see, change, send, or approve? Read OpenAI’s Dots announcement.
The smarter these tools get, the more valuable it becomes to know where the human handoff happens. You do not need to know every model. You do need to know the limits of the assignment.
A cited answer is a starting point, not the finish line
Copilot may draw on work information you have permission to access, depending on your license, connected sources, and organization’s settings. But access is not the same as accuracy. Before forwarding a summary, open the underlying documents or links and confirm the important details yourself.
Try asking:
Use only the approved material I provide or you can access. Separate what the source explicitly says from your interpretation. Flag missing information and disagreements. Show me which source supports each important claim. Do not send or change anything.
And if your organization has not told you what is permitted, ask before uploading client information, employee records, contracts, or confidential material. Read Microsoft on Copilot data access and permissions.
My three-color approach: Stating, Sequencing & Specifying
In my conference demonstration, I gave Claude a synthetic safety-observation spreadsheet and asked it to produce five different business outputs. The impressive part was not five documents appearing on screen. It was how I gave the assignment enough direction to make those outputs reviewable.
I break a more complicated request into three moves. You can use the same approach in ChatGPT, Claude, Copilot, or Gemini, even if all you can do is paste text into a chat.
Tell AI what the job is, why it matters, what sources to use, and what it must not decide.
Ask it to identify evidence, uncertainty, and verification needs before producing polished deliverables.
Describe the output format where consistency matters. Leave room for judgment where it does not.
Here is a version you can actually use this week. Imagine you need to review a proposal, compare several options, or prepare a leadership update. Paste in only information you are permitted to share.
STATING Act as my analytical partner. I need to review [DESCRIBE THE PROPOSAL, DOCUMENT, OR DECISION] for [AUDIENCE/PURPOSE]. Use only the information I supply or explicitly approve. My goal is to identify what deserves attention, not to have you make the final decision. Separate facts from assumptions. Do not invent missing information, contact anyone, or take actions on my behalf. SEQUENCING First, tell me what the material clearly establishes, what is incomplete or inconsistent, and what I should verify before acting. Only then prepare your analysis and recommendation options. Identify any conclusion that depends on an assumption or a source I still need to check. SPECIFYING Give me: (1) a five-bullet leadership summary; (2) a short list of risks and open questions; and (3) an action tracker with Issue, Supporting evidence, Suggested owner, Verification needed, and Status. Mark suggested owners as proposals, not assignments. End with the three questions I should answer before deciding.
A well-written prompt cannot make an unapproved tool safe for confidential data, change its retention policy, or override your company’s permissions. Check those things before you paste.
Where else this works
- The proposal you have to compare before tomorrow’s call: ask for evidence, gaps, and decision criteria before a recommendation.
- The messy meeting notes: ask what was decided versus merely discussed before anyone gets an action item.
- The spreadsheet full of observations: surface patterns and data limitations before producing an executive summary.
- The family trip with three calendars and twelve opinions: identify conflicting dates and missing reservations before building an itinerary.
- The difficult conversation you need to prepare for: separate what you know, what you suspect, and what you need to ask.
Same move, different mess. You are not trying to make AI sound more confident. You are making its work easier to inspect.
Session 08: AI Without the Anxiety
This Friday, October 16, our monthly AI × CRE session is “AI Without the Anxiety: Security, Safety, and Smarter Use.” Anthony Peters, IT Systems Manager at BranchPattern, will go deeper on data boundaries, permissions, grounding, hallucinations, usage rights, and the part of every decision that still belongs to a human. We meet at 10:00 a.m. ET (9:00 a.m. CT) on Microsoft Teams. Bring the question you have been hesitant to ask. I suspect you will not be the only one asking it.
AI Without the Anxiety: Security, safety, and terminology for CRE decision-makers, with Anthony Peters.
Register »A polished chart is not evidence
I love a beautiful dashboard. In my safety demo, AI produced one from made-up observations, plus a leadership brief, a toolbox discussion guide, an employee message, and an action tracker. Very impressive. Also, because I invented the dataset for the exercise, absolutely none of that was evidence about a real facility.
A polished chart can look awfully authoritative for something that began life as a spreadsheet full of pretend problems. That is precisely why I made it synthetic, and precisely why verification comes before sending it to anyone who might act on it.
You don’t have to be the most technical person in the room
You do not have to be the most technical person in the room to use AI well. Be the person who knows what the work is for, what information belongs in the room, what needs checking, and when the call is yours. That is a skill worth taking from work into just about every other part of life.
I’ll keep sorting through the rest with you.