There is a line I think is worth watching right now.
AI is getting better at answering questions, yes. But the bigger shift is that it is getting better at carrying the work farther without you sitting there directing every single step.
That sounds wonderful right up until you remember that occasionally AI is confidently wrong, strangely literal, or operating with the judgment of a very fast intern who has never met your client.
So this week, I want to talk about both sides of that.
- What AI can do for you now.
- And the part I absolutely do not want you handing over.
ChatGPT is getting much better at doing the work, not just talking about it
GPT-6 Astra is built to finish things, not just answer things
OpenAI released GPT-6 Astra last week. Ignore the benchmark Olympics for a minute. Here is the part I think non-technical people will actually notice: Astra is designed to handle much longer, multi-step work across your computer and the web.
Think:
- Research something, compare the options, and organize the findings.
- Fill out repetitive forms.
- Work through a complicated website.
- Organize information across documents.
- Create a presentation or spreadsheet that follows your instructions.
- Carry a task through several steps instead of stopping after every one to ask what you want next.
OpenAI describes Astra as its strongest model yet for computer use and professional work, and it is rolling out across supported ChatGPT paid plans. Read OpenAI's GPT-6 Astra announcement.
Why do I care? Because we are moving away from “help me figure out how to do this” and toward “here is the outcome I need, work through it with me.”
“Help me figure out how to do this.”
“Here is the outcome I need. Work through it with me.”
That is a meaningful change. But it also changes our job.
The better AI gets at carrying out work, the more important it becomes to tell it what good looks like, where it needs to stop, what it cannot touch, and what you need to review before anything becomes final.
More capable does not mean less human judgment. Quite the opposite.
Your Copilot Notebook can now know what happened in the meeting and the email chain
Emails and meetings are now sources inside Copilot Notebooks
Microsoft has added Outlook emails and Teams meeting content as references inside Copilot Notebooks. That means a project notebook can include things such as:
- the files
- the emails
- the meeting transcript
- the meeting chat
- notes and shared content
And Copilot can use that combined context when helping you create a brief, summary, or presentation.
For anyone who has ever thought: “I know we decided this somewhere. I just have no idea whether it was in the meeting, the email after the meeting, the Teams chat during the meeting, or the document somebody sent three days later.” This is for you.
Microsoft lists both as available in Copilot Notebooks across Microsoft 365, though as always, what you can actually see depends on what your organization has turned on. See Microsoft's current Copilot release notes.
Why do I care? AI becomes much more useful when it can see the actual context around the work, not just the one document you remembered to upload.
Before you send it, make AI be your second set of eyes
We spend a lot of time asking AI to create. This week I want you to use it to challenge.
Take something that is almost finished:
- a proposal
- an important email
- a presentation
- a recommendation
- a meeting plan
- a contract summary
- a decision you are about to make
Give it to the AI you already use. Then ask it to find the things you are too close to see.
Act as my second set of eyes before I finalize this. First, tell me what you believe I am trying to accomplish and who this is for. Then review what I gave you and identify: 1. Anything unclear, contradictory, unsupported, or easy to misunderstand. 2. Important information or questions I may be missing. 3. Assumptions I appear to be making. 4. Anything the recipient may challenge, question, or interpret differently than I intend. 5. The three changes that would most improve the result. Separate facts from assumptions. Do not invent missing information. If you think something is wrong, explain why and point me to the exact part that concerns you. Do not rewrite the entire thing unless I ask you to. Finish with: - What I would fix before sending - What I would leave alone - What you may be forgetting Ask me only the questions that would materially change your review.
I love this move because it does not ask AI to replace your judgment. It gives your judgment another angle.
And once you learn it, use the same move almost everywhere.
Five more places I would use it
- Before an important meeting. Give it the agenda and background and ask what questions you are not asking.
- Before choosing between two vendors. Give it your criteria and comparison and ask where your reasoning is weak or incomplete.
- Before sending a difficult email. Ask what the recipient may hear that you did not mean to say.
- Before approving a presentation. Ask what an executive unfamiliar with the project will misunderstand in the first five minutes.
- Before making a personal decision. Travel, a major purchase, family logistics, a home project, a school decision. Ask it to separate what you know from what you are assuming and identify what you have not considered.
That is the pattern:
The AI can help with the first three. You still own the fourth.
Walk into the appointment prepared
Prepare for the doctor's appointment before you are sitting on the paper-covered table trying to remember everything. I have used this one recently for a couple of hamstring surgeries and it was a big time and brain saver, especially when post-surgery I could not think straight.
For eligible U.S. users, ChatGPT Health can connect information you choose from supported medical records and Apple Health and help you understand changes over time or prepare questions for an appointment. It is rolling out gradually. See how Health in ChatGPT works.
I would use AI here to help me organize, understand, and prepare questions. I would not ask it to replace my doctor or make the medical decision.
A useful prompt could be as simple as this:
Help me prepare for this appointment. Based only on the information I give you, summarize what has changed, what I should make sure I mention, and the questions I may want to ask. Separate what is in my records from anything you are inferring. Do not diagnose me.
You can take this a step further and build out a Project in ChatGPT or Claude so you can keep the thread going throughout a health journey.
That is a very good use of AI: help me walk into an important human conversation better prepared.
Have a Family Photo Detective Night
This one requires a shoebox, a phone, or the family member who somehow has every photograph ever taken.
Pick five old family photos. Not the perfectly labeled ones. The mysterious ones. Upload them to your AI and tell it:
We are doing Family Photo Detective Night. Study these photographs carefully, but do not invent facts about the people or events. Point out clues we may not have noticed: clothing, cars, signs, objects, architecture, hairstyles, technology, background details, weather, location clues, and anything wonderfully strange. Then give us: 1. Ten questions to ask the people in our family who might know the real story. 2. Your three best guesses about the decade or setting, clearly labeled as guesses and explaining the clues. 3. The funniest detail in each photograph that a younger person today might not understand. 4. A title for each photo as if it were an episode of a slightly dysfunctional family sitcom. When we give you the real answers, turn them into a short family memory we can save with the photo.
There are AI family-history tools doing increasingly sophisticated versions of this, but you do not need another app to try the basic idea. A multimodal AI chat can already help you inspect what is visible and generate the questions that get the humans talking. Here is one example of people doing it.
I particularly like the rule that the AI has to label its guesses. Because there is a fine line between:
- “That looks like 1978 because of the car and the clothing.”
- “Congratulations, ChatGPT has just invented an entirely new branch of your family.”
Let the machine find the clues. Let Grandma correct the record.
Get good at reviewing what came back
The better AI gets at doing things for us, the more important it becomes to get good at reviewing what came back.
Do not measure your AI skill by how much work you can hand off. Measure it by whether you know:
- what outcome you wanted
- what context the AI needed
- what it was allowed to do
- what still needs checking
- and where your judgment belongs
That is the human-in-the-loop part. And it is becoming more important, not less.
Many of you have attended the AI × CRE AI 101 session. On September 18 we are going one step further with a Build Your Own AI Chief of Staff workshop. Come build the thing that reviews your week with you.