Check the count you're surest about. AI will lump together things that don't belong.
It ranked billing first: 25 complaints. Three about agents who didn't know the hotel's services, it tied to nothing. A spa closed Mondays. A car quoted at 5 minutes that takes 25.
1
33
Check the count you're surest about, at 5:22: youtu.be/N_35z66VXug?t=322
14
Gemini 4 Carbon opened to everyone Tuesday. Wednesday it lost on a customer's agent to Argon, out a month.
Same instructions, same question:
Flash 3:17
Argon 1:41
Carbon 7:18
Argon and Carbon cited every source right.
Pick the model with a stopwatch, not a release date.
93
Answer the questions you'll be judged on before anybody asks them.
Quarter by quarter, I'm measured by MBOs. I put mine in the same Gemini notebook as the supporting files, to make sure I was lined up.
Your rubric, your RFP, your audit list. Same notebook as your data.
1
59
The rubric, the RFP, my MBOs, all in one notebook, at 8:38: youtu.be/_bLOLlQw4yk?t=518
21
Before you build an agent, write the question it should answer "I don't know" to.
My Gemini Enterprise project manager failed three tests on day one. One sentence fixed it: put the source in brackets after every fact, or say you can't.
Ask that one first.
1
42
The build, the three day-one fails, and the one-sentence fix: linkedin.com/feed/update/urn…
8
Your AI adoption average hides the only two numbers worth watching.
Amazon's memo tells leaders to look at the 10th percentile and the 90th. Why the bottom teams aren't using it well. Whether the top teams' delivery went up to match.
Replace the average with those two lists.
1
49
Amazon's memo: aboutamazon.com/news/company…
14
IBM and Microsoft can now end an agent's access when its task ends. Neither can tell you what the task is.
Both ship short-lived, task-scoped access in place of standing access. Deciding the task stays with you.
Before the next agent goes live, write one sentence: its task.
1
47
IBM's Agent Identity announcement: ibm.com/new/announcements/ib…
19
