Skip to content

Session 2

Thursday, 4 June 2026 · 1 hour

Add to Google Calendar

The recurring AI Office Hours event covers every session — adding it once gets you all of them.

Second AI Office Hours. Follow-up from the kickoff: review what landed in the Solution Library and the Ideas queue, two associate demos, walk the next AI Brief digest.

Run of show

Time Beat
0:00 Welcome + recap — what happened since Session 1, what's new on the hub
0:03 Review the Solution Library + Ideas queue — new entries, status changes, what's been triaged
0:13 Associate demo — confirmed (moved from Session 1) — Terri Weaver (Marketing) on the Tri-Engine AI Operations Framework: Gemini Workspace + API + NotebookLM running brand-manager workflows across macro-strategic, micro-tactical, and operational-governance horizons
0:23 Associate demo — confirmed — Caitlin Dezso (Marketing) on the Creative Brief writing Gem: brand background + brief template + desired outcomes → AI-completed creative brief, ready to hand to creative teams or agencies
0:33 Spotlight — Trusting AI outputs — when AI confidently fabricates numbers. How to spot it, validate, and steer AI in the right direction
0:43 Latest in enterprise AI — walk the top AI Brief pick from the new fortnight digest
0:48 Open mic / Q&A
0:58 Topics for next session — what to cover in Session 3
1:00 Hard stop

Pre-read (optional)

For the Spotlight on Trusting AI outputs:

After the session

Q&A distilled from what came up live — the takeaways most relevant to associates.

How do I keep AI from inventing numbers in my research?

Use grounded tools, not free-form chat. NotebookLM restricts answers to documents you upload and cites every claim back to a source passage. Deep Research (in both Gemini and NotebookLM) browses live sources and produces reports with inline citations. Reach for these for research tasks — they're built to keep you accountable to evidence, which is the prevention for the kind of fabricated-numbers case that anchored the Spotlight.

Can I use Gemini in Drive to search across both files and email?

Yes — Gmail as a source in Ask Gemini in Drive went GA on 2026-06-03. You can ground Drive-side Gemini answers in a combined view of email + files + folders and export the findings to a Google Doc. Effectively brings NotebookLM-style grounding into the Drive UX.

What's a "Gem" and how is it different from an "agent"?

A Gem is a custom, prompt-trained Gemini instance you can schedule to run autonomously — same thing most people mean by "agent" today. Terri's daily-business Gem runs at 9 AM, scans Gmail + Tableau, and drafts the morning report; Caitlin's creative-brief Gem acts as a creative lead for the Arrow Garden brand. Gems plug into Workspace connectors (Calendar, Tasks, Gmail, Docs, Drive) out of the box.

Should I use Gemini Flash or a heavier model?

Depends on the task. Flash is fast and inexpensive — fine for everyday summarization. For higher-quality output where depth matters (synthesis, brand voice, multi-step reasoning), switch to a more capable model. Worth checking the model selector on any Gem you build.

How do I audit a Gem's output for hallucinations?

Continuous monitoring, with one nuance: most "wrong" Gem outputs in production aren't hallucinations — they're source-data issues. Terri's experience: anomalies in her daily report almost always trace back to Tableau not reflecting current inventory, not the AI inventing numbers. Tighten the data pipeline before tightening the model.

Why does sharing a Gem's output feel clunky?

Auto-export from a Gem chat to a Google Doc is still rough — today it's mostly copy-paste. The Drive-side Gemini + Gmail-source workflow above is the closest fix: if you start in Drive instead of the Gemini app, your output lands in Drive in the first place.

We keep re-teaching every Gem the same business definitions (fiscal week, weekend). Is there a fix coming?

This came up live — re-training the same metadata into every new Gem is wasted effort. The CoE is exploring a centralized "gold standard" of assumptions that Gems can reference, so business logic gets defined once and reused.

What's the plan on AI-generated fraud (fake images, fake receipts)?

Two-track. Long term, Synth ID — a watermark Google + NVIDIA are embedding into AI-generated assets — becomes the standard once adoption broadens. Short term, the CoE is building a rules-based check (spelling errors, incorrect product forms, color anomalies) to flag suspect submissions until the watermark standard is widely deployed. See the AI Fraud Detection intake card.

When's the next session?

Thursday, 2026-06-18. Same recurring calendar event. Gem show-and-tells are queued up; Workspace Flows will get a walkthrough.