Session 4
Thursday, 16 July 2026 · 1 hour
The recurring AI Office Hours event covers every session — adding it once gets you all of them.
Why no Session on 2026-07-02
We skipped the July 2 slot due to the July 4 holiday. Session 4 lands on July 16; the every-other-Thursday cadence resumes from there.
Fourth AI Office Hours. R&D demo (moved from Session 3), a Stuck point working session where an associate brings a real in-progress problem for the room to debug, walk the latest Solution Library and Ideas queue updates, walk the next AI Brief digest.
Run of show
| Time | Beat |
|---|---|
| 0:00 | Welcome + recap — what happened since Session 3, what's new on the hub |
| 0:03 | Review the Solution Library + Ideas queue — new entries, status changes, what's been triaged |
| 0:18 | Associate demo — confirmed (moved from Session 3) — Tish Sherrick (Research & Development) on the Product Knowledge GEMs: five custom Gemini GEMs covering label/SDS/PDP synthesis, automated old-vs-new label delta reports, UX-optimized articles using Cognitive Load Theory, brand-voice alignment, and NotebookLM-driven training-content acceleration |
| 0:28 | Stuck point — Mary Rind's Gem — Mary (Supply Chain) shares an in-progress Gem and the snag she's stuck on. Working session: room weighs in with suggestions, parallels, fixes |
| 0:38 | Latest in enterprise AI — walk the top AI Brief pick from the new fortnight digest |
| 0:43 | Open mic / Q&A |
| 0:58 | Topics for next session — what to cover in Session 5 |
| 1:00 | Hard stop |
Pre-read (optional)
- How to Participate — ground rules and norms
- Solution Library — what's been added since 2026-06-18
- Ideas — current queue and triage notes
After the session
Q&A distilled from what came up live — the takeaways most relevant to associates.
What did Tish Sherrick demo, and what problem does it solve?
Tish (R&D Knowledge Management) walked through four custom Gems her team built to rewrite a library of ~35,000 knowledge articles — down from 111,000 — into a format that AI agents and Salesforce Agentforce can actually consume. The New Knowledge Gem takes a product label, SDS, reason codes, and product URL and outputs a brand-formatted HTML article ready to paste into Salesforce. The Label Version Audit Gem compares old and new label versions and flags every text change, ignoring image or layout noise. The Quick Text Creator takes a knowledge article and drafts pre-written email and chat blurbs for agent flows in Salesforce. A NotebookLM training script rounds it out — used to generate short video content that trains team members on how to use the Gems. The throughline: these tools help two people work through a massive, tedious task by handling the first draft, while humans stay in the loop for review.
Human review came up directly during the demo — how should I think about that?
A live observation worth flagging: even when a Gem is instructed to pull strictly from an approved label, domain experts need to review the output before it's published. In Tish's demo, a specific term — "feedings" on a pesticide product — was flagged as language that should read "applications" for regulatory reasons, even though it came directly from the label itself. The broader takeaway: the label has already gone through regulatory review, but AI-generated content surfaces what's in the source material, not what should supersede it. Tish confirmed that review is a formal step in her team's process — skipped during the demo for time. The general principle holds across every use case: the human who presses send is accountable.
How is Scotts tracking AI hallucinations?
It depends on the context. For consumer-facing CRM agents, outputs are measured by Sierra (our external partner), and logs are audited manually when issues arise. For Bloom agents (internal enterprise builds), the CoE uses rules-based sandboxes and multi-agent orchestration to keep each agent scoped to specific tasks, with internal tracking for hallucinations. For general Gemini use across the company, there is no systemic audit — we rely on associates to catch issues and report them. Three practices that reduce hallucinations in Gemini: be explicit about which data sources it should use, keep the task scope narrow, and if you're seeing a pattern of bad outputs, contact the CoE immediately — that's a conversation they'll escalate to Google.
Mary Rind's Stuck Point: how do you get Gemini to output real Google Slides?
Mary (Supply Chain) shared a Gem she built to generate slides in the SMG template — but Gemini kept outputting HTML instead of actual slides. The fix: switch to Canvas mode. Click the + next to "Ask Gemini," go to More Tools, and select Canvas. When Canvas generates a deck, an "Export to Slides" button appears automatically — no more HTML workarounds. One important caveat: once Canvas is on in a conversation, it stays on. If you need to do something else, start a new chat. Better yet, in your Gem settings under "Default Tool," set Canvas as the default so every session opens in it automatically.
Rob shared a pre-built SMG Slide Deck Gem configured with the Scotts template and Canvas as the default — try it before building your own.
How do you get better results when uploading documents into a Gem?
Breaking large documents into smaller, labeled files tends to help more than uploading one big file. For the SMG slide template, separating font guidelines into a dedicated "brand fonts" doc and legal labeling requirements into their own doc — rather than leaving everything in the master template — led to noticeably better outputs. Labeling speaker notes in the template to describe what each slide type is (e.g., "this is a presentation title slide") also helped the AI categorize correctly. That said, don't fragment too aggressively — 50 separate docs creates its own confusion. Test how you cut up the information and find what the model responds to best.
Is NotebookLM a better fit than Gems for slide creation?
For use cases where you're synthesizing your own documents into a presentation, potentially yes. NotebookLM restricts its output strictly to what you upload — it doesn't pull from the web or add anything external, which eliminates a major source of hallucination. It also has built-in slide creation and lets you iterate directly in the tool. If the content is already defined and you just need to shape it into slides, NotebookLM is worth trying before building a Gem.
Is there a summary of AI tools available at Scotts and what each one is best for?
Not yet — but it's coming. This came up live as a clear gap: with NotebookLM, Gemini, Gems, Canvas, Bloom, and more in play, associates need a starting point for choosing the right tool. Rob committed to publishing a one-pager.
When's the next session?
Thursday, 2026-07-30. Same recurring calendar event — add it once from any session page and you've got the full series.