Session 6
Thursday, 13 August 2026 · 1 hour
The recurring AI Office Hours event covers every session — adding it once gets you all of them.
Sixth AI Office Hours. Jacob Howarth (Finance, RTR & Controls) demos how he rebuilt a manual SOX maintenance request process — 8 requests/month, 40 minutes each — into a validated Google Apps Script app that checks against live SAP master data, auto-routes approvals by region and object type, and closes each request to an immutable audit-trail PDF. He fed in existing files (the request template, SOX control language, FY25 testing workpaper, SAP master data extracts) rather than starting from prompts, and used Gemini for the bulk of the build and Claude for the sandbox and regression suite. The result is v4.1 live in production, ~48 hrs/year returned to the team, and a July walkthrough with Deloitte and Internal Audit — built entirely on tools every associate already has.
Run of show
| Time | Beat |
|---|---|
| 0:00 | Welcome + recap — what happened since Session 5, 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 — Jacob Howarth (Finance, RTR & Controls) on rebuilding a SOX maintenance workflow in Gemini + Apps Script — a finance manager with no developer background replaced a manual, free-text Google Doc process with a validated app that enforces SAP master data lookups, deterministic approver routing, immutable audit-trail PDFs, and automated reminders. ~48 hrs/year returned to the team; already through Deloitte and Internal Audit review |
| 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 7 |
| 1:00 | Hard stop |
Pre-read (optional)
- How to Participate — ground rules and norms
- Solution Library — what's been added since 2026-07-30
- Ideas — current queue and triage notes
After the session
Session recap
Opening — team project updates (0:00–0:12)
Rob opened with a quick status sweep: social listening / Bloom migration in progress; Roundup competitive-tracking tool underway; CRM brief writer moving; IVR migration from MIAW → Omni with a Blackmagic agent in the works. He flagged that the intake channel is switching from Monday.com to ServiceNow starting September.
Q&A
Q: What does the multi-agent pattern look like in practice — how many Gemini agents did Jacob actually use?
A: Jacob ran four roles — an architect (designs the data model), a prompt-writer (writes the prompts the other agents use), a code-writer, and a code-reviewer — iterated across 17 deployment cycles before going live. The segregation-of-duties rule is baked in: the system won't let a requester approve their own ticket, and for testing he used the email+ alias trick (e.g., [email protected]) to simulate a second approver without a second account.
Q: Did Jacob start from prompts, or did he feed in existing documents? A: He fed in files first — the existing request template, the SOX control language, the FY25 testing workpaper, and SAP master data extracts. That gave the agents real context and dramatically shortened the iteration cycle. He used Gemini for the primary build and Claude for the sandbox and regression suite.
Q: How did it hold up with Internal Audit and Deloitte? A: The July walkthrough covered immutable audit-trail PDFs, auto-routing logic, and the SAP master data validation. Both teams passed it. The key was that every decision in the workflow is deterministic and traceable — no free-text, no manual routing.
Q: Earl (field sales) — is there a way to look up grass seed label and lot information in the field? A: The team confirmed there are existing field-sales tools in the works. Rob asked Earl and Brian to connect with Travis Askew's sales-operations track, where there may be opportunity to surface lot-level data through the same surface rather than a one-off tool.
Q: How are dispatchers using AI? A: No dispatchers were on the call. Rob noted the team will follow up offline with Tina on current dispatcher AI usage.
Q: Heather (Canada) asked for a beginner on-ramp — sessions move fast and terms like "agent" are unfamiliar to many attendees. A: Rob acknowledged the feedback directly. Two things are coming: (1) one-on-one office hours slots with CoE team members will open soon; (2) a structured AI training track (beginner → advanced) is in progress through HR and FOSA. The interview format this session was a deliberate move toward showing process, not just polished output. Jacob also committed to sharing his initial "prompt-to-write-prompts" prompt so attendees can see the bootstrapping step.
Q: Thomas shared a practical beginner tip — what's a low-risk way to start? A: Thomas described setting up a Gemini Workspace Studio routine that emails him a morning digest of all agency emails (campaigns active, recent, and upcoming) — took about an hour to configure and runs at 7:30 a.m. daily. Aditya reinforced that Workspace Studio is the most underrated tool in the SMG Gemini stack and is already linked in the AI Hub under Gemini Tools. He recommended it as everyone's first automation.
Q: Can Jacob's email-summarization Workspace Studio be shared? A: Jacob dropped a link in chat. The template is set to allow copies, so attendees can clone it and configure it for their own inbox without touching Jacob's version.
AI Brief highlights
- Workspace Studio auto-adding relevant content to Gemini Notebooks — a new ambient capture feature worth watching.
- MCP going stateless/headless — Model Context Protocol is moving to a model where the server holds no session state. This is a significant security improvement for enterprise deployments.
- SAP ECC via MCP to Cursor — Scotts has an MCP server connecting SAP ECC data to the Cursor IDE currently in QA. Watch this space.
- Gemini hit 1 billion monthly active users; 70% of voice AI users are now on Gemini. Google Assistant is being deprecated and replaced by Gemini across all surfaces.