# AI @ Justrite — The Roundup <strong>Week 28 / July 2026</strong>
> <em>Everything worth knowing this week, in one place.</em>
---
## 📰 AI Headlines
- <strong>OpenAI Launches ChatGPT Work — The Enterprise "Super App" Arrives</strong> OpenAI released ChatGPT Work on July 9, a dedicated productivity layer built on GPT-5.6 for Enterprise, Pro, and Edu plans. The move targets workplace productivity directly — competing with Anthropic's Claude for enterprise seats. The pricing model is per-seat, and the feature set includes persistent project memory, scheduled task automation, and cross-tool integrations. Why it matters: Your team already has ChatGPT Enterprise licenses. This upgrade turns it from a chat window into something that can run recurring workflows. <a href="https://www.reuters.com">Reuters</a>
- <strong>MUFG Bets Big on AI-Native Banking — Partners with OpenAI</strong> Mitsubishi UFJ Financial Group, Japan's largest bank, announced a major partnership with OpenAI to become an "AI-native" institution. The deal covers back-office automation, fraud detection, and internal tooling. MUFG joins a growing list of financial institutions — Goldman Sachs, JPMorgan, Morgan Stanley — that have signed multi-year enterprise AI deals in 2026. Why it matters: When banks move, manufacturing follows. Finance departments inside industrial companies are next. <a href="https://openai.com/blog">OpenAI Blog</a>
- <strong>Industrial Robotics Becomes Physical AI's Proving Ground</strong> SiliconANGLE reported July 2 that industrial robotics — not chatbots — is where "physical AI" is proving real ROI. Companies are deploying vision-language models directly on factory floors for defect detection, adaptive assembly, and autonomous material handling. NVIDIA's Isaac platform and new edge inference hardware are enabling this shift. Why it matters: The gap between "AI hype" and "AI on the factory floor" is closing — and the winners are companies that test now, not later. <a href="https://siliconangle.com">SiliconANGLE</a>
---
## 🔧 Prompt of the Week
``` You are a senior quality engineer with 15 years in industrial manufacturing.
I'm giving you the last 50 defect reports from our production line. For each defect category, provide:
1. Root cause hypothesis (1 sentence) 2. Pattern you notice across the batch (1 sentence) 3. Recommended corrective action ranked by impact (1-3)
Before analyzing, think step by step about what data points would matter most for a quality investigation in this type of manufacturing. ```
<strong>When to use it:</strong> When you have a CSV of defect data from the shop floor and need a structured quality analysis — without waiting for the monthly review meeting.
---
## 🎥 Watch This (5 min)
<strong>NVIDIA's "Physical AI Explained" — 4 minutes</strong>
<em>NVIDIA's corporate channel published a short explainer on how physical AI differs from language AI, what "digital twins" actually mean for factories, and why edge inference (running AI directly on machines, not in the cloud) is the next wave for manufacturing.</em>
🔗 <a href="https://www.youtube.com/@NVIDIA">NVIDIA YouTube — Physical AI</a>
---
## 📄 Paper Worth Skimming
<strong>"LLMs for Automated Production Planning: A Manufacturing Case Study"</strong> <em>arXiv, June 2026</em>
- 📌 <strong>The problem:</strong> Production planning for custom orders still requires hours of manual scheduling — even in factories with ERP systems. - 💡 <strong>The insight:</strong> Researchers used GPT-4 to parse natural-language customer orders and generate optimized production sequences, matching human planners' output 89% of the time but in under 2 minutes. - 🏭 <strong>Why it matters for us:</strong> Custom manufacturing orders with unique BOMs and routing requirements are exactly where AI scheduling shows first ROI.
🔗 [arXiv: https://arxiv.org/search/?query=LLM+production+planning+manufacturing&searchtype=all]
---
## 🏭 Shop Floor Story
<strong>Quality Lab Uses ChatGPT to Triage Inspection Reports</strong>
<strong>The challenge:</strong> The quality lab at a midwestern manufacturer receives 40+ inspection reports daily across 3 shifts. Only 10% need engineering escalation, but a human reads all 40 to sort them. That's 3 hours every day.
<strong>How AI helped:</strong> The quality manager pastes each report into ChatGPT Enterprise, using a structured prompt: <em>"Classify this inspection report. Escalate if: tolerance exceeded by >5%, first-time failure for part X, or repeat failure within 30 days. Otherwise, log and close."</em> ChatGPT returns a 2-line triage note and a recommended action.
<strong>The result:</strong> 3 hours → 45 minutes. Engineers now only see the 4-5 reports that actually need their attention. The prompt template is saved on AI@Justrite SharePoint.
---
## 🛠 Tool Updates
| What changed | Details | |---|---| | <strong>ChatGPT Work available</strong> | Enterprise-only upgrade. Ask Remus about access — requires admin approval per seat. Persistent project memory, scheduled tasks, cross-tool integrations. | | <strong>Docling v2.3</strong> (IBM) | Updated document parser. Now handles BMP images + improved XBRL finance table extraction. Useful for pulling tables from supplier invoices and financial reports. | | <strong>LiteLLM Proxy</strong> | Currently routing 8 models across 7 providers. New: Anthropic Claude models available via OpenRouter fallback. Contact Remus for API key. |
---
## 🔥 Learning Flame — ⚙️ How It Works
<strong>How an LLM Actually Works</strong>
It doesn't "think" — it predicts the next word. When you type "The factory produced...", the model scans its training data and calculates probabilities: "goods" (87%), "items" (6%), "nothing" (0.01%). It picks the most probable, then predicts the word after that, and the word after that. A paragraph is 200 predictions in sequence. No reasoning — just statistical pattern matching at massive scale.
> <em>📂 Full Learning Flame archive: AI@Justrite SharePoint</em>
---
## 🤖 AI Eat My Homework
<strong>UK Law Firm Slammed Over Junior Lawyer's AI Blunder</strong>
A UK judge harshly criticized the prestigious law firm Pinsent Masons after a junior lawyer submitted court documents containing AI-generated errors in May 2026. The judge called it a "catastrophic AI blunder" and warned other firms about the risks of unreviewed AI-generated legal filings. The takeaway: AI saves time. It doesn't replace review. The same rule applies to financial reports, supplier contracts, and quality documentation — AI output is a first draft, never a final answer.
🔗 <a href="https://www.legalcheek.com">Legal Cheek</a> — "Pinsent Masons slammed by judge over junior lawyer's AI blunder"
---
## 🏆 Automation Challenge
<strong>This month's theme:</strong> "What's the one manual report you'd delete from your week if you could?"
Submit your automation idea here: <strong>[Microsoft Forms Link]</strong>
<strong>Prize:</strong> $50 Amazon gift card for the winning idea — selected by the AI & Automation team. Deadline: <strong>July 31, 2026</strong>.
> 🏅 <em>Last month's winner: Mark T. (Purchasing) — "Automated supplier quote comparison sheet." Prototype in testing.</em>
---
## 🎙️ AMA: Ask Me Anything
<strong>Every Thursday, 2:00 PM CT</strong> — 30 minutes. No presentations, just answers. Bring your toughest AI question or your simplest "how do I even start?" and I'll help.
🔗 <strong>[Teams Meeting Link]</strong>
---
## 💬 Question for You
> What's one task you did last week where you thought, halfway through, > "I bet there's a faster way to do this"?
Reply in the thread 👇
---
<em>Newsletter by the AI & Automation team — Remus Samoila, IT Analyst, Automation & AI</em> <em>Got news, a tool, or an idea? Forward it to me — I'll include it next week.</em>