- WKND AI
- Posts
- Claude On Microsoft Foundry
Claude On Microsoft Foundry
+Copilot For Excel
In partnership with
The best prompt engineers aren't typing. They're talking.
Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them.
Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work.
89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone.
⬆️Support this newsletter by clicking on the link above ⬆️
Hello WKND AI Warriors!
Claude is now built into Microsoft Foundry and Microsoft 365, making it easier for companies to deploy long-running AI agents.
Also, Microsoft just turned Excel into an AI finance analyst that can build models, explain formulas, and pull live ERP data without leaving your spreadsheet.
Plus, Mistral released a new OCR model that reads documents in 170 languages for a fraction of the cost of competing services.
Oh yeah, and OpenAI and Broadcom built a custom AI chip called “Jalapeño” to challenge NVIDIA’s dominance
Today’s newsletter includes:
💼 AI AROUND THE OFFICE
📰 AI NEWS RECAP
🤿 AI DEEP DIVE
💼 AI AROUND THE OFFICE
Invoice Classification & Document Processing Reality
Background
Multiple clients still review invoices manually, assign GLs, and push them into their ERP. One client uses 10 resources to click through thousands of invoice PDFs, pick GL accounts, and feed an Excel that AP then uploads. The AI automated workflow we helped them with?
Text extraction → Claude via API → ruleset matching → auto-journals → confidence scoring and exception handling.
Pain Point
Simply uploading PDFs/images to ChatGPT/Claude is expensive and error-prone. Tokens explode, hallucinations increase, and you still need people in the loop. At the same time, the status quo is 10 people manually coding invoices.
Root Cause
Raw PDFs/images drive up token usage and cost; vision models are heavier.
No confidence thresholds or exception routing; it’s all-or-nothing automation.
Many orgs underestimate the engineering required to create a robust, auditable pipeline.
High-accuracy local models exist, but require serious upfront training and infra.
Resolution
Build a hybrid workflow: use lightweight text extraction, then send structured text to Claude with clear rules; auto-post only high-confidence entries and route low-confidence ones to a small human exceptions team. Consider local visual models where privacy or volume justify the setup.
Next Steps (What You Can Do Today)
Map the current invoice flow end-to-end, including where PDFs are stored and how GL decisions are made.
Run a small POC:
Extract text from 100 invoices.
Prompt Claude with your GL rules and ask it to assign GLs with a confidence score.
Compare against human decisions and track accuracy + token cost.
Define a simple threshold rule: e.g., “≥95% confidence auto-post, 80–95% for review, <80% reject.”
Decide: based on volume and sensitivity, is it worth investing in a local visual model that you fine-tune on historical invoices?
📰 AI NEWS RECAP
Claude Fable 5 and Opus 4.8 are now available in Microsoft Foundry, letting Azure customers build long-running enterprise agents within their existing Microsoft ecosystem. Claude also powers the Researcher agent in Microsoft 365 Copilot and an Agent Mode in Excel.
Microsoft's Copilot in Excel now handles full finance workflows with prebuilt skills for modeling, variance analysis, and scenario planning, pulling live data from ERP and treasury systems. Every output includes full formula traceability so finance teams can audit results without leaving the spreadsheet.
Mistral's OCR-4 supports 170 languages, extracts bounding boxes and block types from PDFs, can be self-hosted, and tops the OlmOCRBench at just $4 per 1,000 pages, up to 15x cheaper than Azure's custom document tier.
Google baked computer use natively into Gemini 3.5 Flash, matching GPT-5.5 and Claude Sonnet 4.6 on OSWorld-Verified at roughly a third of the cost. It includes built-in prompt injection defenses and works across browser, mobile, and desktop environments.
Perplexity's Computer for Counsel connects to legal databases, DocuSign, NetDocuments, and matter management systems to run full workflows like case research, contract extraction, and regulatory monitoring. It returns finished work products rather than isolated answers, available for Pro and Max subscribers.
Google Finance exited beta with investment portfolios, personalized AI briefings tied to your holdings, and a new Android app with daily scheduled digests. It is the first major redesign of Google Finance in years, rolled out globally.
OpenAI and Broadcom unveiled Jalapeño, a custom LLM inference chip built in nine months using OpenAI's own AI models, claiming performance matching Nvidia Blackwell and Google TPUs per watt. Initial deployment begins late 2026, with full scale-up in 2027-2028.
🤿 AI DEEP DIVE
Claude Design 2.0 just dropped,
And its a game changer.
Have you tried it yet?
