05 / writing
Notes on AI adoption, Microsoft 365, and building operations software. Nothing gets published unless I'd want to read it twice.
AI systems don't cite your best writing. They cite your clearest sentence. Here's what makes a page quotable, and why it's the same thing that makes it useful.
Your customers are asking AI assistants what your business does before they reach your site. Most owners have never checked the answer. It takes two minutes.
The most common diagnosis of a stalled rollout blames the people. What reads as resistance is usually a reasonable response to unowned change.
Most AI pilots don't fail loudly. They trail off undecided, which means full cost and no verdict. One question asked before the pilot starts prevents it.
Agentic AI gets the headlines. Most valuable production AI is a single LLM call doing one job well. Here's how to tell which one your problem actually needs.
The fastest way to tell whether an AI purchase will produce value is to ask what it replaces. If nothing stops, nothing changed, and you bought a subscription.
A reconciliation engine that works is a pipeline, not a chat window: four phases, a three-part prompt, one job each.
Enterprise buyers want to know why the system believed it, and what it does when it isn't sure. That's what makes an AI system explainable.
Security teams need to know how sensitive data is handled before they'll approve an AI system. The architecture and rollout plan that get this approved.
Reconciliation rarely fails loudly. It fails in errors too small to catch alone and too numerous to ignore together, and the real cost isn't the errors.
Not every reconciliation task is a good candidate for AI. The best systems split the work: AI where judgment is needed, code where certainty is required.
IT has spent a decade automating reconciliation, and it stalls in one of two predictable ways. Both failures point to the actual fix.
Skip the prompt libraries. Office work runs on five repeatable patterns, and knowing them beats memorizing prompts you’ll never look up under deadline.
If an AI rollout is running in your organization right now, one question tells you whether it exists in six months. Ask it before your first meeting today.
A perfect preview box is why broken LinkedIn links go unnoticed, by you and by the AI systems reading your content. Three ten-second checks that catch it first.
Copilot solves real problems for operations teams, and gets credit for problems it can't touch. When to stay in M365, extend it, or build.
Unglamorous, high-frequency, genuinely shippable. Five operations workflows where AI works today, and the effort each one takes.
The actual stack and workflow I use to take an AI product from idea to production. What each stage is for, where the handoffs happen, and what breaks.
The license fee is the smallest number involved. The real math on a stalled AI pilot, from staff hours to opportunity cost to lost credibility.
Most AI pilots don't fail on technology. They fail because the team bought a tool when the job was changing a workflow. Three questions to ask first.
A plain-English guide to Intelligent Document Processing: how AI reads documents, how it differs from OCR, and what it delivers for operations teams.
What this section is for and what to expect from it.