What was delivered
- An interactive presentation page with six sessions and a Play button on every item, so each point opens a live demo or an explainer.
- An opening film rebuilt in three.js by the local LLM, followed by how it was built: a loop engine that checks every attempt against the original.
- The system and security architecture, the first install and permissions, and a plain guide to daily use: context, skills, extensions and the few commands that matter.
- Live demos: the Monkey Saturn Ride game restyled by chat, a request turned into a ticket through browser automation, and a receipt read by OCR that fills a claim form for the user to check.
- A hands-on build: colleagues paste a plain-language prompt and get a working skill, from a small game with a high-score table to a hybrid skill that raises a server backup ticket.
How it stays safe
- The model runs on the company's own server, with no internet and no cloud AI in the path.
- A gateway checks each key and counts usage, and stores no messages.
- Secrets stay in the operating system's credential store; chat history stays on the user's own PC.
- A human approves every change before it runs, and every server command is audited.
Skills colleagues have already built
Every one began as a plain-language prompt: a daily licence-usage report, a warranty checker that takes a list of serial numbers, a report of failed file-transfer jobs, a tool that draws a project's workflow diagram, read-only questions to the ERP, and stock reconciliation.
Why it matters
Agentic AI stops being a tool for a few specialists. Anyone who can describe the job they repeat every day can turn it into a skill, share it with the team and run it again tomorrow. The AI handles the language, a small script handles the fixed steps, and the person keeps the final say.
The rollout is deliberate: IT and cyber teams first, business users through controlled projects, with governance built in, so more departments can retire manual checking safely.