Canopy Brands · AI Day
I don’t write code.
So how did I rebuild the Xtirpa Academy with a team of AI, and what could that mean for Canopy Brands?
Damiano Cao · Xtirpa
First, the subject
What is Xtirpa Academy?
The home of all technical documentation for Xtirpa confined space and fall protection systems.
Public
Open to everyone
Manuals, datasheets, certifications, brochures
Shared case by case
Restricted documents
Opened per person, by role or invitation
Certified technicians
For trained people
Those who completed Xtirpa training and can perform maintenance and repair
The hook
Two ways to build a platform
The usual way
- Write a brief
- Wait for a quote
- Kick-off meeting
- Wait for the next sprint
- First draft misses the point
- Feedback, back and forth
- Corrections, then more corrections
- Every adjustment is a change request
- Wait for testing and verification
- Pay for every change
- Hope it matches what you meant
What actually happened
- One person who knows the product
- An AI architect and an AI engineer
- Changes discussed, built and checked the same day
- Live at xtirpa-academy.com
- 8 languages, admin, AI assistant
- Every decision made by the expert
Why I’m here
Not a developer. That’s the point.
0
lines of code I wrote
to build the platform
Marketing manager and support
Where? …Global
Product knowledge + a method
Fun fact
The first computer bug was a real moth.
In 1947, engineers found one stuck inside the Harvard Mark II and taped it into the logbook.
My bugs today are caught by AI. I’ll show you three of them.
AI in plain words
An LLM is a brilliant new hire
It has read the whole library. It has never set foot in our warehouse.
Knows a lot
Language, code, law, engineering patterns, in dozens of languages.
Knows nothing about us
Until we give it our documents, our rules and our data.
Needs a manager
Clear tasks, checks, and someone who signs off.
The AI landscape in one picture
Three levels of AI
1
Model
The brain
Claude, GPT, Gemini, Mistral. Understands and writes language.
2
Assistant
The conversation
Chat apps and office copilots. You ask, it answers.
3
Agent
The hands
Claude Code, browser agents. It acts in real tools: files, terminal, databases.
Most companies stop at level 2. The Academy was built at level 3.
Before and after
From a website to a platform
BEFORE · academy.xtirpa.com
After · live today
Open live site ↗8 languages.
One platform.
Click עברית: the whole layout mirrors, right to left.
The tour · the infrastructure
Everything behind xtirpa-academy.com
Why not just WordPress?
Great for blogs. Fragile as a platform.
New WordPress vulnerabilities per year
+42% in one year, 91% of them in plugins
Patchstack, State of WordPress Security 2025 and 2026
The tour · front end
What customers get
Live site
Open live ↗The tour · back end
Control without calling IT
The tour · AI inside
AI is also a feature, not just a tool
The assistant
A round icon in the corner that answers from our own document library.
Translation with a safety net
DeepL for interface text. Legal text drafted by Claude, checked automatically, labelled “machine-translated” until a human reviews it.
The monthly UX report
Written by AI from anonymous numbers. Code inserts the figures; a verifier drops any claim the numbers don’t support.
No lock-in
The AI provider is a setting, not a hard-wired vendor. Open models such as Llama, Mistral or Gemma can even run on our own server: no licence fees, only the hardware.
Live tour
What customers notice first
Live site
Open this page live ↗- Easy access to all shared technical docs
1,244 documents: 280 manuals · 76 certificates · 888 datasheets - Marketing material in 2 clicks
30 brochures · 4 catalogue editions - Language switch that adapts itself
8 languages, even right to left - Private area, fully customisable
Each user arranges their own home screen
The secret
I decide. AI executes. AI checks AI. A human approves.
The engine room · the workflow
How one change travels
1
Me
The need, in plain words
2
Architect
Claude in a chat
Designs it, writes a numbered prompt
3
Engineer
Claude Code
Builds and tests it on the real code
4
Review
AI checks AI
The report is checked against the code, not trusted
5
Release
I apply the database change, deploy, spot-check live
Two AIs with two jobs, and the human at both ends.
The engine room · house rules
Four rules written into every prompt
No secrets in chat
Keys and passwords never pass through the conversation.
No live changes by the AI
The AI builds. Only I touch production.
Tests on a local copy
Automatic tests never run on real data.
“Done” means live and checked
Not “the code looks fine”.
The engine room · the toolbox
Trusted tools, wired together
The engine room · platform health
Cron jobs: what runs while I sleep
The honest part · chat: guess what went wrong
Where AI got it wrong
The visitors who vanished
?
A daily total quietly counted fewer visitors than really came.
Caught by
A rule: every total must equal the raw data. Never reached production.
The one-minute login
?
“Draft and publish” failed live: the sign-in pass expired before the AI finished translating.
Fixed by
Splitting the job into small steps.
The checker who couldn’t count
?
The checker kept rejecting a correct French translation. It was miscounting a bullet list.
Caught by
Testing the checker on the real policies, not guessing.
Why this is better
The usual route versus the method
| The usual route | Our method |
|---|
What’s next
The Academy was the pilot
Next: a new Xtirpa website, talking to the Academy through a locked gate.
Only what is authorised can pass through
Faster loads: each site reads its own data
One clear, single source for every piece of data
One sign-in for customers across both
New website
its own database
Academy
its own database
Governance on one page
Rules first, tools second
AI may
- Draft and summarise
- Translate
- Write and test code
- Analyse anonymous numbers
Always a human
- Publishing
- Legal text and prices
- Customer commitments
- Anything that goes live
Data
- Approved tools only
- Company accounts only
- No passwords or keys in chat
- No customer personal data in public tools
I don’t write code.
I write decisions.
Let AI help you write the next ones.
Damiano Cao · xtirpa-academy.com
This is how work will be done
Microsoft already works this way, at 1,000× the cost
Up to 30% of Microsoft’s code is now written by AI
Satya Nadella, Microsoft CEO, April 2025
2025
Claude added to Microsoft 365 Copilot and Microsoft Foundry
Microsoft, September and November 2025
NOW
AI agents inside Business Central: the Sales Order Agent turns customer emails into orders
Microsoft Learn, Business Central documentation
2026
Microsoft encouraged staff with no coding experience to try AI coding agents. The tool changed; the agents stayed.
The Verge, January and May 2026 (Claude Code, then GitHub Copilot CLI)
Organisations using AI
In at least one business function · McKinsey, The State of AI (2024, Nov 2025)
280× cheaper
Same AI quality per million tokens: $20 (Nov 2022) → $0.07 (Oct 2024)
Stanford AI Index 2025
The question is no longer whether Canopy uses AI. It’s who runs it, and how safely.
Coming soon to xtirpa-academy.com and my other projects
AI can help you build anything
Integrated Mailchimp
Newsletter lists synced with the Academy, consent included
Automated social content
Posts drafted from new products and documents
Real-time 3D configurator
Build the system in 3D, see it in self-hosted AR
Voice-driven quotes
Say what you need; the quote is drafted for you
Business Central connection
Products, prices and stock straight from the ERP
Direct digital invoicing
Invoices issued and sent automatically
ERP and CRM
One view of every customer and order
E-commerce
Order online, straight from the configurator
Shipping integration
Rates, labels and tracking from the carrier
Any API, plus AI
If it has an API, we can connect it with AI
Endless possibilities
Tell AI the process. You decide together; AI builds it; a human approves it.
Q&A
Let’s see how the future unfolds
Special thanks to Prometheus, who stole fire so we would never stop being curious.