of surveyed office professionals used AI tools their employer had not approved.
PagerDuty, June 2026 (international)HERAMIND / AI WITH GUARDRAILS
Use AI the way a sound finance process works.With guardrails you can show.
See what happens to your data, who decides and what gets stopped.
WHY NOW
AI use is running ahead of the rules.
Your people already use AI. Supervision is coming. Set the limits now.
of surveyed organisations have no tested kill switch for AI.
Kiteworks, 2026 (international)New date for EU obligations on stand-alone high-risk AI after the Digital Omnibus. AI literacy has applied since 2025.
Gibson Dunn analysis, 2026Figures from public research, consulted on 3 November 2026. They indicate a direction and are not legal advice.
THE HERAMIND GUARDRAIL MODEL
Seven layers. Each one can be tested.
Per layer: what you get and how we do it ourselves.
01
Purpose and limits
- The question
- What may AI be used for here, and what not?
- What you get
- A register of AI use cases with an owner, a purpose and the tasks that deliberately stay outside AI.
- How we work ourselves
- A language model never takes a financial or legal decision for us and never sends an automatic payment reminder.
- Reference frame
- EU AI Act (risk classes), ISO/IEC 42001
FOLLOW ONE REQUEST
How the guardrails work in practice.
Pick an example and press Start. Fictional data.
AI drafts a reminder for an open invoice.
Log
HOW WE WORK OURSELVES
We apply this to our own work first.
Our own rules, every day.
Several models, each with its own task
Claude, OpenAI, Grok, Cursor and four local models on our own hardware. Sensitive material stays local.
One workspace with sources
Pandora OS brings knowledge, decisions and handovers together, so context does not disappear into separate chats.
One writer per component
Two AIs never work on the same file at the same time. It stays clear who changed what.
A go for each action
Publishing, sending, paying and deploying only happen after explicit permission, every time.
Secrets stay out of sight
Passwords, keys and codes never appear in logs, chats or handovers. Private data is encrypted.
Independent review
Work with material risk gets a separate review before it moves on.
This describes our own way of working. It is not a client result and not a certification.
What we deliberately do not promise
Honest about limits.
- No 100% safety. We design so that a mistake is caught before it does harm.
- No certificate or legal declaration of conformity. Standards are our reference; legal review is for a lawyer.
- No fully autonomous AI for decisions about money, people or rights.
- No vague results: you see what has been tested and what has not.
HOW WE START
Start small, grow with evidence.
Scope and price are agreed after an introduction.
AI guardrail scan
In one to two weeks: where AI is used, where the risk is and what to fix first.
Setting up guardrails
We set it up, test it and hand it over to your team.
AI literacy by role
Short sessions per team, with your own examples.
Can you show where the limit of your AI lies?
Book a 25-minute call. You will know where you stand.