ai-adoption · deliverable

Governance Basics — employee-facing AI guardrails (AVG/GDPR-aware)

Audience: every Helios employee using AI tools · Length: one page on purpose · Rule of thumb: when in doubt, it’s red — and you ask.

These guardrails exist so that “use AI at your desk” never becomes “explain yourself to the Autoriteit Persoonsgegevens.” They are written for employees, not lawyers: short, concrete, and enforceable by common sense.

1. Data classification — the traffic-light table for AI-tool use

Light	What it covers	AI-tool rule	Examples
Green	Public info · anonymized/aggregated content · generic drafting	Go — use external tools; human still checks facts before send	Public regulations, formatting help, “write a meeting agenda”, anonymized process descriptions
Yellow	Internal but non-personal, non-sensitive business content	Go with care — approved internal/approved-external tools only; strip anything identifying	Draft internal memos, non-confidential vendor terms, shift-log structures without names
Red	Personal data · special-category data · confidential business	Stop — never into external AI tools, no exceptions without written sign-off	Customer PII, employee sickness/performance/case details, unreleased financials, bids, credentials

The two-second test: if the content contains a person who can be identified, or a number someone would panic about seeing published — it’s red. Uncertainty resolves upward: treat as red and ask.

2. The disclosure norm

“Client-facing text drafted with AI gets a human name on it. Always.”

If AI helped draft anything that leaves the building (client email, proposal, report, job ad), a named human has reviewed it and signs for it.
The signature is the disclosure. We do not hide AI assistance, and we do not ship unreviewed AI text under a human name — both are norm violations.
Internal drafts don’t need a disclosure ritual, but the check still happens before the content becomes a decision input.

Why this norm carries the whole program: trust in AI adoption dies the day a client receives a confident wrong sentence with our name under it. The name-on-it rule makes ownership visible and keeps the checking habit alive.

3. When to escalate — and who owns exceptions

Situation	Do this	Owner
You pasted red data into an external tool (even by accident)	Report same day to your AI champion + team lead. Same-day self-reporting = no personal consequences; it’s a fix, not a fault.	Champion logs it; DPO assesses within 2 working days
A tool output looks wrong but you can’t prove it	Don’t send. Flag to champion; if client-facing, hold the deliverable and say so — a late deliverable beats a wrong one.	Team lead decides hold/send
You need red/yellow data processed by AI for a real business reason	Request an exception before doing it: business need + data scope + tool + safeguards, in writing.	DPO owns exceptions; team lead countersigns. No verbal exceptions.
A client asks whether AI was used on their deliverable	Answer honestly per the disclosure norm; if unsure of wording, route to comms/team lead.	Team lead / comms
You see a colleague about to paste red data	Say something, kindly, now. “Hey — is that anonymized?” is a complete intervention.	Everyone owns this one

Exception principle: exceptions are written, scoped, and time-boxed. An exception that becomes “how we do it now” is a policy change — and policy changes go through the DPO and the quarterly review, not through habit.

4. Quarterly refresh (keeps this page alive)

Champion network reviews the escalation log: what got reported, what pattern caused it, does this page need a line added?
Audit sample: ≥20 client-facing deliverables checked for the human-name norm.
Tool landscape check: approved-tool list updated; any new tool is yellow-by-default until reviewed.

Fictional disclosure: Helios Facilities Group is a fictional client for this portfolio; thresholds reflect common EU practice. Guardrails drafted with AI assistance; final judgment and copy are human-owned. This is a portfolio artifact, not legal advice.

Prefer the source? Download the raw markdown.