Framework · AI & design · 3 min read
AI usage guidelines
A one-page “how we use AI in UX” covering approved tools, data rules, review rules, transparency and where to ask, so designers can experiment safely without asking permission each time.
Which tools and plans the team may use, agreed with IT security.
Typically: enterprise plans onlyWhich data classes each tool may receive. Anonymise research data first.
Typically: no personal data in unapproved toolsAI output gets the same critique, accessibility and design-system checks.
Typically: a named designer signs offSay when AI was used. Never present AI output as user evidence.
Typically: an "AI-assisted" noteA channel and named people for questions, so "when in doubt" has an answer.
Typically: champions and the DPOA date and an owner, revisited every quarter or when tools change.
Typically: the UX leadUse it when
- Designers already use AI tools, but nobody knows which tools and which data are allowed.
- You are in the first 30 days of an AI plan and need guardrails before you run pilots.
- A new tool appears (or is switched on by default) and people ask "can I use this for research?"
How to use it
- Find out what is already happening. Ask the team which tools they use, for which tasks, and with what data. Check what is already switched on in your organisation's plans. Output: a short list of tools and uses as they are today.
- Agree the tools and data rules with IT security and the DPO. For each approved tool, note the plan or contract and which data classes it may receive (for example public, internal, confidential, personal). Research recordings, transcripts and customer data go only into approved, contracted tools, anonymised first. Output: an approved tools list mapped to data classes.
- Write the review and transparency rules. State that AI output goes through the same critique, accessibility and design-system checks as any other work, that a named designer decides, and when to say AI was used (for example in research synthesis and generated content). Output: three to five short rules.
- Say where to ask. Name a channel, the AI champions and a contact in IT security or data protection. "When in doubt, ask" only works if people know whom to ask. Output: one line with names and a channel.
- Fit it on one page and test it. Cut everything that is detail and link to it instead. Ask one sceptical designer and the DPO to read it and point out gaps. Output: guidelines v0.1, reviewed.
- Publish and review every quarter. Put it where designers work, add the date and owner, and revisit it each quarter or when a tool changes. Output: a dated, owned page on the team wiki.
Worked example
An internal operations platform with 6 product teams and 8 designers. In the first month, the UX lead finds that two designers paste interview transcripts into a personal chatbot account, and an AI design tool was switched on by the company's plan without anyone deciding to use it. The lead drafts v0.1 with IT security and the DPO:
| Section | What the page says |
|---|---|
| Approved tools | The company's enterprise AI assistant and the design tool's AI features on the company plan. No personal accounts for work. |
| Data rules | Public and internal data: approved tools. Customer and operational data: only tools approved for confidential data. Recordings and transcripts: anonymise first, approved research tool only. |
| Review rules | AI output is a draft until a designer reviews it against the spec, the design system and accessibility. The designer who ships it owns it. |
| Transparency | Mark research synthesis and generated content as "AI-assisted". Never present AI output as user evidence. |
| Where to ask | Team AI channel, the two AI champions, and the named data protection contact. |
| Owner and review | UX lead; next review in three months. |
The two designers move their transcripts to the approved research tool. Nobody is blamed: the page simply makes the safe way the easy way.
Watch out for
- Writing it alone. Guidelines that IT security or the DPO have not seen can be wrong, and then they will be overruled. Co-write them.
- A long policy instead of a page. If it does not fit on one page, designers will not check it before they paste data into a tool. Link the details.
- Only saying "no". A list of bans pushes people to unapproved tools. Say clearly what is allowed and for what.
- Never updating it. Tools and plans change often. A page with an old date and no owner stops being trusted.
Template
Copy this into your team wiki and fill it in with IT security and the DPO.
How we use AI in UX v0.1 · Owner: Next review:
1. Approved tools
Tool / plan: Allowed data classes:
Tool / plan: Allowed data classes:
Not allowed: personal or consumer accounts for work data
2. Data rules
Public / internal:
Confidential (customer, operational, pricing):
Personal (recordings, transcripts, employee data): anonymise first, approved tool only
3. Review rules
- AI output is a draft until a named designer reviews and signs it off
- Same critique, accessibility and design-system checks as any other work
4. Transparency
- Say when AI was used in research synthesis or generated content
- Never present AI output as user evidence
5. Where to ask
Channel: AI champions: Data protection contact: