Ask what data is actually needed
Minimisation is useful engineering as well as privacy practice. A task may not need a name, a full document, a full customer history or every spreadsheet column. The less data you give a system, the smaller the surface if something goes wrong.
Perth practice handling sensitive enquiries: a workflow routes customer messages by category. It only sends the subject and a short, redacted summary to the AI step — not the full conversation or account history. Less data, less risk, same outcome.
Minimise before you sendProvider settings matter
Different services and contracts have different retention, training, residency, security and enterprise controls. “It is in the cloud” does not answer the risk question. You need to know what the provider does with the data, what it retains, where it is processed, and how you can delete or export it.
Local AI changes the trade-off, not the responsibility
Running a model on your own hardware can reduce external data transfer, which helps confidentiality. But you still need device security, access control, backups, governance and a lawful basis where personal data is involved. Local processing does not remove UK GDPR / DPA obligations — check current ICO guidance on AI and data protection.
Consequential output needs stronger control
Health, employment, finance, legal or safety-sensitive workflows require specialist review and governance. Caledapt teaches architecture concepts, not legal conclusions. For personal data, review the ICO's UK GDPR guidance and decide whether a DPIA is needed.
Not “can AI use it?” but: what data is involved, why is it needed, where will it travel, what will the provider do with it, what actions can the system take, and who remains accountable?
Public data
- Published material with no personal link in context
- Lowest control burden, still subject to normal use rules
Treat as open, but confirm the source is genuinely public.
Personal data
- Anything that identifies a person, such as names, emails and account history
- Needs a lawful basis, clear purpose and an approved provider
- Minimise and redact before it reaches an AI step
Route through the approved path and check the ICO guidance.
Special-category data
- Health, and other sensitive classes in UK GDPR terms
- Highest control requirement: screening, often a DPIA, extra governance
- Consider local or private processing to keep it under your control
Treat as high risk until a proper review says otherwise.
The routing is not about the model type: a cloud LLM and a local model both need the same data question answered first. Decide what the data is, then choose where it may go.
Checklist before sending data to any AI system
- Identify every piece of data the task would send.
- Minimise and redact where possible.
- Confirm a lawful basis and clear purpose.
- Choose an approved provider and review its retention/training terms.
- Confirm where the data is processed and who can access it.
- Keep a human accountable for consequential output.
Common privacy failure modes
UK GDPR and the Data Protection Act 2018 apply across Scotland. Sensitive sectors (health, care, education, public sector) carry extra obligations. When in doubt, minimise the data and review the ICO's guidance before proceeding.
Frequently asked questions
Can I send customer emails to a cloud AI tool?
Only if you have a lawful basis, a clear purpose, an approved provider and appropriate controls. Personal data needs particular care — check the ICO guidance and consider minimising or redacting first.
Is local AI always more compliant?
No. Local reduces external transfer but does not remove security, access-control or accountability duties. Compliance depends on how the whole system is governed, not where it runs.
Do I need a DPIA for every AI use?
Not for every use, but a DPIA is likely where processing is high-risk to individuals. Use the ICO's screening questions early to decide and document.
The tooling behind this guide is being built in public. It is open source on github.com/BryanFiFife, so you can inspect rather than take our word for it:
Open source does not mean certified. Read the code, run the tests, and treat it like any tool: check it against your own requirements before you depend on it.
A reliable implementation is not “an LLM with a prompt”. It is a controlled business system with a model only where a model adds value.