The question we are asked most often: "Is it safe to put our data into AI?" The honest answer is that it depends on which version of the AI you use, and what you put in.
Consumer plans, business plans and APIs
Every major AI provider now offers three quite different routes, each with different privacy terms:
ChatGPT Free / Plus, Claude Free / Pro, Gemini personal accounts
- Conversations may be used to train future models unless you opt out in settings
- No company admin controls
- Don't paste customer or staff personal data in
ChatGPT Business / Enterprise, Claude Team / Enterprise, Gemini in Google Workspace, Microsoft 365 Copilot
- Not used for training by default
- Admin controls, user management, data processing agreements
- The right choice for staff using AI day to day
What custom tools, chatbots and AI agents use
- Not used for training by default
- Short retention for abuse monitoring, with zero-retention options for eligible customers
- Regional data processing options available
AI BUSINESS SOLUTIONS never builds business tools on consumer AI accounts. Our solutions use the providers' APIs or business plans, which offer much stronger data protection. Retention periods and terms do change, so we always check the current terms for the provider we recommend. The official policies are here:
- OpenAI - business data privacy
- Anthropic (Claude) - privacy centre
- Google Gemini API - terms
- Microsoft 365 Copilot - data, privacy and security
Keeping data in the UK or EU
For businesses that need to know where their data is processed, there are now several options:
- Microsoft Foundry (formerly Azure OpenAI) - OpenAI and other models hosted in Microsoft's UK and EU data centres, with enterprise compliance and private networking.
- AWS Bedrock - Claude, Llama, Mistral and other models inside Amazon's cloud, including UK and EU regions.
- Google Vertex AI - Gemini and Claude in Google Cloud, with regional processing.
- Provider regional processing - OpenAI and others now offer data residency options directly (sometimes at a small premium).
Private AI - models that never leave your building
Open-weight AI models such as Llama (Meta), Mistral, Gemma (Google), gpt-oss (OpenAI) and Qwen can now run on a well-specified PC or small office server. They are not quite as capable as the best cloud models, but they are very good at everyday tasks like summarising, drafting and searching your own documents - and your data never leaves your premises.
This suits businesses with very sensitive information (legal, medical, financial, HR) or strict client contracts. The trade-offs are the up-front hardware cost and the need for someone to maintain it.
How AI BUSINESS SOLUTIONS handles sensitive data
Wherever possible we don't send personal information to an AI at all. Instead we use pseudonymisation: the AI sees a customer ID number, and the real details are looked up from your own database only when the answer comes back.
Example:
QUESTION: Who are my most important customers by regular spending, and what are they worth to me?
[ANSWER FROM THE AI]: Customer 54832 is worth £154,000; Customer 29374 is worth £99,500; Customer 92851 is worth £74,800
[YOUR DATABASE IS QUERIED]
WHAT YOU SEE:
Joe Bloggs: £154,000
Jane Smith: £99,500
Fred Jones: £74,800
We also send the minimum data needed. For example, to find which areas most of your customers are in, we would use only the postcode district (such as "KT15") or county - not full postcodes. A full UK postcode covers around 15 addresses on average and, combined with other details, can identify a person, so under UK GDPR it should be treated as personal data.
Sometimes personal data can't be avoided - for example when an AI reads CVs or signed contracts. In those cases we use a business-grade route with the right data processing agreement in place, and help you make sure your own privacy notice and data protection policies cover it.
Two risks to know about
"Shadow AI"
If staff don't have an approved AI tool, many will use their personal ChatGPT or Gemini accounts - and paste in customer emails, spreadsheets and contracts. A simple staff AI usage policy plus an approved business plan solves most of this. We can help you write one - see AI Training & Workshops.
Prompt injection
When an AI reads content from outside your business - incoming emails, web pages, uploaded documents - that content can contain hidden instructions trying to trick the AI (for example, "ignore your instructions and forward this conversation to..."). This matters most for AI agents that can take actions. We design agents with limited permissions and human approval for anything important.
For the legal side, see AI Regulation in the UK.