Gustav AI tools have entered business workflows faster than any other technology category in recent memory. ChatGPT, Copilot, Claude - they are in the daily routine of millions of people who are using them to draft emails, summarise documents, write proposals, and process customer information.
Most of this is fine. But "most of this is fine" is different from "all of this is fine," and the difference matters for businesses that handle customer data - which is most businesses.
What cloud AI tools do with your input
When you type something into ChatGPT's web interface, that text is sent to OpenAI's servers, processed by their model, and returned to you as a response. The input is also potentially stored, reviewed by human trainers, and used to improve future model versions - depending on your account type and the settings you have configured.
ChatGPT's default settings for free and basic accounts allow OpenAI to use your conversations for model training. You can turn this off in your settings - and if you are using AI for business, you should check whether it is off. ChatGPT Team and Enterprise plans have different terms: by default, conversations on these plans are not used for training.
Claude (Anthropic) and other major AI providers have similar structures: consumer-facing products may use conversations for training by default, while API access and business tiers typically do not.
The free tier of most AI tools was not designed with your customer data in mind. Upgrading to a business plan or using the API changes the data handling meaningfully.
GDPR and what it actually means in practice
GDPR applies to the personal data of EU residents. If your business handles names, email addresses, purchase histories, health information, or any other personally identifiable information about EU customers, GDPR governs how that data can be processed.
When you paste customer data into a US-based AI service, you are transferring that data to a third party in a third country. Under GDPR, this requires either an adequacy decision (the EU has deemed the country's data protection laws adequate) or appropriate safeguards such as Standard Contractual Clauses.
In practice, major AI providers like OpenAI and Anthropic have done significant work to provide GDPR-compliant options - data processing agreements, EU data residency options for enterprise customers, Standard Contractual Clauses. If you are using the consumer web interface, these protections are not automatically in place. If you are using an enterprise plan or the API with a signed DPA, they likely are.
This is not about whether OpenAI or Anthropic are trustworthy companies. It is about whether your data handling is compliant. If your business is audited or a customer asks how you handle their data, "I paste it into ChatGPT" is not an answer that holds up.
What local models actually are
Local models are AI models that run entirely on your own hardware. Tools like Ollama allow you to run models like Gemma, Mistral, Qwen, and Llama locally - on a laptop, a desktop, or a server inside your organisation. When you send a prompt to a local model, nothing leaves your machine. There are no servers in the US, no data processing agreement to worry about, no terms of service governing your input.
The trade-off is capability. Local models are improving rapidly, but the most capable models - GPT-4o, Claude Sonnet, Gemini Ultra - are cloud-only. For many business tasks, this gap does not matter. For others it does.
Local models are particularly well-suited to tasks that involve sensitive data and do not require the highest possible output quality: first drafts of internal documents, processing customer records, generating structured data from forms. They are less well-suited to tasks that require nuanced judgement, multilingual sophistication, or complex reasoning.
The middle ground most businesses miss
Most businesses using AI sit at one of two extremes: they use consumer AI tools without thinking about data handling, or they worry about AI privacy and do not use it at all. The middle ground - which is where most businesses should be - is using AI through the API rather than the consumer interface.
API access to GPT-4o or Claude means your inputs are not used for training by default. You can sign a data processing agreement. You get the full capability of the latest models. The trade-off is that it requires a developer to build the integration - you cannot just open a tab and type.
This is why most businesses end up building a simple internal tool rather than using the chat interface directly. The tool calls the API, handles the data processing in a controlled way, and presents a simple interface to the user. Your team gets the capability; your data stays where it should.
A simple framework for deciding
For most business AI tasks, the question is not "is this private enough?" but "what am I actually sending?" Here is a reasonable framework:
Drafting from your own ideas: Use any tool. You are not sending customer data. Risk is low.
Processing internal documents without customer data: Use any tool, preferably a business plan. Risk is low.
Processing customer names, emails, purchase history, or other personal data: Use the API with a signed DPA, or use a local model. Consumer interfaces are not appropriate.
Highly sensitive data (health, financial, legal): Local models only, or specialist compliant providers. No exceptions.
The good news is that the practical tools exist for all of these scenarios. The businesses that get this right are not the ones that avoid AI - they are the ones that use it thoughtfully, with the right tool for the right task.