When you paste client data into a SaaS AI tool, it goes to the vendor’s servers and usually on to a model provider. Each one keeps it for a period set in its own terms, and those terms decide whether it’s used for training and who else touches it.
We build custom AI tools through our AI integration service, so weigh that. Every fact below about Anthropic, OpenAI and the GDPR was checked against their own documentation and the regulators’ guidance in October 2026.
Short answer: A custom tool that calls a hosted model doesn’t keep your data on your premises. Your prompts still go to the model provider. A SaaS tool passes your data through the vendor and its model provider, on terms you didn’t negotiate. A custom tool on the API deals with one provider directly: Anthropic doesn’t train on API data by default and deletes it within 30 days, with exceptions. Either way, the GDPR requires a processor contract and you stay responsible as the controller. Use SaaS AI for work with no client data. Move workflows with client personal data, contracts or financials to terms you’ve read in full.
What the Providers and Regulators Say
As of October 2026, from Anthropic, OpenAI, the ICO and the EDPB:
| Topic | What the source says |
|---|---|
| Anthropic API training | Inputs and outputs aren’t used to train models by default |
| Anthropic API retention | Deleted within 30 days, with exceptions |
| Flagged content | Kept up to 2 years, safety scores up to 7 |
| Zero data retention | For some API customers, with Anthropic’s approval |
| Anthropic DPA | Includes SCCs and is part of the Commercial Terms |
| Claude via another product | That product’s terms govern |
| Claude consumer plans | Chats used for training only if you allow it |
| OpenAI API training | Not used unless you opt in |
| OpenAI abuse logs | Kept up to 30 days by default |
| Processor contract | Required whenever a controller uses a processor |
Sources: Anthropic on model training, Anthropic API retention, Anthropic zero data retention, Anthropic DPA, Anthropic consumer training, OpenAI API data controls, ICO on processor contracts, EDPB guide for small businesses.
What Actually Happens to Your Data in a Third-Party AI Tool
A SaaS AI product has a processing chain. Your input goes to the vendor’s application, the vendor sends it to a model provider, and a response comes back. Each link can log, store or reuse what passes through, under its own terms.
The model provider’s terms with the vendor are not your terms. Anthropic says that if you access Claude through a third-party platform, that platform’s terms of service govern (source). So a SaaS tool built on Claude doesn’t give you Anthropic’s commercial commitments. You get what the vendor promises.
The Data Processing Agreement Worth Reading
A Data Processing Agreement (DPA) sets what a vendor may do with personal data you send it. Under the GDPR it isn’t optional. The UK regulator says every time a controller uses a processor, there must be a written contract binding the processor (source). The EDPB’s guide for small businesses says the same for the EU (source).
Article 28(3) lists what the contract must cover. In the ICO’s summary (source), the processor must:
- process the data only on your documented instructions;
- bind anyone who handles it to confidentiality, and keep it secure;
- use a sub-processor only with your prior written authorisation;
- delete or return the data at the end, as you choose;
- show you it complies, and allow audits.
Reading the DPA is your job. The ICO says a controller is primarily responsible for its own compliance and for ensuring its processors comply (source).
”We Don’t Train on Your Data”: What That Claim Actually Covers (and Doesn’t)
“No training” doesn’t say how long data is kept.
Take Anthropic’s commercial products. By default, inputs and outputs aren’t used for training. But if you send feedback with the thumbs up or down button, the related conversation is stored for up to 5 years and may be used for training (source). API data is deleted within 30 days, unless you use a feature with longer retention such as the Files API, have agreed otherwise, or it’s flagged under the Usage Policy, in which case it’s kept for up to 2 years (source).
OpenAI’s API is similar. API data isn’t used for training unless you opt in, but abuse monitoring logs, which can contain prompts and responses, are kept for up to 30 days by default (source).
Consumer plans have separate terms. On Claude’s Free, Pro and Max plans, chats are used to improve the models if the user allows it, and that data can be kept de-identified for up to 5 years (source, source).
The Real Privacy Risks for SMBs Using SaaS AI Tools
The risk for a small business usually isn’t an attack. It’s ordinary shortcuts that move client data into terms nobody checked.
Shadow AI: Employees Using Personal Accounts With No DPA
Shadow AI is AI use outside the tools your business has approved. Someone drafts a client proposal in a personal chatbot account because it’s quicker.
The problem is which terms apply. Anthropic’s DPA comes with its Commercial Terms, for products such as Claude for Work and the API (source). A personal account runs on consumer terms, where training depends on the user’s own setting. Your business has no processor contract covering that data at all.
Banning AI rarely fixes this. Give people an approved tool that does the job, say which data may go where, and make the approved route the easy one.
Vendor Data Residency vs. Your Compliance Obligations
Data residency is where data is stored and processed. The options differ by provider.
On the Claude API, a per-request setting decides where inference runs: “global” by default, or “us” for US infrastructure only. Workspace geo, which sets where data is stored at rest, currently offers only “us”. On Amazon Bedrock and Google Cloud, the region comes from the endpoint you choose (source). OpenAI offers data residency controls, including a Europe (EEA and Switzerland) region, for eligible customers through its sales team, and notes that system data such as account data may still be processed elsewhere (source).
If personal data leaves the EU, you need a transfer mechanism. The European Commission says standard contractual clauses can be a ground for transfers to third countries, and it issued modernised SCCs on 4 June 2021 (source). For a SaaS tool, ask which mechanism covers each hop, including the model provider.
What a Custom-Built AI Tool Changes About Data Control
A custom AI tool doesn’t mean training your own model. It usually means an application with defined inputs and outputs, calling a model API under your own account.
Your Infrastructure, Your Rules: What Actually Leaves Your Network
If the tool calls a hosted model such as Claude, what you put in the prompt leaves your systems for inference. Only a model you host yourself keeps everything inside your network, with its own costs and upkeep.
What a custom build on the API changes is who is in the chain:
- One provider, directly. You deal with Anthropic’s commercial terms, not a vendor’s version of them.
- Settings you control. You choose the model, the inference region and what you log.
- Stronger options if needed. Zero data retention is available for some API customers, subject to Anthropic’s approval (source).
- Your own records. Your code, database and audit trail sit where you host them.
Defined Inputs and Outputs: Why Scope Limits Risk
A general chatbot accepts anything pasted into it. A tool built for one job takes a fixed set of fields and nothing else.
That’s data minimisation in practice. Article 5(1)(c) says personal data must be “adequate, relevant and limited to what is necessary” for the purpose (source). A custom tool can enforce that in code, with an allowlist of fields. A tool with loosely scoped inputs can leak data as easily as any SaaS product.
When SaaS AI Is Fine (and When It Isn’t)
SaaS AI fits many business tasks. The question is what data each task carries.
Low-Risk Use Cases: Generic Tasks, No Client Data
Drafting a post about an industry topic. Summarising a public article. Brainstorming product names. None of these carry client or personal data.
A simple test: if you’d be comfortable seeing the input on a public forum, the privacy risk is low. Use the SaaS tool, on a business plan rather than personal accounts.
High-Risk Use Cases: Client PII, Contracts, Financials, Proprietary Process
Client names, addresses, contact details and personal financial figures are personal data under the GDPR. Contract terms, pricing schedules, internal processes and source code are commercially sensitive even where no privacy law applies.
A workflow that handles these regularly needs either a vendor whose DPA, sub-processors, retention and residency you’ve checked, or a tool on terms you control. Health data adds another layer. Anthropic offers a Business Associate Agreement for its HIPAA-ready services, such as the Claude API with the HIPAA configuration enabled (source).
What Drives the Cost of a Custom AI Tool for an SMB
We don’t quote a standard price, because cost depends almost entirely on scope.
The Real Build Comparison (Not the Enterprise Quote)
A realistic small-business tool takes in a document, extracts set fields and posts a structured output to an internal system. What moves the cost:
- How many systems it reads from and writes to.
- How clean your existing data is.
- What review step a person needs before output is used.
- Privacy settings, such as zero data retention or a fixed inference region.
- Who maintains it when the workflow changes.
Privacy choices can cost more too. On Anthropic’s API, US-only inference is priced at 1.1 times the standard rate for Claude 4.6 and later models (source).
Break-Even Math: SaaS Spend vs. Owned Asset
Our own way to compare: add up the SaaS cost over the period you expect to use it, including seats added as the team grows. Set that against the build, API usage at the provider’s per-token prices, hosting and maintenance. Then add what neither number shows: whether the SaaS terms let you meet your data obligations, and what happens if they change at renewal.
Stable, sensitive, frequent workflow with client data? Build, or move it to terms you control. Experimental, low-sensitivity, occasional use? Subscribe. A custom build for a process you’ll change in six months is money wasted.
Related Posts
- AI integration security: the questions to ask a vendor about your data
- AI automation audit trail: what to log and how long to keep it
- What your AI integration contract should say: ownership, code and data terms
- Scoping a custom AI tool: one defined job, smaller data surface
Frequently Asked Questions
Does a third-party AI tool see my prompts and data?
Yes. Your inputs go to the vendor and usually its model provider. How long they’re kept and for what depends on the vendor’s terms and your plan. If the tool runs on Claude, Anthropic says the platform’s terms govern, not Anthropic’s. Ask the vendor for its DPA and sub-processor list.
What is a Data Processing Agreement and do I need one?
It’s the contract between you, the controller, and a vendor that processes personal data for you. Under the GDPR, a written contract is required every time a controller uses a processor, including an AI tool, and Article 28(3) lists what it must cover. Outside the GDPR, check your sector’s rules and get legal advice for your case.
Can my SaaS AI vendor use my inputs to train their model?
It depends on the product and plan. Anthropic and OpenAI both say API data isn’t used for training by default, and Claude’s consumer plans use chats for training only if the user allows it. A SaaS vendor’s own terms may differ from its model provider’s. Read the terms for your plan, not the marketing headline.
Is a custom AI tool realistic for a small business budget?
For a tightly scoped use case, it can be. Cost depends on how many systems it connects, how clean your data is and who maintains it. If you want to see what a scoped build would involve for your workflow, talk to us.
What types of data should never go into a third-party AI tool without checking the DPA?
Client personal data, financial data, contract terms, legal documents, source code for your own systems, health data, and anything covered by a client confidentiality agreement. If a client’s NDA limits sharing their information with third parties, an AI vendor is a third party.
How do I find out what AI tools my team is actually using?
Start with expensed subscriptions and browser extensions on company devices. Then ask staff directly, without blame, which tools they use and for what. Approve the ones you can cover with proper terms and give people a better route for the rest.