Guide · GDPR checklist
Using AI with customer data under GDPR: a practical checklist
For sales, marketing and ops teams putting CRM records or contact spreadsheets into AI tools: a nine-point checklist, the questions to ask any AI vendor, and what running AI locally does and doesn’t change.
Is CRM data personal data?
Usually, yes. GDPR applies to processing personal data of people in the EU, and personal data is information relating to a person who can be identified. A typical CRM, and any export of it, is full of it.
| Typical field | Personal data? | Why |
|---|---|---|
| Contact name and job title | Yes | It names a specific person. |
| Work email, e.g. jane.roe@company.com | Yes | It identifies one person, even at work. |
| Generic email, e.g. info@company.com | Usually not | It reaches a team — unless it obviously belongs to one person. |
| Direct or mobile phone number | Yes | It reaches one identifiable person. |
| Notes about a person | Yes | Free text can hold anything, including details of someone’s health or private life. |
| Company name, industry, headcount | Generally not | It describes a company, not a person. |
| A one-person business | It can be | When the company is essentially the person, its data can identify them. |
The takeaway: a list of companies is a much smaller GDPR question than a list of people. If a task only needs company facts, leave the people out of it.
A GDPR checklist for AI in sales and marketing
Work through these with whoever owns data protection where you work.
Lawful basis and purpose
Confirm a lawful basis, and that the AI task fits the purpose the data was collected for. Why: GDPR requires a lawful basis, and purpose limitation means data gathered for one reason shouldn’t drift into an unrelated one.
Send only the columns the task needs
Researching a company’s industry needs its name and website — not the contact’s name, email and call notes. Why: data minimization. Less data sent is less to worry about later.
Processor and processor agreement
If a vendor processes personal data for you, check that a processor agreement (a DPA) covers the plan you actually use. Why: GDPR expects one (Article 28). Your company’s agreement usually doesn’t cover a teammate’s personal account.
Transfers outside the EEA
Find out where the data is processed and stored, including by the vendor’s sub-processors. Why: GDPR has specific rules for transfers outside the EEA, and the vendor should be able to say where your data goes.
Retention of prompts and chats
Does the tool keep prompts, files and chats — for how long, and can you delete them or turn history off? Why: a chat history full of customer names is an unplanned second copy of your CRM.
Transparency toward people
Check that your privacy notice explains in plain terms how contact data is used, including by AI tools. Why: the people in your CRM didn’t choose your AI vendor; your notice is where they find out.
A DPIA for high-risk processing
Ask your data-protection officer whether this use could be high-risk for the people concerned. Why: GDPR calls for a data protection impact assessment (DPIA) for high-risk processing.
Security and access
Know who can use the tool, who sees what’s sent, where API keys are kept and whether accounts are company-managed. Why: each AI tool is one more place customer data can leak from.
Human review before the CRM
Have a person check AI output, with its source, before it reaches your CRM. Why: AI can be wrong — the wrong Jane Roe, an invented job title — and once saved, a guess becomes a record about a real person.
Questions to ask any AI vendor
Whether it’s a cloud AI assistant, an AI feature in your CRM or a browser extension, ask these — and get the answers in writing.
| Question | Why it matters |
|---|---|
| Where is our data processed and stored? | Transfers outside the EEA have their own GDPR rules, and “the cloud” is always somewhere specific. |
| Is our data used to train or improve models? | Customer records feeding a model you don’t control are hard to square with your original purpose. Terms can differ by plan. |
| How long are prompts, files and chats kept? | Retention you didn’t choose is retention you still have to justify. |
| Will you sign a data processing agreement? | When a vendor processes personal data for you, GDPR expects a processor agreement (Article 28). |
| Which sub-processors do you use? | Your vendor’s vendors handle your data too — ask who they are and where. |
| Which company and jurisdiction hold the data? | Under the US CLOUD Act (2018), US authorities can require US-based providers to disclose data in their control, including data stored abroad — one reason some organizations prefer non-US or self-hosted processing. |
| What can we switch off? | Chat history, training use, connectors, telemetry — each off switch is minimization you can enforce. |
| What is sent in the background? | Usage statistics and error reports can carry fragments of your work. Ask what they contain and whether they’re on by default. |
| Is there a local or self-hosted option? | If the model runs on hardware you control, no third-party AI processor is involved for that step. |
Does running AI locally solve GDPR?
Partly. When the model runs on your own computer or a server you control, no third-party processor is involved for that step, and since the data stays on hardware you control, the transfer question for that step goes too. Your own duties — a lawful basis, minimization, transparency, security — don’t go away.
| Cloud AI assistant | AI on hardware you control | |
|---|---|---|
| Third-party AI processor | Yes — the vendor processes what you send | None for that step |
| Transfer outside the EEA | Depends on where the vendor processes | Data stays on hardware you control |
| Lawful basis, purpose, minimization, transparency | Your duty | Still your duty |
| Security | Shared with the vendor | Yours: the device and who uses it |
| Human review of output | Needed | Needed — smaller local models make mistakes too |
Hybrid tools: ask what goes where
Few tools are purely local. An app can run its model on your laptop and still search the web, connect to your CRM, check for updates, send usage statistics or offer optional cloud features. That’s fine if it’s written down plainly: what runs locally, what goes online, what’s on by default and how to switch it off.
Also easy to miss: a web search about a named person sends that name to the search engine, whichever model wrote the query.
If you’re in Switzerland
Switzerland has its own law: the revised Federal Act on Data Protection (revFADP, in German nDSG) has applied since 1 September 2023. It is a separate law from GDPR, but the practical checklist above applies broadly there too. If you also handle data about people in the EU, GDPR can apply as well. A Swiss data-protection specialist can tell you what applies to you.
Where Octopi fits
Octopi is our app, so treat this as one example, not advice. It turns a plain-English task into a finished spreadsheet, researching companies and people in a real browser and working in Salesforce, HubSpot and Pipedrive.
The AI runs on your Mac
An open-weight model — Qwen2.5-7B-Instruct by default — runs on your Mac through Ollama. Out of the box, your spreadsheets, contact lists, prompts and credentials are not sent.
Said plainly
The web searches and page reads you ask for, your signed-in CRMs, the one-time model download and update checks (GitHub). A task that researches a named person can send that name to a search engine.
On by default — and you can turn it off
Usage statistics and scrubbed error reports are linked to the email you downloaded with; a scrubbed error line can still name a page or field. Switching off in Settings ▸ Privacy unlinks and stops it.
Off unless you turn it on
Each Claude feature has its own consent screen. Planning sends your task sentence and column names — not your rows — so a name typed into the task goes too. Claude’s page reading uses a private, signed-out headless browser. Your CRM and sites you sign in to are always handled on your Mac.
CRM changes shown before they’re made
In the chat, a CRM change appears as a card and waits for your Confirm, unless you turned on “Always allow” for that CRM in Settings (off by default). Writes are checked for duplicates and verified.
Only the columns you asked for
Results land only in the columns the task declared, and every value carries a clickable source a person can check.
Octopi doesn’t make your processing compliant on its own: the lawful basis, the purpose and your privacy notice stay yours. Details: Sovereign AI on your Mac and the privacy policy.
Questions
Can I put customer data into a cloud AI assistant under GDPR?
It depends on the checklist: a lawful basis and purpose, a processor agreement covering your plan, where the data is processed, how long it’s kept and whether it trains models. If any answer is unclear, send less — company facts, not people — or keep that step on hardware you control.
Is a work email address personal data?
Yes, when it identifies a person, such as jane.roe@company.com. A generic address such as info@company.com usually isn’t about one person, though it can be if it obviously belongs to one.
Do I need a data processing agreement with an AI vendor?
When a vendor processes personal data on your behalf, GDPR expects a processor agreement (Article 28), often called a DPA. Check that it covers the plan you actually use.
Does running AI locally make me GDPR compliant?
Not on its own. No third-party processor is involved for that step, but your own duties — a lawful basis, data minimization, transparency, security — stay the same.
Do I need a DPIA to use AI on CRM data?
GDPR calls for a data protection impact assessment for high-risk processing. Whether yours is high-risk depends on what you do with the data; ask your data-protection officer before rollout.
Is GDPR the only rule that matters for AI?
Not necessarily. The EU AI Act entered into force on 1 August 2024, with obligations phasing in over several years, and Switzerland has its revised FADP. Which rules apply to you is a question for your data-protection officer or a lawyer.
Does Octopi send my spreadsheet to the cloud?
No. Its model runs on your Mac, and by default spreadsheets, contact lists, prompts and credentials are not sent. It goes online for the web, page and CRM work you ask for, the model download, update checks, and usage and diagnostics (on by default; switch off in Settings ▸ Privacy). The optional Claude features are off unless you turn them on.
Related
Keep the AI step on your own Mac
Download Octopi and run an open-weight model locally — free, no account needed, with the cloud strictly optional. macOS 13+ on Apple silicon.
Download Octopi free