Skip to main content
Guide

GDPR-compliant AI sales tools: what to actually check before you buy.

An eight-point checklist for EU sales teams. Data residency, AI training, sub-processors, signed URLs, and the DPA you need to ask for.

Try a privacy-first option
EU-hosted by default, 5 free responses to start

The short answer

A GDPR-compliant AI sales tool keeps customer data inside the EU, does not train its models on what your prospects say, lists every sub-processor in writing, and signs a Data Processing Agreement under Article 28. Recordings should sit in a private bucket with one-hour signed URLs. Receivers should not need an account. If a vendor cannot confirm those four things on a sales call, keep looking.
8 checks
to run before you sign any AI sales tool contract
EU-only
data residency is the cleanest way to avoid Schrems II questions
1 hour
is the right expiry for any shareable recording link

Why this checklist exists

Most AI sales tools are US-hosted, route through OpenAI or Anthropic by default, and ask the customer to create an account before answering anything. For an EU sales team, every one of those defaults creates a paper trail your DPO will eventually want to read. The fix is not to avoid AI, it is to pick a tool whose defaults match GDPR, so you do not have to retrofit compliance after the fact.

The eight checks below are what actually moves the needle. Everything else is hygiene. If a vendor cannot answer the first three on a 30-minute call, do not bother with a trial.

The eight checks, in priority order

Run these against any AI sales tool before signing.

1. Where does the data physically sit?

Ask for the region of the database, the file store, and any caching layer. The answer should be an EU region, ideally Frankfurt, Dublin, or Amsterdam. If the answer is "global" or "US with EU failover", that is a US-hosted product with marketing copy.

2. Which AI provider runs the inference?

Get the name. Mistral AI runs in the EU. Azure OpenAI can be pinned to an EU region. Vanilla OpenAI and Anthropic process in the US. Knowing the provider tells you 80% of the data flow story before you read any DPA.

3. Is your data used for AI training?

The contractual answer should be no, by default, with no opt-out box you have to remember to tick. Ask if the AI provider's API tier itself excludes training. For Mistral and the Azure OpenAI service, that is the default. For consumer OpenAI, it is not.

4. Is there a GDPR Article 28 DPA on the website?

A vendor that has thought about GDPR posts the DPA publicly. You should be able to read it before talking to sales. If you have to email "legal@" and wait three days, that is a tell.

5. What is the sub-processor list?

You want a public, dated list. Each entry should name the service, the purpose, and the region. A typical AI sales tool has a database, an object store, an LLM provider, an email sender, and an analytics tool. Five entries is normal. Twenty-five is a red flag.

6. How is recorded media accessed?

The right pattern is a private bucket plus signed URLs that expire within one hour. The wrong pattern is a public CDN URL that anyone with the link can play forever. Ask the vendor to describe the playback flow. If they cannot, the answer is probably the wrong one.

7. Does the receiver need an account?

For prospect-facing tools, the answer should be no. Forcing a signup means collecting an email and a password, which is more personal data than the use case needs. A no-login Magic Link is the GDPR-friendlier default.

8. How fast can you delete a record?

GDPR Article 17 gives data subjects a right to erasure. The tool should let you delete a single response, including the audio file and the transcript, from the dashboard in one click. If deletion is a support ticket, that is a process you will fail to honor under load.

A worked example: how HeySpeak scores

HeySpeak is one async voice feedback tool built for EU sales and research teams. Running it through the same eight checks:

  • Data residency: Supabase Postgres in Frankfurt, Cloudflare R2 private bucket. No US database.
  • AI provider: Mistral AI, 100% EU-hosted, no training on customer audio or transcripts.
  • AI training: contractually excluded for both transcription and summarization.
  • DPA: available on request, written under GDPR Article 28.
  • Sub-processors: Supabase, Cloudflare R2, Mistral AI, Brevo for transactional email, Stripe for payments. Five entries, all disclosed.
  • Recorded media: private R2 bucket, playback via signed URLs that expire after one hour.
  • Receivers: no account, no login, no email collected unless they volunteer it in the recording.
  • Deletion: one click on the dashboard removes the response and the underlying audio file.

The point is not that HeySpeak is the only tool that passes. The point is that any tool you pick should be answerable in the same format. If a vendor cannot fill in this list on a phone call, the compliance work is being deferred to you.

Three patterns that look fine and are not

First, "EU data center" without naming the provider. Many tools proxy through a US service for AI inference even if the database is in Frankfurt. Ask where the LLM call resolves, not just where the row is stored.

Second, opt-out training toggles. If the default is "your data trains our model unless you find this setting", you will forget, and your customer's words will be in the next version. The right default is opt-out at the API tier, set by the vendor, before you ever log in.

Third, public playback URLs. A surprising number of sales-call recorders generate a long random URL and call that "secure". A URL you can paste into Slack and play three months later is not secure, it is obscure. Signed URLs with a short expiry are the actual control.

Keep going

Two related pages on running customer conversations without the meeting and without the privacy headache.

Use case

Async customer discovery

How to run 20 discovery interviews in 48 hours, async, with voice answers and AI summaries instead of calendar invites.

Reference

HeySpeak data protection

The full data flow, sub-processor list, and security details in one place. Bookmark this for your DPO.

Common questions

What makes an AI sales tool GDPR-compliant in practice?
Three things, in order: data stays in the EU, the AI provider does not train on your customer data, and you have a signed Data Processing Agreement under GDPR Article 28. After that, the usual hygiene applies: encrypted in transit and at rest, role-based access, audit logs, and a documented sub-processor list. A vendor that ticks all three core boxes and publishes the DPA is already ahead of most of the market.
Is a US-based AI tool ever GDPR-compliant if it has Standard Contractual Clauses?
Technically yes, practically risky. After Schrems II, SCCs alone are not enough. You need a Transfer Impact Assessment showing the data is protected against US government access, which for most US cloud providers is hard to argue in writing. The simpler answer for an EU sales team is to pick a tool whose primary infrastructure already runs in the EU, so the question never comes up.
Why does AI training on customer data matter for sales tools?
Because sales conversations contain things customers said in confidence: pricing pain, internal politics, names of competitors they are evaluating. If those recordings or transcripts feed a foundation model's training set, fragments can resurface in another company's outputs. The clean version is a vendor that contractually states no training on customer data and routes API calls to a provider with the same commitment, like Mistral or an Azure OpenAI tenant with logging disabled.
What should I ask about sub-processors before signing?
Ask for the full list, where each one is hosted, and what data flows to each. A typical AI sales tool has at least four: a database, a file store, an LLM provider, and an email or SMS sender. Each one is a place your customer's voice or transcript can leak. If the vendor will not share the list or hosts critical pieces in the US without justification, that is your answer.
How long should recordings or transcripts be retained?
Default to as short as the use case allows. Sales discovery rarely needs more than 90 days of raw audio, and you can keep summaries longer if you anonymize them. The tool should let you delete a single response on demand and should expire any shareable playback link within an hour, so a forwarded URL cannot be replayed a week later.
Do receivers of an AI sales tool need to consent?
Yes, and the consent has to be informed. The receiver needs to know they are being recorded, who controls the recording, and that AI will transcribe or summarize it. The cleanest pattern is a clear notice on the page where they record, no buried checkbox, no pre-ticked consent. If the tool also forces them to create an account before they can answer, that is more personal data than the use case justifies.
Where does HeySpeak fit in this checklist?
HeySpeak is built EU-first by default. Audio sits in a private Cloudflare R2 bucket, the database is Supabase in Frankfurt with row-level security, and transcription runs on Mistral AI which is 100% EU-hosted. Playback uses signed URLs that expire after one hour. Receivers do not create an account, so no personal data beyond the recording itself is collected. We do not train on customer audio and do not share recordings with third-party analytics.

Try a privacy-first feedback tool in under a minute.

EU-hosted by default. No receiver login. Five free responses to start, no credit card.

Create your first link