AI auto-replies to customer emails: who is liable when it gets it wrong?
Who is liable when AI gets a customer email wrong? What the AI Act and the GDPR require, and which email types can safely be switched to automatic.
Key takeaways
- When an AI replies automatically, your company remains the party the customer deals with; liability depends on the facts.
- There is no dedicated EU AI liability law, so contract law, tort law and the GDPR apply.
- AI Act Article 50 sets transparency obligations for certain AI systems; have it checked whether they cover your emails.
- Only routine emails without a price, deadline or commitment should run automatically; everything else stays a draft.
- Starting with drafts and switching email types on one at a time cuts the risk before it appears.
When AI gets a customer email wrong, your company remains the party the customer deals with. The email leaves your inbox with your name under it, and to your customer it is a statement from your company. That is why an AI assistant only pays off with clear rules: routine can run automatically, while prices, deadlines and commitments stay drafts until a person approves them.
Why do customer emails sit unanswered for hours, and why is autopilot so tempting?
Because the person who could answer is somewhere else. The owner of a small installation business spends the day on site and with customers, then faces a full inbox in the evening. Many of those emails are questions already answered a hundred times: When can you come out? What do you need for a quote? Did my photos arrive?
Those emails are often suitable for carefully bounded automation. A reply in minutes instead of the next evening shortens response times and helps prospects while they are still looking. Our trades page shows how that looks day to day.
The worry follows right behind: what if the AI quotes a customer a price that is wrong? Or confirms a visit that does not exist?
Replying yourself, an autoresponder or an AI assistant with approval: who legally writes the email, and what changes day to day?
Legally, your company writes the email in all three cases. The difference is how much checking sits between the email arriving and the reply going out.
| Replying yourself | Classic autoresponder | AI assistant with approval | |
|---|---|---|---|
| Who writes it | You or your team | One fixed text, the same for everyone | The AI, in your tone, fitted to the request |
| Who checks before sending | You | Nobody, the text never changes | You, with one click or tap on your phone |
| Typical risk | The reply comes too late | Says nothing, but harmless | A wrong detail, if sent without approval |
| Response time | Hours to days | Instant | Draft in under 60 seconds |
A classic autoresponder is limited to a fixed message. That narrows what it can get wrong, but it does not replace review. AI is useful because it says something, and that is exactly why it needs rules an autoresponder never did. What separates a simple helper from a real assistant is covered in our article on the AI virtual assistant for business.
Who is liable when an automatic AI reply states a wrong price, appointment or commitment?
In practice your company remains the party the customer deals with; whether and to what extent you are liable depends on the facts and applicable national law. The EU has no liability rule specific to AI. The Commission withdrew its proposed AI Liability Directive, as recorded in the Official Journal of the EU. The rules that already cover every email apply: contract law, tort law and data protection law. Whether a wrong price binds you, for example, also depends on the exact wording. If it comes to that, your lawyer assesses it.
Passing the loss on to the software maker is not realistic here either. The new Product Liability Directive expressly treats software as a product, but it does not cover damage to data used for professional purposes. A wrong quote is therefore still governed by the applicable contract and national liability rules; whether you have claims against the software provider depends on the facts.
It changes once personal data is involved. If the AI sends customer A's address to customer B, that is a data protection breach. Under GDPR Article 82, you are liable as the controller. The provider running the AI for you is only liable if it breached its own processor obligations or acted against your instructions. You are only off the hook if you prove you are not in any way responsible.
The practical conclusion: the question is not who pays in the end. The question is which emails may go out without anyone checking them.
What do AI Act Article 50 and the GDPR require once replies go out without your approval?
The AI Act calls a company that uses an AI system under its own authority in a professional context a "deployer" (Article 3(4)). Put an assistant on your company inbox and you are one. Article 50(1) sets transparency obligations for certain AI systems. Whether it applies to automatically sent customer emails should be confirmed legally before deployment.
Have two questions checked: whether Article 50(1) covers emails that go out automatically, and whether it also applies to replies a person has approved first. If you want to play it safe in the meantime, the simplest step is a short line in the signature of your automatic replies, for example: "This reply was written by our AI assistant. If anything is unclear, just reply and a person will get back to you."
The GDPR comes in as soon as the AI reads emails containing customer data, which is almost always. You need a data processing agreement with the provider and should know where the data is processed. Verselo provides the DPA at launch, and the assistant runs in a data centre in Frankfurt, separated from every other client. Your data is not used to train AI models. What the server location changes is explained in EU server vs US cloud.
Which email types can run automatically, which always stay drafts, and how do you set this up safely in 14 days?
The rule of thumb: only emails without a price, deadline or commitment run automatically, the kind where a mistake at worst triggers a follow-up question.
| Email type | Setting | Why |
|---|---|---|
| Acknowledgement with the next step | Automatic | Only says what happens next |
| Request for missing measurements, photos or documents | Automatic | Asks, promises nothing |
| Appointment requests and proposed times | Draft at first | An appointment is a commitment you confirm yourself |
| Opening hours, directions, standard facts | Automatic | Fixed facts you set up once |
| Prices, discounts, quotes | Always draft | Every number can be a commitment |
| Deadlines, delivery dates, contract changes | Always draft | This is where obligations arise |
| Complaints and unhappy customers | Always draft | You decide tone and goodwill |
| Newsletters, invoices, system emails | Silent | No reply needed |
The Verselo AI assistant is built exactly this way. It starts in draft mode for everything. In the control cockpit you set each email type to draft, automatic or silent, one at a time. It never sends prices or commitments without asking. An adjustable reply buffer gives you time to stop an automatic email, and if you reply yourself, it holds back. Every rule change is saved and can be rolled back with one click.
Getting there takes 14 days:
- Week 1: we connect Gmail, Microsoft 365 or your IMAP inbox, and the assistant learns your tone from your sent emails.
- Week 2: every reply waits as a draft. You read along, correct and see which email types come out right every time.
- Day 14: you switch the first routine types to automatic, with the notice in the signature. Everything in the bottom half of the table stays a draft.
To be straight with you: if you get a handful of customer emails a day, a good autoresponder plus templates is enough. The assistant pays off when many similar requests come in and you are not at your desk during the day.
Conclusion
When AI auto-replies to your customers, that email comes from your company like any other that leaves your inbox; whether the transparency obligations in AI Act Article 50 cover your emails is worth a legal check. The risk is lower when only routine emails without a price, deadline or commitment run automatically and everything else waits for your approval as a draft. Which of your email types can go automatic is what we work out in a 30-minute call.
Frequently asked questions
Who is liable when an AI sends a customer the wrong answer?
In practice your company remains the party the customer deals with, since the email leaves your inbox under your name. The EU has no dedicated AI liability law: the Commission withdrew its proposal. Existing contract, tort and GDPR rules apply. Whether and to what extent you are liable depends on the facts and applicable national law, and that is a question for your lawyer.
Do I have to tell customers that an AI answered their email?
Article 50(1) of the AI Act sets transparency obligations for certain AI systems. Whether it covers your automatically sent emails should be confirmed legally before deployment. A short line in the signature of your automatic replies is the simplest precaution.
If the AI sends customer data to the wrong person, is the assistant provider liable?
You are liable first, as the controller. Under GDPR Article 82, a processor is only liable if it breached its own processor obligations or acted outside or against your instructions. That is why a data processing agreement belongs in place before you start.
Which customer emails can I safely switch to automatic?
Acknowledgements, requests for missing documents and fixed facts such as opening hours. Appointment requests stay drafts at first, as do prices, discounts, deadlines, complaints and anything contractual.



