AI email assistant: consent, audit trail, confidentiality
An assistant that reads and writes your email touches everything: client contracts, employee files, competitor quotes. Three things must be written down before go-live - the approval rule, the audit log, the confidentiality perimeter. Here is the specification, clause by clause, and the four regulators to know.
The short answer
An AI email assistant works when three things are written down before go-live: who approves a message before it leaves, what trace remains of every action, and which data leaves the company. Skip them and the tool saves minutes a day while costing one incident a year. The GDPR already governs the processing and the processor relationship[1]; the EU AI Act adds an AI literacy duty, in application since 2 February 2025[2][3], and transparency rules applicable since 2 August 2026[3].
What follows is a specification, not legal advice: every requirement points to the text or the authority behind it.
What an AI email assistant actually does
Three levels of autonomy hide behind the same word, and they do not carry the same risk.
- Level 1 - read, sort, summarise. The assistant labels, files, summarises a thread, extracts a deadline. Nothing reaches a third party if the model is hosted by a provider under contract. The risk is confidentiality.
- Level 2 - propose. The assistant drafts a reply, offers a slot, pre-fills a CRM record. A human approves. The risk is an error copied without rereading.
- Level 3 - send. The assistant replies on its own, confirms a meeting, follows up. The risk becomes contractual and commercial: a price quoted, a deadline promised, a commitment made.
A tool sold as level 2 that fires automatic acknowledgements is in fact level 3 for those messages. So the question to put to the vendor is not does the AI send? but list every case where a message leaves without a human click.
Consent: who decides a message leaves
An assistant writing on behalf of the company needs an explicit approval rule, and that rule has to live inside the tool, not in a document on a shared drive. Three models exist, and they combine.
- Per-message approval. Every send waits for a click. Slow, safe, right for clients and sensitive subjects.
- Per-template approval. A reply template is approved once, then the assistant sends it without a new click as long as the content, the recipient and the amount do not change. This is the useful compromise for acknowledgements and B2B appointment setting.
- Per-perimeter approval. The assistant is free inside a closed list of recipients or subjects, and silent everywhere else.
Money (price, discount, refund) - a deadline or committed date - contracts and signatures - deletion of data - regulation and legal matters. An email assistant that touches any of the five hands control back to a named human, never to a shared inbox. It is the cheapest rule to set and the most expensive one to forget.
The audit trail: the log to demand
"The AI replied" is not evidence. A log is. The GDPR puts the burden of demonstrating compliance on the controller[1] and obliges the processor to assist; without a readable log, that demonstration rests on the vendor's word. Demand an export, not a screenshot.
| What the log must carry | Why |
|---|---|
| Timestamp and message identifier | Tie a send to a thread, recover the real order of events |
| Model and version used | Answers change when the model changes |
| Text as actually sent | A draft is not a send |
| Who approved, or which approval rule applied | Separate a human action from an automatic one |
| Data passed to the model | Measure what left, answer a subject access request |
| Log retention period | A log wiped after 7 days does not cover a dispute at 3 months |
| Self-service export (CSV or API) | Do not depend on a support ticket to prove anything |
A vendor who ticks all seven lines saves you days when a client contests a message. One who ticks three is selling a tool, not a service.
Confidentiality: what leaves the company
A business inbox almost always holds personal data, and often third-party data: a client contract, an employee complaint, a competitor's quote. Four requirements have a text behind them.
- A processing agreement under Article 28 GDPR[1], naming every sub-processor. An AI vendor hosting with a third party must name it.
- No reuse for training without a legal basis and without information. The EDPB set out in 2024 how to assess the anonymity of a model and the lawfulness of training data[4], and the French CNIL published recommendations on legitimate interest applied to AI system development[10]. In an email assistant contract the clause to look for is short: no customer content is used to train or improve a model.
- Minimisation (Article 5(1)(c) GDPR[1]): the assistant does not need five years of archives to draft today's reply.
- Location and transfers: if processing leaves the EU, Chapter V GDPR applies[1]. The CNIL factsheet for small businesses makes one very concrete point: whatever you type into a prompt goes to the provider[6].
To size the exposure before negotiating, a working assumption - replace it with your own numbers:
Illustrative assumption: 12 inboxes x 60 messages x 30 days = 21,600 messages exposed at any given moment. Recalculate on your real volume before deciding anything.
That one line settles most arguments: cutting retention from 30 days to 2 divides the surface by fifteen and changes nothing about answer quality. Two organisational reflexes then weigh as much as the contract: keep whole inboxes out of scope (HR, management, legal), and treat a data breach as a planned scenario - the GDPR provides for notification to the supervisory authority within 72 hours where it is required[1].
The same reasoning applies to outbound mail, where the risk is deliverability rather than confidentiality: see our guide to AI-assisted B2B email prospecting.
Four countries: who supervises, under which text
CyberNovaLabs.io works in the Netherlands, Belgium, France and Luxembourg. The GDPR and the AI Act are the same there; the competent authority, the language of requests and the available guidance are not.
| Country | Data protection authority | Published AI guidance |
|---|---|---|
| France | CNIL | Factsheets on using generative AI in small businesses[5][6] and recommendations on AI system development[10] |
| Belgium | APD / GBA | Artificial intelligence theme and a brochure on AI systems and the GDPR[7] |
| Netherlands | Autoriteit Persoonsgegevens | Algorithms & AI theme[8] |
| Luxembourg | CNPD | Artificial intelligence thematic dossier[9] |
On the AI Act side, two duties already reach an email assistant: AI literacy - people trained on the tool they use - in application since 2 February 2025[2][3]; and the transparency rules, applicable since 2 August 2026[3], which cover among other things informing people interacting with an AI system and marking certain generated content[2]. The full timeline is published by the European Commission[3] and unpacked in our AI Act guide for SMEs in 2026. Do not carry a rule from one country to another without checking it: that is the only sentence in this article that holds for all four.
The specification in 12 questions
Put them to the vendor in writing, before signing. An evasive answer counts as a no.
- In exactly which cases does a message leave without a human click?
- How do we revoke that autonomy in one minute, without a support ticket?
- Which sub-processors, in which countries?
- Is our content used to train a model? Written clause.
- How long is content retained, and where?
- Does the log carry the seven lines in the table above?
- Can we export that log ourselves?
- Which inboxes can be excluded, and at what granularity?
- Who is informed, and how fast, in case of a data breach?
- Which model version is running today, and how are we told when it changes?
- Does the recipient know an AI system is involved, and by what means?
- What happens at the end of the contract: return, deletion, proof of deletion?
Test this specification over two weeks on a single inbox at level 2, then open level 3 one template at a time. Same method as in the five tasks to automate first in an SME: a narrow perimeter, a measurement before, a measurement after.
At CyberNovaLabs.io
We install email assistants that stay at level 2 by default, with an exportable log and human approval on the five sensitive subjects, then open level 3 one template at a time. Perimeter, retention and approval rule are written down before the first message is handled. See our AI and business automation offer, or tell us your email volume and your sensitive inboxes: we reply with a costed perimeter. Services quoted on request; best-efforts obligation.
Sources
- Reglement (UE) 2016/679 (RGPD) - texte consolide, EUR-Lex
- Reglement (UE) 2024/1689 (AI Act) - texte, EUR-Lex
- Commission europeenne - Regulatory framework for AI (calendrier d'application)
- CEPD / EDPB - Avis 28/2024 sur certains aspects de protection des donnees lies aux modeles d'IA
- CNIL - Utiliser l'IA generative dans les TPE et PME
- CNIL - Fiche IA 4 : quelles precautions prendre (PDF)
- Autorite de protection des donnees (Belgique) - Intelligence artificielle
- Autoriteit Persoonsgegevens (Pays-Bas) - Algoritmes & AI
- CNPD (Luxembourg) - Dossier thematique Intelligence artificielle
- CNIL - Recommandations sur le developpement des systemes d'IA et l'interet legitime
Qualified meetings, no lock-in.
Criteria in writing before launch, pay per meeting or monthly, stop with a simple email.
Read next
NGINX: five critical flaws patched urgently (18/11/2026)
Five critical flaws in NGINX, the software receiving traffic for more than 225 million websites, were fixed on 18 November 2026. Four authorities warned the same day. Your site can be exposed without having a single defect of its own.
Read the article →November 20, 2026 · 8 min readNIS2 Luxembourg: are estate agencies and syndics in scope?
Real estate appears in neither Annex I nor Annex II of the Luxembourg Act of 5 May 2026. But your supervisory authority already exists: the AED, on money laundering. Position as of 20 November 2026.
Read the article →November 12, 2026 · 8 min readNIS2 Belgium: are law firms in scope?
Legal services appear in neither Annex I nor Annex II of the Belgian Law of 26 April 2024. A Belgian law firm is therefore not a NIS2 entity, whatever its size. What remains: professional secrecy, the GDPR, and your NIS2 clients' security questionnaire.
Read the article →November 11, 2026 · 7 min readSMS sign-in abused in Belgium: the CCB warning (10/11/2026)
No patch to install: a setting. The CCB observed successful attacks on several Belgian organisations where SMS was the first factor — a number, a code, no password. Microsoft stops delivering those codes on 1 February 2027.
Read the article →November 10, 2026 · 7 min readWordPress flaw exploited during the update itself (04/11/2026)
WordPress 7.1.4 has fixed, since 28 October, a remote code execution triggerable during a core upgrade. CISA flagged it as exploited on 3 November, NCSC-NL warned on 4 November. For an SME, patching is not enough: you also have to look back.
Read the article →November 10, 2026 · 9 min readNIS2 Netherlands: are accounting firms and fiduciaries in scope?
The Dutch law implementing NIS2 took effect on 15 August 2026. It does not name accountancy. Even so, a Dutch accounting firm or fiduciary inherits duties through three doors: the GDPR, the Wwft, and the contracts of clients who are themselves in scope. Here is what applies, and what does not.
Read the article →