NevTan Mail is secure, ad-free business email on your own custom domain, with a built-in calendar, meetings, admin controls, and a privacy promise that your mail is never used to train AI. This guide gives you a way to answer that question about any provider — including ours — by reading their actual terms rather than their marketing. You will learn the three different things "AI and your email" can mean, why the two biggest viral privacy scares of recent years were both wrong, and the nine questions to ask before you hand a provider your business correspondence. For the wider product picture, see our complete NevTan Mail platform guide.
"Training AI on your email" is one of three distinct things, and they get conflated constantly. Processing means your provider's systems read message content to deliver features like spam filtering and search — every provider does this, and it is not training. Grounding means a generative assistant reads your mailbox at the moment you ask it something — the data informs one answer and is not retained in a model. Training means your message content becomes part of the dataset used to adjust model weights, permanently and irreversibly. Only the third is what people mean when they are worried. Check the provider's Data Processing Agreement or Product Terms for an explicit statement on training, confirm whether it covers all plans or only paid tiers, and check whether the commitment is contractual or only a blog post.
The Three Things That Get Confused
Almost every argument about email and AI is two people meaning different things by the same words. Separate them and the question becomes answerable.
1. Processing (content analysis). Your provider's systems read your messages to make the product work. Spam and virus filtering requires reading message bodies. So does search indexing, threading, attachment previews, calendar invite detection, and categorising mail into tabs. This has been true since webmail existed and it is not controversial — it is the only way filtering can function. If your provider could not analyse message content, it could not tell you why your emails go to spam or stop anyone else's from reaching you. Content analysis for product function is not model training.
2. Grounding (retrieval at query time). You ask an assistant "what did Priya say about the deadline?" and it reads your recent messages to answer. Your data informs that single response and then the exchange ends. Nothing about the model changed. This is how most inbox assistants work, and it is also where the genuinely important question sits — not "does it read my mail" (it must, to answer) but "what happens to the prompt and the retrieved content afterwards?"
3. Training (model improvement). Your message content is added to a dataset used to adjust the weights of a model that will then serve other people. This is the irreversible one. Once content is in a training set, it cannot be meaningfully extracted, and the resulting model may reproduce patterns from it. This is what people mean when they ask the question in this article's title, and it is the one to get a clear answer on.
A fourth category sits alongside these: human review. Some providers sample interactions for quality or abuse monitoring, which can mean a person reads a prompt and response. Worth asking about separately, because a "we don't train on your data" commitment does not automatically exclude it.
What Actually Happened in the Two Big Scares
Both major email-AI privacy panics of recent years turned out to be based on a misreading, and understanding why is the fastest education in how to read these policies.
Gmail, November 2025. A post claiming Gmail users had been "automatically OPTED IN to allow Gmail to access all your private messages & attachments to train AI models" went viral, reaching millions of views, and was picked up by security vendors — including Malwarebytes, which subsequently issued a correction. The post pointed at Gmail's "Smart features and personalisation" setting.
Google responded directly on 21 November 2025, stating that the reports were misleading, that it had not changed anyone's settings, that Gmail Smart Features have existed for many years, and that it does not use Gmail content to train the Gemini AI model. Spokesperson Jenny Thomson gave the same statement to The Verge.
The confusion was category one being mistaken for category three. Smart Features genuinely do analyse your email — that is how you get package tracking, smart replies, tab sorting, and flights added to your calendar. That analysis powers Gmail's own functions; it is a different thing from contributing to generative model training.
Microsoft 365, November 2024. An almost identical cycle. Posts claimed Microsoft had "slyly turned on an opt-out feature that scrapes your Word and Excel documents to train its internal AI systems," pointing at the "optional connected experiences" setting. Microsoft told Reuters: "These claims are untrue. Microsoft does not use customer data from Microsoft 365 consumer and commercial applications to train foundational large language models," adding that connected experiences enables features like co-authoring and cloud storage and has no connection to how it trains large language models.
The lesson is not "stop worrying." It is that a setting labelled "analyse your content" tells you nothing about training, in either direction, and you have to go to the contractual terms to find out. Which is exactly what the checklist below does.
The Nine Questions to Ask Before You Sign Up
Work through these on any provider's site. Most can be answered in fifteen minutes.
1. Is there an explicit statement about training, in the terms rather than a blog post?
You are looking for a sentence that names model training directly. Microsoft's, for example, lives in its Product Terms under "Use of Content for Training" and commits that its generative AI services will not use Customer Data to train any generative AI foundation model except on the customer's documented instructions. That is a contractual commitment in a document that forms part of the agreement.
A statement in a marketing page or a tweet is weaker. Both can be true and sincere; only one is enforceable. Check the terms of service and the privacy policy of any provider you are evaluating, not the homepage.
2. Does the commitment cover your plan, or only enterprise?
This is the most common gap and the one that catches small businesses. Several providers make firm no-training commitments to commercial and enterprise customers while treating consumer and free-tier accounts under a different policy. If you are a five-person company on a low-cost or free plan, confirm which document applies to you.
3. Does it cover every AI feature, or only the flagship assistant?
A commitment about "Copilot" or "Gemini" may not extend to a separate smart-compose, summarisation, or categorisation feature built on a different model. Ask whether the statement covers all AI functionality in the product.
4. Is training opt-in, opt-out, or off entirely?
These are very different. "Off entirely" means no toggle exists because the practice does not happen. "Opt-out" means it happens by default until you find the setting — and that setting may live in more than one place. "Opt-in" is the middle ground. If there is a toggle, find out what the default is for a newly created account, because that is what your colleagues will have.
5. What happens to prompts and outputs, as distinct from message content?
Your messages and your interactions with the assistant are often governed separately. A provider may never train on your mailbox while still using your conversations with its chat assistant to improve that assistant. Both questions need asking.
6. Is human review part of the pipeline?
Ask whether any sampling of content, prompts, or outputs is reviewed by people for quality or abuse monitoring, and whether that is contractually bounded.
7. Where is the data processed and stored?
Relevant for GDPR, for data residency obligations, and for anyone with sector-specific compliance duties. Note that a provider's regional data commitments may carve out specific AI services — Microsoft, for instance, documents that certain third-party models deployed in its Copilot offerings are excluded from the EU Data Boundary and in-country processing commitments. Carve-outs like that are easy to miss and important to find.
8. What happens to your data when you leave?
A no-training promise is worth less if you cannot get your mail out. Check for a full-mailbox export in a standard format. NevTan Mail exports a complete .mbox in Google Takeout format at no cost on every plan, specifically so leaving is possible — an exit route is itself a privacy feature. Our guide to why teams are switching to modern webmail covers what else to check before a move.
9. What else is connected to the mailbox?
Your provider's policy governs your provider. Every integration you authorise — a CRM, a scheduling tool, an automation platform — operates under its own terms, and may have its own training provisions. If you have connected tools via API, each one is a separate policy to read; our comparison of IMAP, SMTP, and API integrations covers what access each method grants. The same applies to calendar and task data where those share a workspace, as described in email, calendar, and tasks in one workspace.
AI Features and No Training Are Not a Trade-Off
A common assumption is that refusing to train on customer data means going without AI features. It does not. Grounding — category two above — lets an assistant read your mailbox to answer your question without that content ever entering a training set.
NevTan Mail includes an AI assistant on the free plan, alongside business email on your own domain for up to ten users:
NevTan Mail Free plan pricing card, showing: free for every user with no card needed, 1 GB storage per user, up to 10 users per workspace, business email on your domain for up to 10 users, NevTan AI assistant, and NevTan Office
NevTan Mail's free plan includes the AI assistant. Our privacy commitment is that your mail is never used to train AI — the assistant works by reading your mailbox to answer what you asked, not by contributing your correspondence to a training set.
Our features page sets out what each plan includes, and the principles behind the product explain how we think about the trade-off. The broader shift this sits inside is covered in how AI is changing the inbox.
Where to Look: Reading the Documents
Four places, in order of how much weight to give them.
Data Processing Agreement or Product Terms. The strongest source. These form part of your contract. Search the document for "train", "training", "model", and "machine learning". If the provider is a processor under GDPR, this is where its obligations are defined.
Privacy policy. Useful, but written for a general audience and often ambiguous. Phrases like "to improve our services" are doing a lot of unspecified work; "improve our services" can cover everything from bug-fixing to model training.
Product-specific AI documentation. Many providers maintain a dedicated privacy FAQ for their AI features, which is often clearer than either of the above. Microsoft's Copilot documentation, for example, states that prompts, responses, and data accessed through Microsoft Graph are not used to train foundation LLMs.
Support articles and blog posts. The weakest. They reflect current practice and current intent, and can change without notice to you. Useful for understanding; not something to rely on.
Green flags and red flags
Green flag | Red flag |
|---|---|
Training commitment appears in the DPA or Product Terms | Only on a marketing page or in a support article |
Policy is identical across free and paid tiers | Separate, weaker terms for consumer or free accounts |
Commitment names all AI features, not one product | Covers only the flagship assistant |
Training is absent, not a default-on toggle | Opt-out buried in settings, enabled by default |
Human review is addressed explicitly | Silent on human review entirely |
Free full-mailbox export in a standard format | Export is paid, partial, or proprietary |
Changes to AI terms are notified in advance | Terms may change "with continued use" |
No advertising business model | Revenue depends on profiling users |
That last row deserves emphasis. The structural question underneath all of this is how the provider makes money. A provider whose revenue comes from subscriptions has no commercial reason to extract value from your correspondence. A provider whose revenue comes from advertising has a standing incentive that no individual policy fully neutralises. This is not an accusation about any specific company's current practice — it is a reason to read the terms of ad-funded providers more carefully, and to re-read them when they change.
Real Example: How to Run This Check in Fifteen Minutes
A consultancy handling client financial data needed to document, for its own compliance file, whether its email provider could train on client correspondence.
They did not email sales. They opened the provider's Data Processing Agreement, searched for "train", and found a clause. They then checked three things about that clause: which document it lived in, whether it applied to their subscription tier or only to enterprise agreements, and whether it covered all AI features or named one product.
The clause was contractual, applied to their tier, and covered generative AI services generally. They screenshotted it with the date, saved the PDF of the DPA version they had agreed to, and filed both. Total time: under twenty minutes.
Two things made this work. First, they went to the contract rather than the marketing. Second, they saved a dated copy — because terms change, and "what did the agreement say when we signed" is the question that matters if it is ever disputed. Set a calendar reminder to re-check annually; policy changes are frequently announced by email that nobody reads.
Common Mistakes to Avoid
Mistake 1: Treating content analysis as evidence of training. Spam filtering, search, and threading all require reading message content. That is the product working. Solution: Judge the training question on the training statement, not on the existence of content analysis.
Mistake 2: Trusting a marketing page over the contract. Both viral scares above involved people reading a settings label instead of the terms. Solution: Search the DPA or Product Terms for "train".
Mistake 3: Assuming the enterprise policy applies to your plan. It frequently does not. Solution: Confirm which document governs your specific tier.
Mistake 4: Checking once and never again. AI terms have changed repeatedly across the industry. Solution: Save a dated copy and diarise an annual review.
Mistake 5: Forgetting integrations. Your provider's policy does not bind the tools you have connected to your mailbox. Solution: Audit the authorised apps list and read each one's terms.
Mistake 6: Confusing encryption with privacy from your provider. TLS protects mail in transit between servers; encryption at rest protects stored data from certain classes of compromise. Neither prevents a provider from processing content it holds the keys to. Solution: Treat encryption and data-use policy as separate questions — our guides to email encryption for business and TLS and phishing protection cover what each actually protects.
Mistake 7: Assuming free means exploited. Free tiers exist for several reasons, including straightforward customer acquisition. Solution: Read the terms rather than assuming — but do check whether the free tier is governed by the same document as paid plans. Our guide to cheap business email for startups covers evaluating low-cost and free options properly.
Frequently Asked Questions
Is Google training Gemini on Gmail content?
Google has stated publicly that it does not. On 21 November 2025, responding to viral claims that users had been opted into AI training, Google said the reports were misleading, that it had not changed anyone's settings, that Gmail Smart Features have existed for many years, and that it does not use Gmail content to train the Gemini AI model. Smart Features do analyse email content, but to power Gmail's own functions such as smart replies and calendar detection, which is separate from generative model training.
Does Microsoft train AI on Microsoft 365 data?
Microsoft has denied this and made it contractual. In November 2024 it told Reuters that claims it used Microsoft 365 customer data to train models were untrue, and that the "connected experiences" setting has no connection to how it trains large language models. Its Product Terms commit that Microsoft Generative AI Services will not use Customer Data to train any generative AI foundation model except on the customer's documented instructions, and its Copilot documentation states that prompts, responses, and data accessed through Microsoft Graph are not used to train foundation LLMs.
What is the difference between an AI assistant reading my email and training on it?
Reading at query time — grounding — means the assistant retrieves relevant messages to answer the question you just asked, then the exchange ends without altering any model. Training means your content is added to a dataset used to adjust model weights, permanently, in a model that will serve other users. The first is how most inbox assistants function; the second is the irreversible one. Ask about prompt and output retention separately, since those are often governed by different terms than your mailbox.
Can I use AI email features without my data training a model?
Yes. The two are not linked. A grounded assistant can read your mailbox to answer a question without that content entering a training set. NevTan Mail includes an AI assistant on the free plan and commits that your mail is never used to train AI.
Does spam filtering mean my provider reads my email?
Yes, and it has to. Spam and virus filtering works by analysing message content, headers, and sender reputation. There is no way to filter mail without processing it. This is content analysis for product function, and it is distinct from model training.
Where do I find my provider's AI training policy?
In order of reliability: the Data Processing Agreement or Product Terms, which form part of your contract; any product-specific AI privacy documentation; then the privacy policy; then support articles and blog posts. Search those documents for "train", "training", "model", and "machine learning".
Does GDPR stop my provider from training AI on my emails?
Not automatically. GDPR governs the lawful basis, purpose limitation, and the controller–processor relationship, and a provider acting as a processor is contractually restricted to processing on the controller's documented instructions. In practice, whether training is permitted turns on what your specific agreement says, which is why the DPA is the document to read.
Should I avoid free email tiers for business?
Not on principle — but check whether the free tier is governed by the same terms as paid tiers, since this is the most common place where privacy commitments differ. Also weigh the structural question: an ad-funded provider has a standing incentive to profile users that a subscription-funded one does not.
Email That Is Never Used to Train AI
The question this article opens with should have a one-sentence answer from any provider you are considering, and it should appear in their terms rather than their advertising.
Ours: your mail is never used to train AI. No ads, no data selling. You can read the specifics in our privacy policy and our terms of service, and we would encourage you to read them with the nine questions above in hand rather than taking the sentence on trust — that is the entire point of this guide.
Alongside that you get secure business email on your own domain, a calendar and meetings built into the inbox, role-based admin controls, TLS-encrypted connections with content encrypted at rest, guided SPF/DKIM/DMARC setup covered in our guides to domain authentication and deliverability best practices, and a free full-mailbox export so you are never locked in. Teams running campaigns alongside their business mail should also read transactional vs marketing emails for how to keep the two separated.
For more on privacy, security, and getting your mail delivered, browse the NevTan Mail blog.
Ready for an inbox that does not treat your correspondence as training data? Start free with NevTan Mail — ten users on your own domain, no card needed.
