Legal Review Checklist for an AI Email Tool Contract

A security questionnaire tells you how a vendor protects your data. The contract tells you what they are allowed to do with it, and what you can do when something goes wrong. A legal review checklist for an AI email tool is where those two documents meet: the Data Processing Agreement, the terms of service, and the order form.

This is the contract-side review, written for the person who reads the DPA before signing rather than the person who runs the trial. It is a set of questions to ask any vendor, not legal advice.

Bring your own counsel and procurement to the actual redline. The point of this page is to make sure you walk into that redline knowing which clauses decide the outcome.

The short answer#

Read the DPA first, because it binds the vendor's conduct rather than describing its intentions. Check three things in it before anything else: that the vendor processes your mail only on your documented instructions, that it will not use your content to train models, and that sub-processors are named with notice before any change.

Then read four commercial clauses in the main agreement: who owns the output the tool generates, whether the vendor indemnifies you for that output, how high the liability cap sits against a realistic breach, and what happens to your data when the contract ends.

None of these show on a pricing page. All of them are in the contract, and every one is negotiable before signature — which is the only time you have any leverage.

The clauses that actually matter#

An AI email tool is a data processor that reads your entire inbox and, increasingly, acts on it. That makes a handful of clauses carry most of the risk. Here is the shortlist, in the order counsel usually reads them.

• No-training term — your message content is never used to train or improve the vendor's or any sub-processor's models, stated as a binding clause, not a website line.
• Sub-processors and change notice — who else touches your mail, including which model providers, and how much notice you get before that list changes.
• Documented instructions — the vendor acts only on your instructions, the GDPR Article 28 baseline for any processor.
• Output ownership — whether the drafts and summaries the tool produces belong to you, in writing.
• AI indemnity — whether the vendor covers you if its output infringes a third party's rights.
• Liability cap — the ceiling on what the vendor pays if it causes you loss, read against what a mailbox breach would actually cost.
• Termination and data return — how you get your data out, in what format, and when every copy is deleted.
• Governing law and data residency — whose courts and which region's servers, which together decide your remedies and your regulatory exposure.

The no-training term and sub-processor notice#

Most AI email tools send your message content to third-party model providers to generate a draft or a summary. So two clauses do the heavy lifting: what those providers may do with your content, and who they are.

The no-training term should say, in the DPA and not only on the marketing site, that your content is never used to train or fine-tune the vendor's or any sub-processor's models. Pair it with zero-retention inference terms, so the model provider keeps nothing after it returns the answer. A sentence on a features page is a promise; a clause in the DPA is a term you can enforce.

The sub-processor clause is the other half. Under the GDPR's general written authorisation model, a processor may add sub-processors but must inform you of intended changes and give you a chance to object. Ask for the named list — including the model providers — and a notice period before it changes. 'Affiliates and service providers' is not a list.

Ownership, indemnity, and the liability cap#

The commercial clauses decide who carries the cost when the AI is wrong. Start with ownership: most vendor terms assign the output the tool generates to you, but confirm the contract says so, rather than granting the vendor a broad licence over your drafts and inputs. Whether purely AI-generated text is itself protectable by copyright is a separate and unsettled question your counsel can speak to.

Next, indemnity. Ask whether the vendor indemnifies you against third-party intellectual-property claims arising from its output, and read the conditions — some indemnities fall away if you disable a safety feature or edit the output heavily. An indemnity you cannot practically satisfy is a clause, not a protection.

Then the liability cap. SaaS contracts commonly cap the vendor's total liability at the fees you paid in the prior twelve months. Weigh that number against what a breach of a mailbox full of client correspondence would actually cost you, then ask for a carve-out — a higher cap or none — for data-protection breach, breach of confidentiality, and the IP indemnity.

The contract-review scoring table#

Score each clause as you read it. Weights run to three; treat a red on any triple-weighted clause as a reason to walk, not something a strong showing elsewhere averages away.

Read the low-scoring rows, not the total. A tool can have an excellent product and a contract that quietly leaves your data-return rights and your liability exposure worse than they need to be — and those are exactly the clauses nobody notices until they matter.

Worked example: one clause, reviewed the way counsel would#

Here is a single termination clause annotated the way a reviewer would mark it. The clause is not wrong; it is thin, and the gaps are what you fix in the redline.

The Data Act (Regulation (EU) 2023/2854) entered into force on 11 January 2024 and became applicable on 12 September 2025. For EU buyers it adds rules that let customers switch between data-processing services and reduce barriers to porting data — leverage worth naming in a termination discussion. Check the current text and dates on EUR-Lex, since implementation guidance is still settling.

Red flags in an AI email tool contract#

• The DPA is 'available on request' and never arrives before signature. If you cannot read it, you cannot review it.
• No sub-processor list, or one that names categories such as 'cloud providers' instead of companies. You cannot assess who you cannot see.
• 'We do not train on your data' appears on the website but nowhere in the contract. Marketing is not a term.
• The liability cap covers a data breach and there is no carve-out. The ceiling is set below the loss.
• Termination gives you an archive, not an export — a format no other tool can open.
• The vendor may change sub-processors or terms 'at any time' with no notice and no right to object.
• Governing law is a jurisdiction where you have no realistic way to enforce the contract.

What we would check, and where AI Emaily fits#

So which tool clears this checklist? Weight it by who you are. If your procurement policy requires a completed SOC 2 Type 2, a signed HIPAA Business Associate Agreement, or ISO/IEC 27001 before you can begin, then a vendor that already holds the one you need clears your gate faster, and you should buy that one. Be honest with yourself about which of those is a hard requirement and which is a preference.

We build AI Emaily, an AI email client for Gmail, Outlook, and IMAP, so here is our own contract-side scope, plainly. We publish a Data Processing Agreement, do not train on your mail, use zero-retention inference terms with model providers, keep a published sub-processor list, request minimum OAuth scopes, and offer a bring-your-own-key option for sensitive work. OAuth tokens and any provided keys are envelope-encrypted, and we completed Google's independent CASA restricted-scope assessment for our Gmail scopes, including gmail.modify, in July 2026. Your mail keeps living in the Gmail, Outlook, or IMAP mailbox you own — we connect to it rather than becoming your mail host.

What we do not have: a completed SOC 2 report — our security page lists it as on the roadmap, not done — no signed HIPAA BAA, and no ISO/IEC 27001 or 42001 certification. If any of those is a non-negotiable requirement, we are the wrong pick today, and we would rather say so than imply a status we have not earned. We are the right pick if you evaluate on terms you can read and test now: the DPA and its no-training clause, an approval gate before any send, an audit log with undo behind every action, and export you run yourself. See the full picture on our homepage at aiemaily.com, the specifics at aiemaily.com/security, and the plans, including the 7-day free trial, at aiemaily.com/pricing. Have your counsel and procurement review the actual terms before you sign — this checklist frames the questions; it does not answer them for your situation.