Blog/ Buyer guides

35 Questions to Ask an AI Email Vendor Before You Buy

Nafiul HasanNafiul Hasan· 16 min read
AI Emaily blog cover for questions to ask an AI email vendor, showing a buyer's question bank for vetting AI email software before purchase

The short answer

Take thirty-five questions in seven groups: data handling and training, autonomy and safety, accuracy and failure modes, provider coverage, team and admin, commercial terms, and continuity and exit. Judge the shape of each answer, not its warmth. Not yet is fine, I do not know is fixable, and I will not say is the one that ends the evaluation.

35 questions to ask an AI email vendor, each with why it matters and the weak answer to listen for, including the eight we answer badly.

On this page
  1. 01The short answer
  2. 02How to run this without sounding like an auditor
  3. 03Group 1: Data handling and training
  4. 04Group 2: Autonomy and safety
  5. 05Group 3: Accuracy and failure modes
  6. 06Group 4: Provider coverage
  7. 07Group 5: Team and admin
  8. 08Group 6: Commercial terms
  9. 09Group 7: Continuity and exit
  10. 10The four shapes of a weak answer
  11. 11When several weak answers land at once
  12. 12Turn the list on us: where AI Emaily answers badly
  13. 13The follow-up email is the deliverable

These are the questions to ask an AI email vendor out loud, in a live conversation, while someone from their side is on the line and can be pinned to an answer.

A scorecard is what you test yourself in a trial. A security questionnaire is a document their compliance team fills in at its own pace. A sales call is the one setting where you hear what a vendor says before anyone edits it.

So every question below carries two things a bare list leaves out: why it matters, and what the weak answer sounds like. The weak answers are the useful column. We build AI Emaily, and eight of these are ones we answer badly.

The short answer#

Ask across seven groups: data handling and training, autonomy and safety, accuracy and failure modes, provider coverage, team and admin, commercial terms, and continuity and exit.

Then grade the shape of the reply rather than its content. Vendors rarely lie outright. They answer an adjacent, friendlier question, and the buyer nods because something was said.

One weak answer is a bad day. Two inside the same group is a pattern, and the pattern is what you are buying.

How to run this without sounding like an auditor#

You are not reading thirty-five questions down a phone line. Pick twelve and leave room to push on the ones that wobble.

  1. 1

    Ask the default, not the capability

    Every product can be configured safely. Ask what ships switched on, and what happens if nobody opens the settings page.

  2. 2

    Follow every claim with show me

    In the product, on this call, on a mailbox with something awkward in it. A live demonstration is worth ten feature-page bullets.

  3. 3

    Send the written follow-up within a day

    Restate their answers in your own words and ask them to correct anything wrong. Privacy pages get edited; a saved reply does not.

  4. 4

    Score the shape, not the warmth

    The friendliest call in a process is often the one that leaves you with the least in writing.

Thirty-five questions is not a call agenda

Read as a list this becomes an interrogation nobody answers well, honest vendors included. Ask twelve live, send the rest by email where a written answer is the whole point, and keep three back for the second call.

Group 1: Data handling and training#

This group decides whether the rest of the evaluation is allowed to happen. The recurring evasion is answering a narrower question than the one asked, so watch questions 2 and 3.

Ask thisWhy it mattersA weak answer sounds like
1. Which sub-processors see the contents of my messages, and where is that list published?GDPR Article 28(2) requires written authorisation for sub-processors and notice when they change. Without a published list, neither is possible.We use industry-standard infrastructure providers.
2. Is my mail used to train, fine-tune or evaluate any model, including one that serves only me?We do not train on your data usually covers the foundation model only. Fine-tuning, evaluation sets and human review are separate paths.We would never sell your data. Different question.
3. Do you hold zero-retention terms with every model provider, and is that in the DPA?The vendor's own promise does not bind the API behind it. That commitment has to exist one layer down, in writing.We use a major model provider, so it is covered.
4. Who at your company can read a message of mine, and is that access logged?Some human access is unavoidable for support work, so you want the control described rather than denied. Google's Limited Use rules restrict it.Nobody here can see your email.
5. After I delete a message, what survives in caches, indexes, embeddings and logs, and for how long?Delete removes the message record. Derived data often sits outside the delete path, which is exactly where a data subject request lands.Deleted is deleted.
6. Will you sign a DPA, and a BAA if my mail carries health information?HIPAA requires a written business associate contract before a vendor handles protected health information. No BAA means regulated mail is out of scope.We are HIPAA compliant, with no BAA on offer.

Get this group in writing

Google's Workspace API policy prohibits using Workspace user data to create, train or improve a model beyond that user's own personalised one. Any vendor on Gmail scopes is bound by it, so a refusal to restate it in an email tells you about the vendor, not the rule.

Group 2: Autonomy and safety#

Everything here is about what happens when you are not looking. Anchor each question to the shipped default: a configurable product and a safe product are different claims.

Ask thisWhy it mattersA weak answer sounds like
7. At the default setting, list every action the product takes without me clicking anything.The entire risk profile in one answer, and the default is the setting most colleagues keep forever.It is fully autonomous. Or: it only does safe things.
8. Where is the setting that changes that list, and is it per account or global?Autonomy should be a dial per mailbox. A personal account and a client-facing one do not carry the same risk.It adapts to how you work.
9. On send, does a draft wait for my click? What ships as the default?Ask about the default, not the capability. Almost every tool can be configured safely and rather fewer arrive that way.You can switch auto-send off. Which means it is on.
10. Which actions can I undo, which are permanent, and how long is the window?Archive, label and unsubscribe are recoverable at very different costs. A sent message is not recoverable at any.You can always delete it afterwards.
11. What stops a crafted incoming email from instructing the agent?The agent reads mail from strangers, then acts. OWASP lists prompt injection and excessive agency separately; an email agent meets both.A pause, then: the model would not fall for that.
12. If it sends something wrong, what does your contract say about liability?Not because you plan to litigate, but because the answer shows whether anyone there has thought about the failure case.That has never happened.

Group 3: Accuracy and failure modes#

Demos run on threads that work. This group is you asking, politely, to see the ones that do not.

Ask thisWhy it mattersA weak answer sounds like
13. Show me it failing. Which kind of thread does it handle worst?The best single question in the bank. Nobody who has run this on a busy mailbox for a year lacks an answer.It handles everything well.
14. What are the published limits: thread length, attachment types, message size?Every system has limits. Published ones let you test the edge deliberately; unpublished ones surface on the thread you cared about.There is no limit.
15. When it is not confident, does it guess, flag or stay quiet? Show me.Confident wrong output costs more than a visible gap, because you stop checking. You want to see the uncertain state rendered.It is very accurate.
16. Where does the drafting voice come from, and can I point at the input behind a sentence?An accuracy question and a data question at once. A configured voice is steerable and inspectable; one inferred from your archive is neither.It learns your writing style from your past emails.
17. Which model is behind this, can I choose it, and how will you announce changes?A routing change moves tone and behaviour overnight with no release note. You want a changelog rather than a Monday morning.We route to the best model for each task.

Group 4: Provider coverage#

Coverage is the cheapest disqualifier in the process and the one buyers check last. Ask these with your least standard mailbox in mind.

A grid of checkmarks with three empty squares outlined in green, representing the surfaces and providers an AI email vendor has not shipped yet
Every grid has empty squares. The question is whether they name them before you find them.
Ask thisWhy it mattersA weak answer sounds like
18. Which mailbox types connect natively: Gmail, Microsoft 365, Exchange on-premises, IMAP, shared aliases, delegated mailboxes?Shared and delegated mailboxes break more integrations than any other case, and teams usually want them covered first.We support any email account.
19. Does anyone have to migrate mail or change their address?A migration is a project with a rollback plan, not a signup. It also changes who has to approve the purchase.We will migrate you for free. Still a migration.
20. Is this a client that replaces my interface, or an overlay on Gmail and Outlook?The sharpest architectural line in the category, and most comparison pages blur it. It decides whether you have a rollout or an install.We work inside your existing workflow. Which one is that?
21. Which OAuth scopes do you request, and what is each one for?Scopes are the actual permission grant, not a summary of it. Anyone on the call should justify each without reading a screen.Just the standard ones.
22. Which surfaces ship today and which are roadmap: desktop, mobile, Linux, offline?Coming soon without a version is a wish. Ask which release, then what the last three features did against their dates.It is on the roadmap for this year.

Group 5: Team and admin#

Skip this group if you are buying one seat. Weight it above everything else if a security team has to approve the purchase. A control that is per user and optional is not a control.

Ask thisWhy it mattersA weak answer sounds like
23. Can an admin set autonomy policy for everyone, or is every safety setting optional?Your rollout is only as safe as the least careful person's settings unless policy is enforced centrally.Each user configures their own, which people love.
24. What identity support exists: SSO or SAML, SCIM provisioning, deprovisioning at offboarding?Without deprovisioning, a leaver keeps a live mailbox connection until somebody remembers. That gap is what security reviews reject on.You invite people by email, it takes seconds.
25. Can an admin export the audit trail for another user, and what is in an entry?A log only the vendor can read is not evidence. Ask what an entry records: actor, target, reason, timestamp.Support can look into it for you.
26. What happens to a departing employee's drafts, rules and context configuration?Configuration written over months is company knowledge, and losing it with the seat is a cost nobody prices into a renewal.The seat is freed for the next person.
27. Do you have a SOC 2 Type 2 report or a completed CAIQ, shareable under NDA?Type 2 covers a period rather than a moment. The CSA's CAIQ is a self-assessment published to the STAR Registry, not an audit.We are SOC 2 aligned. Or: the audit is in progress.

Group 6: Commercial terms#

Four questions about the shape of the bill rather than its size. Prices move constantly here, so verify any number on the vendor's own pricing page at signature time.

Metering by how much mail arrives falls hardest on the busiest inboxes, which are the people buying triage. At least one vendor here does exactly that.

Ask thisWhy it mattersA weak answer sounds like
28. What exactly does the price meter: seats, messages received, AI actions, tokens?The model decides your bill more than the headline number does, and it is the part nobody models before signing.Simple per-seat pricing, followed by an asterisk.
29. What happens when I hit the AI allowance mid-month: stop, throttle or overage?The question that produces a surprise invoice, or a dead assistant in the third week of your busiest month.Most customers never reach it.
30. What are the renewal terms: auto-renew, notice window, cap on price increases?Renewal terms are negotiable before signature and never afterwards. Ask while you still hold the only leverage you get.We have never raised prices.
31. What is the support commitment in writing: hours, response target, who answers?A named target is something to hold someone to. Responsive is a feeling, and it changes as a company grows.We are very responsive, ask our customers.

Group 7: Continuity and exit#

This group stopped being theoretical. Notion Mail announced its shutdown in June 2026, the inbox retires on 22 September 2026, and unexported drafts, snippets and auto-label rules are deleted the day before.

Nobody who bought it asked question 33 and got a bad answer. They never asked it.

Ask thisWhy it mattersA weak answer sounds like
32. Who owns the company, who funds it, and is email the main product?Ownership moves here. Superhuman now sits inside Grammarly's group and the name covers both a mail client and a bundle, so establish which entity signs.We are backed by great investors.
33. If you shut down or are acquired, how much notice do I get, and is it in the contract?Notion Mail users got roughly a quarter's warning. A notice period costs a vendor nothing to promise, and refusing shows where the thinking stops.We are not going anywhere.
34. Run the export now, on this call. What is in the file, and what reads it?An export you have watched run is a fact. Messages-only output in a proprietary format is an archive, not an exit.Export is available on request.
35. If I cancel tomorrow, what stops working in my mailbox?The answer reveals where your mail actually lives. If cancelling costs you the mail itself, you bought hosting, not an assistant.Everything stays as it is. Including the rules?

The four shapes of a weak answer#

Nobody gives thirty-five clean answers, us included. What matters is telling the gap that is fine from the gap that ends it.

  • Not yet. An honest roadmap gap. Ask which release, then check the last three things they shipped against the dates they gave.
  • I do not know. Usually the wrong person in the room. Ask who does know and for a written answer by a date, then see whether it arrives.
  • The reframe. You asked about training and got an answer about selling data. Repeat the original question, unchanged, and let the silence work.
  • I will not say. The only fatal shape. A vendor who will not put retention, training, sub-processors or a notice period in writing has answered all four.

Two follow-ups do most of the work

Can you show me that now, in the product? And: can you put that in an email? Between them they turn a good call into two artifacts, and they are why a vendor who was going to be vague gets precise around minute twenty.

When several weak answers land at once#

Cluster them by group before reacting. One wobble anywhere is noise; two inside the same group is the group telling you something.

Weak on data handling plus weak on autonomy should end the evaluation. A system that acts on its own and will not describe what it does with what it read is a problem no later setting fixes.

Weak on commercial terms plus weak on continuity is usually an unfinished contract. Ask for the notice period, renewal cap and overage behaviour in the order form. Agreement repairs two rows in an afternoon; refusal answers question 33 properly.

Turn the list on us: where AI Emaily answers badly#

A question bank whose author passes it cleanly is a brochure. We build AI Emaily, so here are the eight of these thirty-five where our own answer is weak, checked July 2026.

QuestionOur answer, July 2026
6. DPA and BAAWe sign a DPA. We do not sign a BAA. If your mail carries protected health information we are not the right vendor, and nothing else here changes that.
22. SurfacesDownloadable desktop apps for macOS and Windows, both an Electron shell around the web interface rather than a native binary. macOS is Apple Silicon only, and not in the Mac App Store. Native iOS app, Android as a PWA, no Linux build. Offline covers reading and drafting, not a full local archive.
24. IdentityNo SSO or SAML, no SCIM, no automatic deprovisioning. Team seats with shared inboxes and comments, invites by email. For a security review that gates on identity, this is a no.
27. AttestationsNo SOC 2 report and no published CAIQ. SOC 2 is on the roadmap, which is exactly the answer question 27 tells you to discount.
29. AllowancePer seat, and we do not meter how much mail arrives. Plans carry a monthly AI credit allowance, so a heavy month can reach the ceiling. Bring your own model key and the allowance stops applying.
31. SupportEmail support from a small team that knows the product. No published response-time target and no phone line.
33. Notice periodWe publish none. We are independent and founder-led with email as the only product, and your mail stays in a mailbox you own because we never host it. Read the missing written commitment as the gap it is.
34. ExportSelf-serve from Settings, no ticket, and it includes the vendor-side layer: Context, rules, templates and drafts. But it is our JSON, no other client imports it, and rules and Context are a rebuild by hand.

The other twenty-seven we answer well. Autonomy is an explicit per-account setting, Manual, Copilot or Autopilot, with Copilot as the default. Sends stay approval-gated, every autonomous action lands in an audit log with its reasoning, and the destructive ones undo. We do not train on your mail and hold zero-retention arrangements with model providers.

On question 16, the drafting voice comes from a Personal Context you write plus client profiles you set, never inferred from your past messages, so you can point at the line that produced a sentence. On 18 and 19, Gmail, Outlook, Microsoft 365, iCloud, Fastmail, Proton and standard IMAP connect in one view with no migration.

Three questions belong to somebody else. If a security team gates on 6, 24 and 27, an assistant inside Google Workspace or Microsoft 365 beats us outright: your IT already runs the identity and audit controls, and both now bundle AI. On 34, any client that hands you MBOX or EML beats our JSON.

The follow-up email is the deliverable#

The call produces notes nobody trusts in six months. The email produces a record, and asking a vendor to confirm their own words is itself the last question in the bank.

Sent within 24 hours of the call
SubjectWritten confirmation of four answers from today
OpeningRecording what we heard. Correct anything wrong in reply.
Q2No part of our mail trains, fine-tunes or evaluates any model, including a per-customer one.
Q3Zero-retention terms exist with every model provider, and appear in the DPA.
Q6You sign our DPA. You cannot sign a BAA.
Q33Notice on discontinuation is 90 days, added to the order form.
ClosingWe will treat no reply by Friday as agreement.

Either it comes back corrected, which is a good vendor doing its job and leaves you with a document. Or it never comes back, which answers question 31 for you.

Frequently asked

Nafiul Hasan

Written by

Nafiul Hasan

Nafiul Hasan is an entrepreneur and AI automation system builder with 10+ years of experience turning messy, manual workflows into reliable automated systems. He designs and ships AI enterprise solutions end-to-end — the agent logic, the data plumbing, and the product people actually use — and founded AI Emaily to give busy professionals their attention back. He writes here from the builder's seat: what works, what breaks, and how to put AI to work without giving up control.

EntrepreneurAI Automation System BuilderAI EnthusiastBuilds AI Enterprise Solutions10+ years experience
More from Nafiul
Ready when you are

Ask us all thirty-five.

AI Emaily runs as Manual, Copilot or Autopilot per account, keeps sends approval-gated by default, logs every action with undo, and connects to the Gmail, Outlook or IMAP mailbox you already own. The eight we answer badly are named above. Start free at app.aiemaily.com/signup.

  • 7-day free trial
  • Cancel anytime
  • Every provider