A private email assistant for architects is one whose written terms you could show a client. It doesn’t train general AI models on your mail. It names every company that processes your mail. It says how long it keeps the summaries and search index it builds, and it deletes them when you leave. It limits which of its own staff can look, and says when. Those terms then have to fit what you’ve already signed: the confidentiality clause in your owner-architect agreement, any NDA a developer handed you, and on federal work, the rules for controlled building drawings.
- Who can already see your client work: your team, every consultant on the cc line, your mail provider and whoever runs your admin account.
- Who an assistant adds: the vendor’s staff, the AI companies it sends excerpts to, and the copies it makes.
- What to ask for: five answers in writing. They’re in a table below.
- Where to keep AI out: controlled building information on federal work, unless the contracting officer agrees.
Client work in a practice inbox
Most of what a client tells you is ordinary. Some of it would hurt them if it got out. A small practice’s inbox usually holds a bit of each.
- The plan of a private house, with the alarm panel, the safe room and the camera positions marked.
- A developer’s site purchase, months before it’s public.
- A tenant’s fit-out for a store that hasn’t been announced.
- An owner’s budget, and the bid numbers that came in against it.
- A school’s security upgrade, or a courthouse’s.
- A board member’s email about why the project is really over budget.
None of that arrived as “confidential” by accident. Clients send it because you need it to design. The question is who else ends up holding it.
Who can see it before any AI is involved
The list is longer than most practices picture. It’s also mostly people you chose.
| Who | What they can see | What governs it |
|---|---|---|
| Your staff | Any mailbox they share or are delegated to | Your employment terms and your own access settings |
| Consultants and contractors | Everything on threads where they’re copied | Their agreements with you, if they carry a confidentiality clause |
| Your mail provider | The mailbox itself, Google or Microsoft | Your Workspace or Microsoft 365 terms |
| Your workspace admin | Every mailbox in the account | Whoever holds that login. In a seven-person firm, often you |
| Anyone a thread is forwarded to | That thread and everything quoted under it | Nothing, once it’s gone |
The last row is where most leaks start. Someone forwards a long thread to the structural engineer, and a client’s budget email is quoted twenty messages down. Someone sends a drawing set to their personal Gmail to look at over the weekend. Neither one involves software at all.
Who an email assistant adds
An assistant reads your mail so it can sort, summarize and draft. That’s the whole job, so there’s no version where it doesn’t see the content. What changes is the list of who else holds a piece of it.
- The assistant’s company. It stores your mail, or a copy of it, to do the work.
- The AI providers it uses. Excerpts of a thread go to a language model, usually run by another company.
- What it builds. Summaries, a search index, extracted dates, drafts. These are new copies of client information, and they can outlive the email.
- Its staff. Support and security people may be able to look, under rules the vendor writes down, or doesn’t.
Training on your mail gets the attention. How long those copies live decides more. Keeping client data out of AI model training reads a training clause line by line. Where your mail goes when AI reads it follows one message through each step.
What your agreement already says about client confidences
If you work on the AIA’s owner-architect form, your confidentiality duty is already written down. Here’s B101-2017 at Section 10.8, as it reads in a public copy of the standard form:
“If the Architect or Owner receives information specifically designated as “confidential” or “business proprietary,” the receiving party shall keep such information strictly confidential and shall not disclose it to any other person except as set forth in Section 10.8.1.” AIA Document B101-2017, Section 10.8
The duty covers what the owner designates, so look at what’s been marked. Section 10.8.1 then lets you share it with “employees, consultants, or contractors in order to perform services or work solely and exclusively for the Project,” as long as they’re bound by the same restrictions. Read that against any software that will process the owner’s marked information. Whether a software vendor fits that list is a question for your agreement, and for your lawyer if the project is sensitive.
Owners often rewrite this clause, so read your signed copy. The University of Wisconsin System’s form replaces it with its own list of permitted recipients and adds the state’s public records law.
Drawings and models have their own clause. B101 Section 1.3 says the parties “shall agree upon protocols governing the transmission and use of Instruments of Service or any other information or documentation in digital form” (AIA sample). It points to AIA E203-2013, which adds a category called Confidential Digital Data. That’s data the sender “designates and clearly marks as “confidential."" The rule for it is the same shape:
“If a Party receives Confidential Digital Data, the receiving Party shall keep the Confidential Digital Data strictly confidential and shall not disclose it to any other person or entity except as set forth in Section 2.2.1.” AIA Document E203-2013, Section 2.2 (AIA sample)
If you’re an AIA member, Rule 3.401 of the Code of Ethics and Professional Conduct also applies. It says members “shall not knowingly disclose information that would adversely affect their client or that they have been asked to maintain in confidence.” An email assistant for architects, when there’s no PA covers what that rule means when someone else reads your mail.
What to look for in a developer’s NDA
A separate NDA usually reaches further than B101. When one arrives, look for four things before you connect any assistant to the mailbox that will carry the project.
- How confidential information is defined. Many NDAs cover everything the client discloses, marked or not.
- Who you may share it with. Look for “representatives,” “advisors” or “service providers,” and whether they must sign similar terms.
- Anything about AI or third-party tools. Newer NDAs sometimes name generative AI. Some ban it, and some ask you to list the tools you use.
- Return and destruction. If the NDA says to destroy everything at the end, that reaches the copies an assistant made too.
Federal work and controlled building drawings
On GSA-controlled space, some drawings come under a federal rule as well as your contract. GSA Order PBS 3490.3 treats sensitive building information as Controlled Unclassified Information (CUI), in the Physical Security category. It names “architects, engineers, consultants, contractors” among the firms that may receive it. Its contract clause sets the rule for sending it:
“Electronic transmission of CUI information outside of the GSA network must use session encryption (or alternatively, file encryption) consistent with National Institute of Standards and Technology (NIST) SP 800-171.” GSA Order PBS 3490.3 CHGE 1, contract clause section 2.a (full order)
The same clause asks three more things of a contractor:
- A list of everyone you sent it to. The name of each firm, the person responsible there, their contact details and what you sent. “For Federal buildings, final payment may be withheld until the lists are received.”
- A controlled place to keep it. Electronic and paper copies stay where unauthorized people can’t reach them.
- Return or destruction at the end. You return it to the contracting officer or destroy it under NIST SP 800-88. Then you notify the contracting officer to receive final payment.
An AI assistant that reads the mailbox becomes another place the drawings sit, and another set of companies they pass through. That’s hard to square with a dissemination list and a destruction certificate. So this is where something else fits better. Keep marked CUI out of any mailbox an outside AI service reads, unless the contracting officer has agreed in writing. Use the agency’s own transfer method for the drawings. Let the assistant handle the rest of the project’s mail. Sending large drawing sets by email covers links, transmittals and size limits for everything else.
What to demand from a private email assistant for architects
Ask these in writing. A good vendor answers each with a section number in a published document.
| Ask | The answer worth accepting | Why a practice needs it |
|---|---|---|
| Do you train AI models on my mail? | No training of general or foundation models without your permission, in the contract | A client’s floor plan inside someone else’s model can’t be pulled back out |
| Which companies process my mail? | A named, dated list of subprocessors | It’s the list you’d give an owner who asks who has seen their drawings |
| How long do you keep what you build from it? | A retention table with its own row for summaries and indexes | An NDA’s destruction clause reaches the copies, not just the email |
| What happens when I leave? | Deletion or return within a stated period, with backups explained | Projects end, and so do vendors |
| Which of your staff can read my mail, and when? | A written list of the circumstances | “Nobody, ever” isn’t true of any product that offers support |
Three more answers matter in a practice.
- Which mailbox it reads. An assistant that works on your existing Gmail or Microsoft 365 account leaves the archive where it is. One that asks you to move mail makes a second copy of every project.
- What it does on its own. Look for a setting per kind of action, with review as the default. A reply to a client about a design change should wait for you.
- What it does with mail from strangers. An email can be written to give instructions to an AI. A good assistant holds unknown senders aside before it reads them for meaning.
Is an AI email client safe for client data? has the longer list of questions for any vendor.
Copilot and Gemini, already in your mailbox
Your mail provider may already have an assistant switched on. Its terms are the ones you already have with Microsoft or Google.
Microsoft 365 Copilot. Microsoft’s privacy documentation says “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs, including those used by Microsoft Copilot.” The same page says Copilot “only surfaces organizational data to which individual users have at least view permissions.” So it can reach anything a user can already open. A project folder shared with the whole firm years ago is now easy to find. Check old sharing before you turn it on.
Gemini in Gmail. Google’s Workspace privacy hub says “Your content is not human reviewed or otherwise used for Generative AI model training outside your domain without permission.” That’s for Workspace accounts.
Both commitments cover business accounts. A personal Gmail or Outlook.com account is under different terms. So the common slip in a practice is someone pasting a client’s email into a chatbot on their phone. Your firm’s AI policy matters more than any vendor’s. Point compared with Outlook and Point compared with Gmail set the built-in tools beside Point.
Retention: how long any of it lives
Two clocks run on the same project mail.
Yours. You keep project correspondence for your own record. Your insurer and your state’s rules shape how long. That’s your archive, and it stays in your mailbox.
The assistant’s. The vendor keeps your mail and what it built from it. Ask whether deleting an email also deletes its summary and its place in the search index. Ask what happens to backups. If an NDA or a federal contract says to destroy everything at the end, those copies are part of everything.
The second clock is the one to get in writing. A vendor that never trains, but keeps every summary for years, holds more of a client’s business than you might want.
Four projects, sorted before you connect
Here’s how one practice of seven might sort its live work.
| Project | Where confidentiality comes from | What the assistant reads |
|---|---|---|
| A private house | B101 §10.8; the owner marked the security plan confidential | Everything, with the security drawings sent as a locked message or by link |
| A developer’s mixed-use scheme | An NDA covering all disclosures, with no AI clause | Everything, once the NDA’s “service providers” wording is checked |
| A federal office fit-out | GSA’s CUI clause | Ordinary project mail. CUI drawings go through the agency’s transfer route, not this mailbox |
| A school addition | B101 as modified by the district, plus public records law | Everything. Much of it may be public anyway |
That’s one afternoon’s work, and it answers the question an owner might ask later. Write it down where your team can see it.
How Point handles a practice’s client work
Point is an AI email client. Every question above lands on Point too. Here’s where each answer lives.
- Training. Section 6.5 of the privacy policy says Point doesn’t train generalized third-party or foundation models on your mail unless you expressly authorize it. Point may use it to improve the service for you.
- Who processes it. The subprocessor list is versioned and dated. Amazon Web Services hosts the service. Anthropic and OpenAI do the AI processing, under business terms with no-training controls where those exist. Google’s Generative AI API builds the search embeddings.
- How long. The retention table gives summaries, indexes and other derived data their own row. They’re generally kept while the mail they came from is kept.
- When you leave. The data processing addendum in the terms commits to deleting or returning your data within 60 days, if that’s specified in your agreement. Backups may last until they’re overwritten.
- Who at Point can look. Section 6.8 limits it to support you asked for, security, legal compliance, troubleshooting, running the service, or with your consent.
- What Point isn’t for. Section 5.1 of the terms rules out classified information, and data under contractual controls Point hasn’t accepted. Controlled building information on a federal job sits there.
Point reads the Gmail or Microsoft 365 account you already have, so there’s nothing to migrate. For the one drawing set that needs it, there’s a lock in the composer. “Lock a message end to end and it stays between you and the recipient. Not even Point can read it.” A locked message isn’t summarized or ranked, because Point can’t read it. Encrypted email, and what an AI assistant can’t read explains what each kind of lock covers.
How much Point does on its own is your setting, one kind of action at a time. Review is the default, so a client reply waits for you. At the top of the dial, Point does that kind of work without asking. Every action goes in the activity log. Mail from senders Point has never seen is held aside before anything reads it. When an email is written for the AI covers why.
Point keeps separate businesses apart. Inside one practice, projects share a feed, so the sorting table above still does its job. Keeping one business isolated from another shows where that line runs.
Point holds no SOC 2 report or ISO certificate yet. What it has is dated documents you can compare against the copy you kept. Point is in private beta. Email management for architects shows the rest of the practice inbox with Point reading it, and the benefits page has the full list.
Common questions
What is a private email assistant for architects?
It’s an AI or human assistant that reads a practice’s email under terms you could show a client. In writing, it doesn’t train general AI models on your mail. It names who processes your mail and says how long it keeps what it builds. It deletes on exit and limits staff access. It should also fit the confidentiality clause in your owner-architect agreement and any NDA.
Can I use an AI email assistant if my client made me sign an NDA?
Often, yes. Read the NDA first. Check how it defines confidential information, who you may share it with, and whether it mentions AI tools. If it bans third-party processing, keep that project’s mail out of the assistant or ask the client in writing.
Does an encrypted email client keep client work private from an AI assistant?
End-to-end encryption keeps a message from anyone without the key, an assistant included. That protects the message, and it also means the assistant can’t summarize or sort it. Most practices encrypt the few messages that need it and let the assistant read the rest.
Is a secure email client enough for federal drawings?
Not on its own. GSA’s rules for controlled building information require approved encryption for electronic transmission. They also require a list of every recipient and return or destruction at the end. Use the agency’s transfer method, and keep those drawings out of any mailbox an outside AI service reads unless the contracting officer agrees.
Do Copilot and Gemini train on my practice’s email?
Microsoft says Copilot doesn’t use prompts, responses or Microsoft Graph data to train foundation models. Google says Workspace content isn’t used for generative AI training outside your domain without permission. Both apply to business accounts, not personal ones.
Who at an AI email company can read my mail?
That depends on the vendor’s written policy. Look for a list of the circumstances, such as support you requested, security investigations and legal compliance. A vendor that says nobody can ever look either has no support process or hasn’t thought it through.