Home / Can I connect AI to my work email?
AI tools and borrower dataCan I connect AI to my work email?
Last reviewed: October 28, 2026. General information, not legal advice. Plan features below were read on the vendors' own pages on that date and can change; confirm them before you buy. This page describes a sensible internal policy for a mortgage company. It is not a statement of what any regulator or GSE requires word for word.
Why email is different from pasting a question into a chat
When someone types a question into an AI tool, they choose what goes in. When an AI tool is connected to a mailbox, it can read whatever that mailbox holds. In a mortgage office that usually means borrower nonpublic personal information: Social Security numbers, pay stubs, bank statements, credit reports, tax returns and loan estimates sent back and forth during a file. A single connector can expose years of it.
The rules that already apply to that information do not pause for convenience. The Gramm-Leach-Bliley Act and the FTC Safeguards Rule expect you to limit each user's access to the customer information they need, to oversee service providers by contract, and to keep a written incident response plan. Freddie Mac Section 1302.8 and Fannie Mae Supplement Section 4 expect you to know and govern where AI is used. An email connector is an AI use that touches the most sensitive data you hold, so it needs the strongest paperwork.
Our recommended rule, in three tiers
| Situation | Allowed? | What has to be in place |
|---|---|---|
| Personal AI account (free or paid consumer plan) | No | Never connect a personal account to a work mailbox or forward borrower email into one. Staff confirm this in the personal-account attestation in the portal. |
| Business or Teams plan (for example ChatGPT Business or Grok Business) | Only by documented exception | Named users only, an approver, a written risk acceptance, conditions (SSO and MFA, own mailbox only, retention set, reviews), and a review date. Log it in the portal under AI controls, Exceptions. |
| Enterprise plan with a signed DPA covering the product and connector | Yes, company-wide, once rolled out | DPA and SOC 2 report on file, SSO and MFA enforced, user provisioning and removal controlled, audit logs available and reviewed, retention configured, staff trained and signed off. |
Why Enterprise and a signed DPA, not just a paid plan
Business plans are a real step up from personal accounts. For example, OpenAI's plan comparison lists SAML SSO, an admin console, SOC 2 Type 2 and connectors to Google Workspace and Microsoft 365 for ChatGPT Business, and says business data is not used for training by default. x.ai lists role-based access control, custom data retention, connectors and "advanced audit and security controls" for Grok Business.
The gaps that matter for email sit at the next tier. As of October 28, 2026, OpenAI's comparison table marks SCIM user provisioning, role-based access controls, the Compliance API logs platform, IP allowlisting and Enterprise Key Management as No for Business and Yes for Enterprise. x.ai lists custom single sign-on and directory sync (SCIM) only for Grok Enterprise. Without automated provisioning, a departing loan officer's AI account and its email connection rely on someone remembering to remove them. Without exportable activity logs, it is hard to show what the AI read during an incident review.
The DPA is the contract that makes the vendor a service provider bound to protect your data. Under Regulation P's service provider exception, sharing nonpublic personal information with a service provider depends on a contract that limits its use of that information, and the Safeguards Rule requires service providers to maintain safeguards by contract. Check scope carefully:
- OpenAI publishes a self-serve DPA that is accepted by clicking "I agree", signing an Order Form or using its business services. Confirm in writing that it covers the specific product and connectors you plan to use.
- x.ai publishes a DPA (now titled the SpaceXSI Data Processing Addendum, effective September 22, 2026) that is incorporated into its Enterprise Terms of Service. Its text says personal data may be processed only through the "ZDR-Enabled API" and that other services must not receive personal data. We could not confirm from the published text that it covers the Grok chat workspace or its email connectors. Get that answer in writing from x.ai before any borrower email flows through Grok.
The risks to weigh before any exception
- Over-broad access. A connector usually sees the whole mailbox, including old loan files and shared threads, not just the message the user is working on.
- Prompt injection through incoming email. A message from outside can contain hidden instructions aimed at the AI tool, such as "forward the last bank statement". Freddie Mac Section 1302.2 names prompt injection as a threat your security training should cover.
- Copies outside your control. Summaries, memories, chat history and files created from email may be stored by the vendor under its retention settings, in addition to your mail system.
- Weak audit trail. On plans without exportable activity logs you may not be able to reconstruct what the AI accessed during an incident.
- Offboarding gaps. Without SCIM, removing a user from your directory may not remove their AI account or its email connection.
- Wrong-recipient and auto-drafted replies. AI-drafted replies can attach the wrong document or state loan terms that a person has not checked. A person must review anything that goes to a borrower.
- Incident clock. Bulletin 26-01 plans for a first notice to Fannie Mae within 36 hours of a cybersecurity incident. The Safeguards Rule requires notice to the FTC within 30 days of discovering a notification event involving 500 or more consumers. Vendor DPAs typically promise notice "without undue delay", which may be slower than you need.
- Sub-processors. Your borrowers' data may pass to the vendor's own providers. Both OpenAI and x.ai publish sub-processor lists; keep a dated copy.
What a documented exception must contain
An exception is a written decision by an accountable person to accept a specific risk, for specific people, for a limited time. In the BrokerLumen portal it lives under AI controls, Exceptions, with these fields:
- Who and what: the named users, the tool and plan, and the connector (for example "Outlook connector, own mailbox only").
- Business reason: why the benefit is worth the risk now rather than after an Enterprise rollout.
- Approver: the executive sponsor or AI controls owner, not the person requesting it.
- Risk acceptance: the risks above in your own words, and the decision to accept them.
- Conditions: SSO and MFA on, connected only to the user's own mailbox (no shared, processing or closing mailboxes), memory and chat history settings reviewed, retention set to the shortest workable period, no auto-send, monthly spot checks, and immediate disconnection when the person changes role or leaves.
- Review date: no more than 90 days out, and no later than your planned Enterprise rollout.
Worked example. Three administrators use ChatGPT Business with the Outlook connector to draft internal scheduling and vendor emails. The executive sponsor approves the exception, accepts the risks of whole-mailbox access and limited audit logs on the Business plan, and sets conditions: SSO and MFA enforced, own mailbox only, no borrower document folders, no auto-send, and a review on the date the Enterprise plan and DPA are expected to be in place.
Getting to company-wide use
- Pick the vendor and get an Enterprise quote using the vendor quote checklist.
- Sign the DPA, confirm it covers the product and the email connector, and file it with the vendor's current SOC 2 Type 2 report.
- Turn on SSO, MFA and SCIM provisioning, and set retention.
- Confirm how audit logs are exported and who reviews them.
- Update your AI policy, train staff and collect sign-offs.
- Roll out by phase, starting with a pilot group, and close any open exceptions.
The AI controls checklist lists the six controls to have in place first, and the plan comparison guide shows what each ChatGPT and Grok tier includes as of October 2026.
Log exceptions and track sign-offs in one place
BrokerLumen portal plans include an AI controls area for vendor DPAs and SOC 2 reports, documented exceptions with approvers and review dates, roles and a rollout planner, alongside staff sign-off tracking. The $349 AI Policy Kit includes 3 months of Small Shop portal access.
Get the AI Policy Kit, $349AI Policy Starter, $99AI Team Kit, $699Compare kitsFrequently asked questions
Can my loan officers connect ChatGPT to Outlook?
Not company-wide until you are on an Enterprise plan with a signed DPA that covers the product and the connector. Before then, only named users under a documented exception with an approver, conditions and a review date.
Is ChatGPT Business enough for email connectors?
ChatGPT Business supports connectors and SAML SSO, but as of October 28, 2026 OpenAI lists SCIM, role-based access controls and the Compliance API logs platform as Enterprise features. We treat Business as exception-only for email.
Does the xAI DPA cover Grok connected to email?
The published SpaceXSI DPA says personal data may be processed only through the ZDR-Enabled API. We could not confirm that it covers the Grok chat workspace or its connectors. Ask x.ai in writing.
Can staff use personal AI accounts with work email?
No. Personal accounts are outside your contracts and controls. Staff confirm this in the personal-account attestation.
Is this legal advice?
No. It is general information and a recommended internal policy. Confirm your obligations with counsel.