Research brief · AI and compliance for professional offices · November 2026

Email, AI and the GDPR

Can I connect my work email to ChatGPT? Every professional office that keeps client data in email asks it: an accountant looking for the latest tax circular, a lawyer for last year's clause, an insurance broker for an insurer's green card instruction. The same inbox holds tax returns, contracts, CVs, ID copies and, in a clinic or an insurance office, a medical report sent this morning. We read OpenAI's current product and contract pages, the GDPR, the EDPB guidelines and DORA, and set out what a connection really opens, what stays stored, where it runs, and when a private company brain is the safer design.

By Christos Papadimitriou · House of 47 Research
Last updated:

Read the answer
1 scopegmail.modify, what ChatGPT's Gmail app requests: read, compose and send
30 daysfor deleted ChatGPT Business and Enterprise chats to leave OpenAI's systems, unless the law says longer
Art. 9health data are prohibited unless an exception applies, for care under medical secrecy or by explicit consent
7 officesclinics, law firms, accountants, HR, estate agents, advisers, brokers: same inbox problem
The short answer · key findings

Test inbox yes. Real inbox, not yet

For a test mailbox that holds only public circulars and invented messages, yes. For a real mailbox with client files, the answer is "not yet", until a short list of technical and legal checks is done. Six findings explain why.

Same problem, different offices · what makes the inbox sensitive

Seven offices, one inbox problem

The question is the same in every profession that keeps client files in email. What changes is which data make the mailbox sensitive and which extra rules apply. The GDPR covers all of them; DORA only the financial ones.

OfficeWhat sits in the inboxWhy it is sensitive
Clinics and dental practicesDiagnoses, test results, X-rays, referralsHealth data, a special category (Art. 9), under medical secrecy (Art. 9(2)(h), 9(3))
Law firmsCase files, contracts, evidence, sometimes criminal recordsProfessional secrecy under national rules (Art. 90); criminal-offence data only under Art. 10
Accountants and tax advisersTax returns, payslips, bank statements, tax IDsConfidential financial data, attractive for fraud; payroll for clients often makes you a processor (Art. 28)
HR and recruitmentCVs, interview notes, references, sick notesCandidate and employee data; sick notes are health data (Art. 9); national employment rules (Art. 88)
Real-estate agenciesID copies, proof of income, buyer and seller detailsIdentity documents; estate agents are obliged entities under EU anti-money-laundering rules (Directive (EU) 2015/849)
Financial advisersPortfolios, income, risk profiles, IDConfidential financial data; investment firms are in DORA's scope
Insurance brokersPolicies, claims, number plates, medical reportsHealth data in life and health cover (Art. 9), claims files; DORA except SME intermediaries
Permissions · what the app actually sees

The consent screen decides, not the chat

When you connect Gmail to ChatGPT, OpenAI requests one Google scope: gmail.modify. Google describes it as "read, compose, and send emails from your Gmail account", everything except immediate permanent deletion. Outlook Email works through Microsoft Graph permissions that a Microsoft Entra administrator grants for the whole organisation.

OpenAI is clear about what the in-chat settings do. ChatGPT can reach the content "available to the account you connect". The app permission options, from "Always ask" to "Allow read actions", decide when ChatGPT asks before using that access. In OpenAI's words, "app permissions do not grant an app new access", and they do not take any away either. To reduce access, you disconnect the app or an administrator disables it.

Four details matter for any office:

  • Defaults. In ChatGPT Business, apps are enabled by default; the workspace owner has to switch off the ones the office does not allow. Enterprise starts with a selected set and adds role-based controls, and can exclude Outlook messages by sensitivity label.
  • Memory. If Memory is on, eligible information from connected Google apps may be used to personalise ChatGPT. Disconnecting Gmail does not delete existing chats or saved memories.
  • Shadow AI. Often the riskiest connection is the one nobody approved: 78% of AI users bring their own AI tools to work (Microsoft and LinkedIn, 2024), and 20% of breached organisations reported incidents involving shadow AI (IBM, 2025). A personal account linked to the office mailbox comes with no DPA and no Article 28 authorisation, while the office remains the controller.
  • Hostile email. An incoming message can contain instructions aimed at the assistant. OpenAI says its safeguards against prompt injection "do not eliminate" the risk.

One point in the common advice is now out of date. ChatGPT no longer offers administrator-managed or individually authorised sync for apps; they use live access. The copies that remain are therefore the ones the work itself creates: chat history, uploaded or generated files, project contents and memories. Other AI tools may still sync and index a mailbox, so ask every vendor.

If the goal is fast search of circulars and price lists, do not connect the main mailbox. Create a separate mailbox that receives only official circulars, newsletters and supplier price lists, for example through a forwarding rule by sender domain, and connect that.

Retention · what stays after the answer

No training is half the answer

OpenAI's enterprise privacy commitments state that data from ChatGPT Business, Enterprise, Edu and the API Platform are not used to train its models by default. That settles training. It does not settle storage.

When the AI reads a claim email and writes a summary that mentions the compensation amount, the amount and the details now live in the chat history, and possibly in a project's files or in a memory. Workspace admins control how long data is kept. Deleted conversations are removed from OpenAI's systems within 30 days, unless the law requires longer; for Business, also when "reasonably necessary to protect our services or any third party from harm".

The API is not automatically different. By default OpenAI keeps abuse monitoring logs for up to 30 days. Zero Data Retention and Modified Abuse Monitoring exist, but only after OpenAI's prior approval, and some endpoints (files, vector stores, assistants, stored conversations) keep application state until you delete it, even with Zero Data Retention on.

The question for every office, especially when an employee or partner leaves: what copies remain inside the tool's chats and files, who can see them tomorrow, and when are they finally deleted?

Europe or America · storage versus processing

Where it is stored, where it runs

An EU-hosted mailbox protects the mailbox. Once an AI reads it, two separate questions apply: where the content is stored (chats, files, history) and where the model runs (on whose servers the extract travels at the moment the answer is generated).

OptionStored at restModel runsContract and transfers
Personal ChatGPT (Free, Plus, Pro)Outside the EEA, including the US, per the EU privacy policyOutside the EEAOpenAI Ireland is the controller, not your processor; content may train models unless you opt out. No DPA. Not for client data.
ChatGPT BusinessNo EU data residency option listed for BusinessNo inference residency option for BusinessDPA with OpenAI Ireland; onward transfers under SCCs or an adequacy decision
ChatGPT Enterprise / EduEurope (EEA + Switzerland) for new workspacesEU inference residency for eligible workspaces; non-GPU steps may run globallyDPA; connected third-party apps follow their own terms
OpenAI API, EU projectEEA + Switzerland at eu.api.openai.comIn-region for supported endpoints and modelsNeeds approval for MAM or ZDR and a Modified Retention amendment; 10% uplift on models released from 5 March 2026

Per OpenAI help and platform pages read on 12 November 2026. Data residency does not cover system data (account, billing, usage) and does not cover third-party apps connected to ChatGPT.

Transfers to the US need a Chapter V mechanism. The EU-US Data Privacy Framework covers transfers only to US organisations certified under it; the EU General Court upheld it on 3 September 2025 (T-553/23, Latombe), and an appeal is pending before the Court of Justice (C-703/25 P). OpenAI's DPA names Standard Contractual Clauses or an adequacy decision for onward transfers by OpenAI Ireland. Check the DPF list for any provider rather than assuming, and, under SCCs, assess the destination as the EDPB recommends.

Who decides · roles, Article 28, DORA

Who approves, and why ISO is not enough

Large company, group or insurerin-house staff

No connection on your own initiative. Use only the tools IT and the Data Protection Officer have approved for your role.

Office acting for othersmanagement sets the rules

Management decides, but must also check its contracts. Where the office processes data on someone else's behalf, an accounting firm running payroll for client companies or an insurance agency working for an insurer, it is a processor, and Article 28(2) forbids engaging another processor, such as OpenAI, without that controller's prior specific or general written authorisation. The role is decided by the facts, not the label (EDPB 07/2020); for its own clients an office is often a controller.

Sole practitioner or small officeyou decide how you work

You are the controller and carry the full GDPR responsibility, whether you are a lawyer, an accountant, a dentist or an independent broker.

Only insurance and finance: DORARegulation (EU) 2022/2554

Applies from 17 January 2025 to financial entities, including insurers, insurance intermediaries and investment firms, which must manage the risk of every ICT provider, an AI vendor included. It does not apply to insurance intermediaries that are micro, small or medium-sized enterprises (Art. 2(3)(e)); small, non-interconnected investment firms follow a simplified framework (Art. 16). Accountants, law firms, clinics and agencies are outside DORA as such.

The certificate trap

ISO/IEC 27001 certifies an information security management system. SOC 2 is an auditor's attestation report on controls, not a certificate. OpenAI reports SOC 2 Type 2 for Business, Enterprise and the API. Both are useful evidence of security. Neither covers your own GDPR obligations.

What the law asks

Article 28(3) requires a processing contract, the Data Processing Addendum or DPA. It is necessary, not sufficient: you still need a legal basis (Art. 6, plus Art. 9 for health), a valid transfer mechanism (Chapter V), an updated privacy notice and, for health data at scale, an impact assessment.

Decision tables · two choices

Which inbox, which design

QuestionTest inboxReal office inbox
ContentsPublic circulars, price lists, invented clientsTax returns, contracts, CVs, ID copies, claims, medical reports
Legal basisNone needed for public or invented dataArt. 6 for each purpose, Art. 9 for health
ContractBusiness plan terms are enoughDPA, plus Art. 28 authorisation if you process for another controller
Where it runsNot criticalChecked: storage region, inference region, transfer mechanism
VerdictConnect and testNot yet: after the checks, or through a design that never exposes the inbox
QuestionDirect connection to ChatGPTOwn AI brain via API
What the model can reachEverything the connected account can seeOnly the passages your system selects
Who filtersThe model, after it has accessYour system, before anything leaves
Personal data sentWhatever the search touches: names, tax IDs, plates, medical filesMinimised: names, IDs and case numbers removed or masked first
Copies left behindChats, files, memories at the provider until deleted, plus up to 30 daysYour own logs in your own database; provider retention per contract, ZDR only where approved
EU processingEnterprise or Edu with residency; not BusinessChoose an EU-region endpoint where the provider offers and approves it
Staff leavingTheir chats and memories must be found and deletedKnowledge stays in company files; access is revoked in one place
EffortMinutesA project: weeks, with tests before launch
An insurance example, the green card · one request, four steps

What the model sees, and what it does not

1 · Findon the office's own system

The brain finds the insurer's circular in the office's own store. The client's name, plate and history stay there.

2 · Stripbefore anything leaves

Names, plates and internal case numbers are removed. A case number still counts as personal data if it can identify the client (pseudonymised data remain personal data, GDPR Recital 26), so it goes too.

3 · Askthrough a business API

Only the two lines of the circular and the question "Where is it issued?" go to the language model, through a business API that does not train on the data, in an EU region and with zero retention where the provider has approved it.

4 · Drafta person approves

The model returns a draft reply. The office adds the client's details locally and a person approves before it is sent.

Minimisation lowers what reaches the model to the minimum, ideally none. Removing names alone does not make data anonymous; test what is left.

First safe test · three steps, low risk

Try it this week, on nothing real

That is the useful role for AI here: a fast librarian for circulars and guidelines, kept away from client secrets.

The design that removes the problem · company brain

A brain that filters before it asks

Why the design matters

Every check above exists because a direct connection hands the whole inbox to the tool and then relies on settings, contracts and people to limit the damage. A private company brain reverses the order: the office's knowledge stays in its own store, permissions decide what each assistant may search, and the model only ever sees the passage it needs.

Our own bet

The AI Company Brain from theagency47 is built that way. Knowledge lives in files the office owns, with a search index on top and live systems queried, not copied. Each client gets its own database in the EU, a data processing agreement is signed before any client data is touched, sensitive categories can be excluded by sender, domain, folder or keyword, names, phones and IDs are masked before text reaches a model where the workflow allows, and a person approves what agents send.

Stated plainly, as on theagency47's trust page: data are stored in the EU, while model inference may run outside the EU under Standard Contractual Clauses unless an EU-region endpoint is scoped and approved for the client. Business APIs are used, whose terms exclude training. The first stage is a fixed-price build; see the agents for accounting, law firms, clinics and insurance brokers.

Do not ask the AI to ignore your clients' files. Build it so it never sees them.

Sources · read on 12 November 2026, primary documents first

Sources

OpenAI, Enterprise privacy at OpenAI (updated 8 Jan 2026): training, retention, DPA, SOC 2

OpenAI, Data Processing Addendum (effective 1 Jan 2026): roles, sub-processors, SCCs, Schedule 1

OpenAI, Europe privacy policy (updated 24 Aug 2026): consumer transfers and controller

OpenAI Help, Google app data controls FAQ: gmail.modify, Memory, deletion

OpenAI Help, Connected apps in ChatGPT: app permissions

OpenAI Help, Admin controls, security and compliance for plugins and apps: defaults, sync, label exclusions

OpenAI Help, Data residency and inference residency for ChatGPT

OpenAI Platform, Data controls in the OpenAI platform: abuse logs, ZDR, EU region

Google, Gmail API scopes

GDPR, Regulation (EU) 2016/679: Arts. 4(11), 6, 7, 9, 10, 28, 35, 88, 90, Chapter V, Recital 26

EDPB, Guidelines 07/2020 on controller and processor

EDPB, Guidelines 05/2020 on consent

EDPB, Recommendations 01/2020 on supplementary measures for transfers

European Commission, EU-US data transfers and the Data Privacy Framework · DPF participant list

IBM, Cost of a Data Breach Report 2025: shadow AI in 20% of breaches

Microsoft and LinkedIn, 2024 Work Trend Index: 78% bring their own AI

EU General Court, T-553/23 Latombe v Commission, 3 Sep 2025 · appeal C-703/25 P

DORA, Regulation (EU) 2022/2554, Article 2 scope and exemptions · EIOPA on DORA

ISO, ISO/IEC 27001:2022 · AICPA, SOC 2

DORA, Article 16 simplified framework · AML, Directive (EU) 2015/849

Εθνική Ασφαλιστική, MyEthniki app: green card issuance for active motor policies

theagency47, AI Company Brain · Trust and EU data

Changelog · what changed in this brief, newest first

Changelog