Most people have accounts, loyalty cards and mailing-list entries at hundreds of companies, plus profiles at data brokers they never heard of. The GDPR gives them rights to access, erase and object, but exercising them is tedious: find the right contact, word the request, track the one-month deadline, read the answer. An AI agent finds the likely holders, drafts and sends the requests, tracks replies and escalations, explains the returned data and watches for re-listing.
Problem & value
- Today: individuals use templates (e.g. datenanfragen.de), write letters by hand or pay a broker-removal subscription. Requests get lost, controllers answer late or incompletely, and nobody checks whether data came back.
- Legal baseline: Art. 12(3) GDPR requires a response without undue delay and within one month (extendable by two months for complex or numerous requests, with notice). Art. 15 gives access plus a copy, Art. 17 erasure, Art. 21 objection (absolute for direct marketing), Art. 20 portability. Requests are free of charge in principle; controllers may refuse only manifestly unfounded or excessive ones [1][2].
- AI value: discovery (which controllers?), drafting in the controller's language, classification of replies (complete, partial, refusal, identity-check demand), summarising a 200-page export, and deadline discipline. No reliable public figures exist on average time saved; measure it yourself.
Variants
- Consumer app: inbox scan, one-click requests, dashboard (Mine, Incogni-style services).
- Broker-removal subscription: recurring opt-out requests to people-search and marketing-data brokers, mostly US-centred, increasingly EU/UK.
- Self-hosted personal agent: local LLM + mail client, nothing leaves the home (see PERS-005).
- Family/guardian mode: requests on behalf of children or elderly relatives (authority proof required).
- Specific targets: credit agencies (SCHUFA in Germany), adtech/RTB data, employers, former landlords, social platforms.
- Employer side of the same request: SEC-018, DOC-018, DOC-021.
How it works
- discovery (inbox scan
- sender domains; account export; broker search by name/address)
- controller resolution (privacy contact / DPO email / portal from register or privacy policy)
- request drafting (Art. 15 / 17 / 21 / 20, language, identity minimisation)
- HUMAN APPROVAL
- send (email, web form, postal letter)
- deadline tracker (day 30, reminder, extension notice, escalation draft)
- reply triage (classify: data delivered / refused / ID check / silence)
- data export analysis (summary, surprising categories, third-party recipients, Art. 15(1) info)
- follow-up (erasure confirmation, complaint to supervisory authority draft)
- monitoring (re-listing checks, new breach/sign-up alerts)
- Discovery uses email headers and receipts, classified by an LLM or small classifier; broker search via scraping or browser agents.
- Drafting is template-driven with LLM adaptation; legal wording stays in vetted templates, not free generation.
- Human-in-the-loop: user approves every outgoing request and every identity document submission. The agent never sends a copy of an ID without explicit confirmation, and should propose redacted copies where the law allows.
Data preparation
The main data source is your own mailbox: registration mails, newsletters, receipts and account notifications show which companies hold your data. Further sources are the browser's saved passwords and logins, app lists, loyalty cards and old accounts. Typical gaps are old mail addresses, deleted mailboxes, accounts without recoverable e-mail, and unknown data brokers that never wrote to you.
Preparation steps: - Make a list of all mail addresses and phone numbers you used over the years. - Export or index the mailbox locally (IMAP plus notmuch or an export to mbox/EML, PREP-007), and search for sender domains instead of reading mail bodies. - Build a table of controllers: name, contact for data requests, request date, deadline, answer (PREP-011). - Pull controller contact details from the datenanfragen.de database instead of researching them manually. - Separate sensitive accounts (health, finance) and decide whether to include them.
Effort is low: the AI step is small, and most time goes into building the account inventory and tracking answers. Expect a first pass in a few hours and follow-up over weeks because controllers have a one-month response period under Art. 12 GDPR. See GUIDE-003 for the inventory approach.
Implementation paths
Which path for whom
- Private: A. The free datenanfragen.de generator and consumer-protection templates need no mailbox access; paid removal services only make sense if you accept a new holder of your data.
- Developer: C. Local mail indexing with a small open-weight model on HW1 keeps mailbox content on your device and the task (classification, drafting) is simple.
Path A — Buy (recommended for most people)
- Mine (saymine.com): inbox-based discovery and automated requests, Israel-founded; MineOS sold the consumer app to McAfee in Nov 2025, which is folding it into its identity and privacy protection offering, so check current EU availability. Incogni (Surfshark/Nord group) and similar subscription removal services target brokers. datenanfragen.de (Germany/EU, non-profit, open source) provides a free generator and a controller database. Also consider noyb's resources and consumer-protection organisations' templates (e.g. Verbraucherzentrale).
- Check: where is your mailbox data processed, what OAuth scopes are asked for (read-only?), whether the vendor itself becomes a new holder of your data.
Path B — Build on cloud APIs
- LLM with structured outputs (EU-region deployments of Claude, GPT, Gemini, or Mistral as EU provider) to classify mails and draft requests; Gmail/Microsoft Graph/IMAP for mail access; calendar/ICS for deadlines.
- Orchestrate with n8n or a small agent framework; keep the controller database and templates as versioned data. Send only sender domains, not mail bodies, to the model where possible.
Path C — Self-host (open source)
- Local mail indexing (notmuch, IMAP + SQLite) and a small open-weight LLM (Qwen, Llama, Mistral families, Gemma) via Ollama or llama.cpp; Mistral Small-class models run on HW1/HW2 and are adequate for classification and drafting.
- Datenanfragen's open-source generator and controller data can be embedded; Playwright-based browser agents handle web-form opt-outs.
- Ops effort is low; the hard part is keeping the controller database current.
Path D — Custom training
Rarely useful. A fine-tuned reply classifier is only worth it for a service operator with thousands of labelled controller replies.
State of the art (2026)
- LLMs reliably classify reply types, draft requests in many EU languages and summarise exports. Browser/computer-use agents can fill opt-out forms but are brittle against CAPTCHAs, e-mail verification loops and changed layouts.
- Controllers increasingly provide self-service portals; EDPB Guidelines 01/2022 on access clarify scope (all data, no need to justify the request, limited room to refuse) and are the best reference for drafting [2].
- Typical failure modes: wrong controller or DPO address; requests that treat Art. 17 as unconditional (it has exceptions such as legal retention duties); hallucinated legal citations; missed identity-check replies that restart the clock; brokers re-importing data after removal.
- Platform-level opt-out mechanisms (e.g. the EU Data Act and DMA portability duties for gatekeepers, browser-level signals such as Global Privacy Control) complement but do not replace requests.
Shared resources & combinations
- Same stack (llm-general + agent-framework + workflow-automation): AGENT-001 (inbox triage), AGENT-003 (browser agents), PERS-008 (understanding official letters, escalation to authorities), PERS-015 (consumer-rights self-help).
- PERS-005: local runtime so mail content never leaves the home.
- Organisation-side counterpart: SEC-018 (DSAR automation), DOC-018 (redaction), DOC-021 (eDiscovery). Requests from this agent land in those systems.
- Reply parsing reuses document-parsing pipelines (ocr-docparse via DOC family) for postal answers.
Effort, cost & KPIs
- Path A: minutes to set up; subscription pricing varies by vendor (check current price pages); free generators cost nothing.
- Path B: a few days for a personal prototype; API cost per request is small.
- Path C: a weekend to a week; one existing laptop or mini-PC.
- KPIs: controllers contacted, response rate within one month, share of full vs partial answers, erasure confirmations, complaints filed, re-listings detected, user time per request.
People & change
This is a personal use case. The people involved are the user and, if it concerns them, family members; the counterparties are controllers and their data protection contacts. Escalation goes to the data protection supervisory authority or a consumer organisation such as Verbraucherzentrale or noyb if a controller does not answer.
Data protection: the sensitive point is the tool itself. Prefer read-only mailbox access, send only sender domains rather than mail bodies to cloud models, and check that a removal service does not become another holder of your data (GUIDE-005). No DPIA is needed for private use. If a company offers this as a service for employees or customers, a DPIA, an AVV with every processor and the DPO would be required.
Works council: no, since no employer system is involved. If used on a company device, follow the company AI policy (GUIDE-007) and do not process company mail.
AI literacy: understand that a model can misclassify a sender and that the legal text of a request still has to be right; read each drafted request before sending, keep copies and set reminder deadlines yourself (GUIDE-006). Start with five or ten controllers as a pilot to see response quality. The typical concern is "will this create more spam or trouble"; the answer is to send the requests from a dedicated address and keep control over what is sent. The human always approves sending (GUIDE-008 for habits that help).
Risks, regulation & compliance
- EU AI Act: minimal risk; consumer tool with no Annex III use. AI literacy (Art. 4) applies to providers and deployers. If the agent interacts with humans via generated text, Art. 50 transparency is not triggered by letters to controllers, but a chatbot front end would be limited-risk.
- GDPR: the tool processes the user's own data (and contact data of DPOs). Vendors become processors or controllers; read their privacy policy, minimise OAuth scopes, avoid uploading full mail bodies. Requests on behalf of others need a mandate.
- Abuse: mass or repetitive requests can be refused as excessive; do not use the agent for harassment or to trace third parties. Never request data of another person.
- Identity fraud: requests are an attack vector (fake access requests to harvest data); controllers may demand ID proof, and users must send the minimum necessary.
- Mitigations: vetted legal templates, human approval, redacted ID copies, logging, escalation path to the supervisory authority (Art. 77) and optional collective redress/noyb contact.
Getting started
- Pick 10 likely holders (banks, telecom, big platforms, credit agencies, retailers) and send Art. 15 requests with a free generator.
- Record sending date and day-30 deadline in a calendar or spreadsheet.
- Try one inbox-based discovery tool or a local script over sender domains.
- Use a local or EU-hosted LLM to summarise the first export and spot unexpected recipients.
- Escalate silent controllers with a reminder, then a complaint to the competent supervisory authority.
Sources
- Regulation (EU) 2016/679 (GDPR), Arts. 12, 15, 17, 20, 21, 77 — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- EDPB — Guidelines 01/2022 on data subject rights, right of access — https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-012022-data-subject-rights-right-access_en
- Datenanfragen.de (open-source GDPR request generator) — https://www.datenanfragen.de
- Mine (data-discovery and request service) — https://www.saymine.com
- Incogni (data-broker removal) — https://incogni.com
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Global Privacy Control specification — https://globalprivacycontrol.org
- noyb — European Center for Digital Rights — https://noyb.eu
- Security Systems News — MineOS completes sale of consumer privacy business (to McAfee) — https://www.securitysystemsnews.com/article/mineos-completes-sale-of-consumer-privacy-business