Catalog / Agents & Process Automation / AGENT-019

Personal admin agent (bills, subscriptions, bureaucracy)

An assistant that watches a person's mail, bank feed and accounts, then tracks bills and deadlines, spots and cancels unwanted subscriptions, and drafts replies to authorities and providers for the person to approve.

Parts needed to run it yourselfAgents & automationDocument readingLanguage model

Private households lose time and money to scattered paperwork: bills in three inboxes, forgotten subscriptions, silent price increases, authority letters with deadlines. A personal admin agent reads the incoming material, builds one picture of obligations and recurring costs, and prepares the next action (pay, cancel, object, file) for the person to approve with one click. The value is fewer missed deadlines and avoided waste, not full autonomy.

Problem & value

  • Today: letters, emails, app notifications and bank statements are handled ad hoc. Subscriptions renew silently, tariffs rise after promo periods, deadlines (objection periods, tax, insurance notice periods) are missed. Cancellation flows are often deliberately cumbersome.
  • AI value: unify unstructured sources, extract amounts, dates and obligations, detect recurring patterns in transactions, and draft the right letter or form. Agents with browser control can also walk through portals that offer no API.
  • Evidence for the pain: the EU consumer-law "digital fairness" work and Commission sweeps of online subscription practices document dark patterns and hard-to-cancel contracts [3]. Germany already requires a cancel button for online-concluded continuing contracts (Kündigungsbutton, §312k BGB) [4]. No reliable public figure for average household savings exists; measure your own.
  • Honest status: assistants that read and draft are mainstream; agents that act in the world (log in, cancel, pay) are early and fragile.

Variants

  • Consumer / household: single person or family; mail + bank + calendar.
  • Self-employed / micro-business: adds invoices, tax advance payments, insurance and licence renewals; overlaps DOC-001/DOC-002.
  • Caregiver mode: admin on behalf of an elderly relative (needs power of attorney, strong consent and audit log).
  • Employer benefit: offered as a financial-wellbeing service (B2B2C), with strict data separation from the employer.

How it works

  1. sources (mail, scans, bank feed, calendar)
  2. ingestion + OCR/VLM parsing + classification (bill, contract, authority letter, offer, spam)
  3. extraction (sender, amount, due date, IBAN, reference, deadline, legal basis)
  4. knowledge store (obligations, contracts, subscriptions, deadlines) with provenance links
  5. planner/agent (what is due, what is unusual, what should I do?)
  6. proposed actions: payment proposal | calendar entry | cancellation letter | objection draft | portal task
  7. human approval (always for payment, cancellation, legal statements)
  8. execution via API, email draft, or browser agent; result logged
  • Recurring-charge detection: rules on transaction history (same payee, regular interval, stable amount) combined with an LLM to normalise merchant names and match them to contracts.
  • Letter understanding: VLM/LLM reads the letter, classifies the type and deadline, and explains it in plain language (see PERS-008).
  • Action layer: tool calls to calendar, email drafts, banking APIs (read-only) and a computer-use browser agent for portals (AGENT-003). Tools are exposed via MCP or function calling.
  • Human-in-the-loop: payments, cancellations, legal declarations and anything sent in the person's name need explicit approval; the agent has read-only bank access.

Data preparation

Data is spread across the mailbox and attachments, paper letters in folders, photos on the phone, bank exports, contracts and insurance policies in cloud drives or drawers, and passwords in various places. Typical gaps: paper originals never scanned, unnamed PDFs, duplicates, contracts without end dates, no overview of recurring charges and scattered access (mail, bank, portals).

Preparation steps: (1) scan paper with a phone or document scanner and run OCR (PREP-006, DOC-014); (2) set up a simple folder scheme and consistent naming (PREP-010); (3) export 12 months of bank transactions as CSV and tidy them (PREP-011); (4) list contracts with provider, price, term and notice period; (5) define categories (insurance, tax, utilities) before automation (PREP-019); (6) define what the agent may read and what stays off limits (PREP-024). Tools: Paperless-ngx, a scanning app, a spreadsheet. Rough share of effort: 30-40 percent; the agent is only as good as the archive it reads. See DATA-003 for document extraction.

Implementation paths

Which path for whom

  • Private: C. Bank, health and tax letters are very sensitive, and local tools (Ollama or LM Studio with a small model, plus Paperless-ngx) run on a normal PC or home server.
  • SMB: B. No mature product covers bills, deadlines and authority letters for a company's specific systems, so a build on cloud APIs integrated with mail, bank feed and accounting fits best.
  • Developer: B. Cloud LLM APIs plus PSD2 aggregators and mail/calendar APIs give the quickest, most capable personal agent.

Path A — Buy (off-the-shelf)

  • Subscription managers and budgeting apps: Finanzguru, Outbank, Bankin' (EU), Rocket Money and Trim (US-centric); some banks embed recurring-payment overviews. Cancellation-service providers (e.g. Aboalarm in DE, EU) send legally formatted letters.
  • Document and mail inboxes: Paperless-style cloud services, digital mailbox providers (Briefkasten services, e.g. Dropscan in DE, EU), cloud assistants in Microsoft 365 Copilot, Google Gemini and ChatGPT agent modes for email and calendar chores.
  • Choose A when: you want reminders and subscription overviews, not full delegation. Check hosting location, bank-connection provider and whether data is used for advertising.
  • Models: a frontier LLM with tool use and structured outputs (Claude, Gemini, GPT, Mistral as EU provider) in an EU region where offered; a VLM for letter photos.
  • Bank data: PSD2 account-information aggregators such as GoCardless Bank Account Data (formerly Nordigen), finAPI (DE), Tink (EU) or Enable Banking; consent must be renewed periodically under PSD2 [5].
  • Orchestration: n8n or a small agent framework; Gmail/Microsoft Graph/IMAP, CalDAV. Browser tasks via Playwright with a computer-use model; connectors via Model Context Protocol [2].
  • Choose B when: you want custom rules and are comfortable operating a small service.

Path C — Self-host (open source)

  • Models: open-weight LLMs (Qwen3, Mistral/Mistral Small, Llama, gpt-oss) served with Ollama or vLLM; Docling or PaddleOCR for scans; Qwen-VL for letter photos.
  • Stack: Paperless-ngx (document archive with OCR and tagging), Firefly III or Actual Budget (finances), n8n, Home Assistant-style home server; see PERS-005.
  • Hardware: HW2 is enough for 14-30B quantised models; the volume of a household is tiny.
  • Trade-off: privacy is maximal; bank access still needs an aggregator or manual CSV export, and browser automation is the main maintenance burden.

Path D — Custom training

Not needed for a single household. Only a provider building a product for many users might fine-tune a small classifier for letter types or merchant normalisation using user corrections.

State of the art (2026)

  • Agent protocols: MCP standardises tool/data connectors; vendors ship email, calendar and browser agents that act with the user's logins [1][2].
  • Reading is solved, acting is not: document extraction is strong; multi-step portal navigation (login, 2FA, captcha, changing layouts) still fails often and needs supervision.
  • Prompt injection is the central security problem: a malicious email or PDF can instruct an agent that has mail, bank and browser access. Mitigate with least privilege, separating untrusted content from instructions, and confirming sensitive actions.
  • Failure modes: wrong due date or IBAN, misread merchant, hallucinated legal claims in letters, cancelling a contract that is still needed. Always show the source snippet.
  • Open-banking limits: transaction text is messy and consent expires; not every bank is covered.

Shared resources & combinations

Effort, cost & KPIs

  • Path A: hours to set up; subscription fees of the apps.
  • Path B: a few days to a few weeks for a working prototype; model API cost is low at household volume.
  • Path C: weekend to several weeks; hardware and upkeep are the cost.
  • KPIs: deadlines missed (target zero), recurring costs identified and cancelled (EUR per year), share of drafts accepted unchanged, time spent on admin per month, wrong-action incidents.

People & change

For private use, the person alone decides; involve a partner or family member who shares accounts and agree what the agent may see. For an SMB owner using it for company admin, involve management (what the agent may do, spending limits), the bookkeeper or tax advisor, IT (accounts, secure credentials, GUIDE-013) and the data protection officer if employee or customer data is in the mail (GUIDE-005). A bank-feed connection needs an AVV with the provider.

Works council: no, as long as it is used for personal or owner-level admin and does not monitor employees; if rolled out to staff mailboxes, §87(1) no. 6 BetrVG would apply, so keep it out of shared mailboxes.

Start with a pilot: read-only mode on one mailbox for a month, then drafts only. Typical concerns: "it could pay or cancel something wrongly" and "who sees my bank data." Address them with strict approval for every payment, cancellation or message, no passwords in prompts, local processing for the most sensitive items, and a log of every action. Basic AI literacy matters: the agent can misread letters (GUIDE-006, GUIDE-007).

Risks, regulation & compliance

  • EU AI Act: typically minimal risk for personal use; AI literacy (Art. 4) applies to providers and professional deployers. If the agent chats with third parties or generates content in a way that could be mistaken for a human, Art. 50 transparency may apply to the provider. Credit-scoring or eligibility assessment features would move a product towards Annex III high-risk. Annex III high-risk obligations apply from 2 Dec 2027 (Digital Omnibus on AI, Reg. (EU) 2026/1744) [6].
  • GDPR: mail and bank data are highly sensitive (special categories can appear: health insurer, church tax, union fees). The household exemption covers a private person's own use; a provider offering the service is a controller or processor and needs a legal basis, DPA, EU hosting and deletion on request. Acting for a relative requires authority.
  • PSD2: only licensed or registered account-information providers may access bank data; read-only consent, 90-day-style re-authentication cycles apply.
  • Security: least-privilege tokens, no payment initiation by default, secrets outside prompts, full action log, kill switch.
  • Legal: letters drafted by AI are not legal advice; in Germany the Rechtsdienstleistungsgesetz limits legal services by providers. Contract terms and notice periods must be verified.

Getting started

  1. Pick the top pain (e.g. subscriptions or deadlines) and export 12 months of bank CSV.
  2. Run recurring-charge detection and review the list manually; cancel the obvious waste.
  3. Set up a forwarding address or Paperless-ngx inbox; let the agent extract deadlines into a calendar.
  4. Add draft-only letter generation with source snippets; approve every send.
  5. Only after weeks of clean logs consider browser automation, with read-only banking and no auto-payment.

Sources

  1. Anthropic — Model Context Protocol — https://modelcontextprotocol.io
  2. MCP specification (authorization and tools) — https://modelcontextprotocol.io/specification
  3. European Commission — Digital fairness / consumer protection and dark patterns — https://commission.europa.eu/law/law-topic/consumer-protection-law_en
  4. §312k BGB (Kündigung von Verbraucherverträgen im elektronischen Geschäftsverkehr) — https://www.gesetze-im-internet.de/bgb/__312k.html
  5. GoCardless Bank Account Data (formerly Nordigen) documentation — https://developer.gocardless.com/bank-account-data/overview
  6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  7. Paperless-ngx — https://github.com/paperless-ngx/paperless-ngx
  8. Directive (EU) 2015/2366 (PSD2) — https://eur-lex.europa.eu/eli/dir/2015/2366/oj