Catalog / Everyday Automation & AI Micro-Tools / AUTO-007

Automatic filing & renaming of email attachments

A small flow saves incoming email attachments (delivery notes, certificates, invoices, drawings) to the right folder with a clean, AI-generated file name and basic tags - for office teams that still drag attachments around by hand.

Quick win: proven, manageable effort, first results within weeks, low regulatory risk.

Parts needed to run it yourselfAgents & automationDocument readingLanguage model

Every office has the same chore: an email arrives with scan0042.pdf, someone downloads it, renames it, and drags it into a project or customer folder - or forgets, and the file lives only in the mailbox. A small flow (Power Automate, n8n, Paperless-ngx) can do this on arrival: read the attachment, let a language model decide the document type, sender and reference, build a consistent file name, store the file in the right place and tag it. Unsure cases go to a review folder. This is deliberately small: days of work, no new platform, and a good first step for data preparation (PREP-009, PREP-010).

Problem & value

  • Today: attachments are saved by hand or not at all; names are inconsistent and files sit in personal folders or the mailbox. When a colleague is absent, nobody finds the document.
  • AI value: rules ("sender contains X") break on new senders and generic file names. A model reads the content and fills document type, reference, date and project, giving a searchable archive and a cleaner base for later AI search or a Copilot rollout.
  • Boundary: this is filing, not accounting. Invoice extraction is DOC-001; central mailroom classification at scale is DOC-008.
  • No reliable benefit figures exist for this niche; measure your own baseline.

Variants

  • Shared mailbox (office@, orders@): best start. Personal mailbox: per-user flow with a category.
  • Craft/construction: certificates, delivery notes, drawings per project number.
  • Public sector: incoming mail into the e-file/DMS, with staff council involvement.

How it works

  1. new email (shared mailbox, trigger)
  2. skip noise (signature images, logos, tiny files, newsletters)
  3. read attachment (text layer, or OCR for scans)
  4. LLM: document type, sender/party, date, reference, project (JSON, fixed schema)
  5. lookup: map party/project against customer or project list
  6. build name YYYY-MM-DD_party_type_reference.pdf + target folder
  7. save to SharePoint / Nextcloud / file share, set tags
  8. low confidence or no match
  9. "to-check" folder + daily list for a human
  • AI part: a small or mid-size model with a fixed JSON schema is enough. The prompt contains the allowed document types and your naming rules; the flow, not the model, builds the final name and path. Never let the model invent folder paths.
  • Human in the loop: unsure or unmatched files are not guessed; they go to a review folder. The original mail stays in the mailbox, so nothing is lost if the flow errs.

Data preparation

Which data: a folder structure and naming convention, lists of customers, suppliers and projects, and 30-50 real example attachments with the correct target. Where it lives: shared mailboxes, SharePoint/Nextcloud/file shares, project numbers in Excel or the ERP.

Typical gaps: no agreed convention, private project lists, scans without text, folders with too wide access rights.

Preparation steps: 1. Agree 5-10 document types and one naming pattern (PREP-019). 2. Export customer, supplier and project lists to one clean table (PREP-005, PREP-011). 3. Tidy target structure and tags (PREP-009, PREP-010); check access rights (PREP-023). 4. Keep the example attachments as test set (PREP-018).

Tools: Excel/SharePoint lists, Power Automate or n8n, scanner with OCR.

Rough share of effort: about 50 percent - the rules and lookup lists take longer than the flow. See GUIDE-003.

Implementation paths

Which path for whom

  • SMB - Path B: the flow must fit your own folders, project numbers and naming rules, so no product fits well; Power Automate (M365) or self-hosted n8n with an EU-region model is built in days. Path C where documents are sensitive.
  • Enterprise - Path B: a flow platform usually exists (Power Automate, Logic Apps, n8n); one governed pattern is reused by many teams.
  • Public - Path C: mail holds citizens' data; sovereignty rules favour self-hosted or German-cloud setups.

Path A - Buy (off-the-shelf)

Outlook rules plus Power Automate templates (no AI), or DMS products with mail intake: DocuWare, d.velop, ELO (German vendors). Choose A when senders and names are stable or a DMS already exists.

Path B - Build on cloud APIs

  • Power Automate: trigger "When a new email arrives in a shared mailbox (V2)" (set Include Attachments to Yes; for personal mailboxes "When a new email arrives (V3)"), loop over attachments, AI Builder or a model API, SharePoint "Create file" [5]. New cloud flows stay in the EU Data Boundary for EU tenants, but Copilot inferencing can route outside it unless an admin switches this off, and some third-party models are excluded [3][4]. Verify the AI Builder location in your tenant.
  • n8n (self-hosted in the EU or n8n Cloud EU): Email Trigger (IMAP) with attachment download, then an LLM node [1]. EU model options: Mistral, Azure OpenAI in an EU region, IONOS AI Model Hub, STACKIT.
  • Google Workspace: Apps Script or Gmail filters plus Drive and the Gemini API.
  • Leave posting of invoices to DOC-001.

Path C - Self-host (open source)

  • Paperless-ngx: mail rules fetch IMAP mail, filter by sender or subject and assign tags, correspondent and document type [2]. GoBD process documentation is your task.
  • n8n + Ollama: a small model (3-8B class) classifies and names; short text in, JSON out, so HW1 suffices.
  • Nextcloud: Flow plus Assistant with a local model.
  • Choose C for HR, health, legal or customer data, or when the same server serves AUTO-008, AUTO-012 and AUTO-016.

Path D - Custom training

Not needed; a schema plus 30-50 examples in the prompt is enough.

State of the art (2026)

  • LLMs with structured output read most business PDFs reliably; weak points are poor scans, multi-document PDFs and unusual formats (CAD, ZIP). Unknown type goes to review.
  • Models guess confidently: use a confidence field, a "none of the above" option and cross-check against lookup lists. Add a hash check against duplicates.
  • Prompt injection: a hostile PDF can contain instructions. The model only fills a schema; the flow has no rights beyond "save file" (GUIDE-013).

Shared resources & combinations

Effort, cost & KPIs

  • Path A: hours to a day. Path B: one to three days for a first mailbox incl. conventions and tests. Path C: two to five days incl. server setup.
  • KPIs: share filed correctly untouched, share in review, minutes saved per week.

People & change

Involve early: department lead and end users decide folders and names; IT builds and secures the flow; data owners confirm lookup lists. The DPO checks personal data in attachments and the AVV (Art. 28 GDPR); a DPIA is normally not needed, but check if applicant or employee documents arrive (GUIDE-005, GUIDE-010).

Works council: depends. The flow logs senders and times; on a functional mailbox without individual evaluation risk is low, yet objective suitability for monitoring (§87(1) no. 6 BetrVG) and §90(1) no. 3 (timely information on AI-supported work) still apply. Personal mailboxes need more care; public bodies involve the Personalrat (GUIDE-004).

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; a short briefing on misreads and the review folder is enough (GUIDE-006).

Pilot: one shared mailbox, one team, two to four weeks in parallel with manual filing, one champion; present it as relief from drag-and-drop (GUIDE-008, GUIDE-009).

Concerns: "my files vanish" (the original mail stays), "AI reads our mail" (one documented mailbox).

Risks, regulation & compliance

  • EU AI Act: minimal risk, no Annex III use; Art. 4 literacy support applies (GUIDE-014).
  • GDPR: attachments contain personal data; limit the flow to defined mailboxes, EU-hosted processing, AVV, deletion rules; health or HR documents argue for Path C.
  • Records: accounting documents need tax-law retention (GoBD); a renamed copy must not replace the original.
  • Security: least-privilege service account, malware scan before processing, action log (GUIDE-013).
  • Mitigations: review folder, confidence threshold, monthly sample check (GUIDE-011).

Getting started

  1. Pick one shared mailbox; agree document types and one naming pattern.
  2. Collect 30-50 real attachments with the correct folder and name as test set.
  3. Build the flow in Power Automate or n8n with a review folder.
  4. Run it in parallel with manual filing for two weeks, then adjust.
  5. Clear data protection and works council points, then extend.

Sources

  1. n8n documentation - Email Trigger (IMAP) node - https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-base.emailimap
  2. Paperless-ngx documentation - Basic usage (mail accounts and rules) - https://docs.paperless-ngx.com/usage/
  3. Microsoft Learn - Data locations in Copilot Studio - https://learn.microsoft.com/en-us/microsoft-copilot-studio/data-location
  4. Microsoft Learn - Services that transfer data out of the EU Data Boundary on an ongoing basis - https://learn.microsoft.com/en-us/privacy/eudb/eu-data-boundary-ongoing-partial-transfers
  5. Microsoft Q&A - Rename and save PDF attachments from Outlook - https://learn.microsoft.com/en-us/answers/questions/5823916/rename-and-save-pdf-attachments-from-outlook
  6. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj