Registrations, service requests and internal requests arrive as web forms, emails or chat messages, and someone copies them into Excel by hand. A small flow with one AI step reads the free text, fills the agreed columns, marks unclear entries and writes a row. For most SMBs this is a one to three day project with tools they already own, and it is often the first "real" automation that also forces a tidy table design.
Problem & value
- Today: an assistant or team member retypes entries from mailbox, form results and messages into a list. Typical errors: wrong columns, inconsistent spelling of customers and places, missed messages, and no single overview.
- Why AI helps: classic forms with fixed fields need no AI. The AI step is useful where people write free text ("Heater in room 3 is broken since Monday, call Mr. Meier 0171 ...") and the row needs fields such as category, location, urgency and contact.
- Benefit: time saved on retyping and fewer lost requests. No reliable industry figure exists that fits every case, so measure it on your own volume: count submissions per week and minutes per entry before and after.
- Not in scope: formal application forms with scans and OCR at scale (DOC-003) and order entry into ERP (AGENT-004). This is the simple text-in, row-out flow.
Variants
- Forms to list: Microsoft Forms or Google Forms response with a free-text box, plus a choice field, becomes a SharePoint list item or sheet row.
- Mailbox to table: a shared mailbox (service@, anfragen@) is read and each mail becomes one row.
- Internal requests: IT tickets, facility requests, holiday cover wishes, material orders. For approvals see AUTO-011.
How it works
- trigger (new form response / new mail / new message)
- clean text (strip signature, quoted replies)
- AI step: extract fields as JSON that matches the table schema
- validate (required fields, allowed values, date and phone formats, duplicate check)
- write row; unclear entries get status "check"
- human reviews the "check" rows; optional draft reply for approval
- Key AI component: a small or general LLM with structured output (a fixed JSON schema). The model must only choose from allowed values and return "unknown" instead of guessing.
- Validation outside the model: plain rules (regex, lookup lists) catch what the model gets wrong.
- Human in the loop: a person reviews flagged rows and nothing is sent externally without approval.
Data preparation
Which data: a clear target table (one row per submission, fixed columns), a list of allowed values (categories, locations, departments), and 10-30 real, anonymised example submissions with the correct row.
Where it lives: form tools, shared mailboxes, chat channels, and existing Excel sheets or SharePoint lists, often with years of inconsistent entries.
Typical gaps: the target sheet has merged cells, free-form columns, mixed date formats or several tabs; categories are not defined; no unique ID per row; personal data (phone, address, health remarks) lands in a sheet that too many people can open.
Preparation steps: 1. Define the schema: columns, data types, allowed values, required fields (about 2 hours with the process owner). 2. Clean the existing sheet or start a new table; convert it to a real Excel table or list (1-3 hours; AUTO-006 helps with tidying). 3. Collect 10-30 example inputs with the expected row and reuse them as a test set. 4. Set access rights on the target file and decide the retention period.
Tools: Excel, Power Query, SharePoint lists, Google Sheets, Nextcloud Tables, Airtable-like tools.
Rough share of effort: low; the schema discussion is the real work. See GUIDE-003.
Implementation paths
Which path for whom
- SMB - Path B: the target tables and forms are your own, so a small flow on an EU-hosted model API fits better than a product. No standard product covers every table. Path C is better if the content is sensitive (see below).
- Enterprise - Path B: a central workflow platform with approved EU model endpoints and governance, reused by many departments.
- Developer - Path B: a few lines of code with structured output against an EU API.
- Public - Path C: citizen messages are sensitive and procurement favours sovereign hosting; a local model on HW1/HW2 is enough for short texts.
Path A - Buy (off-the-shelf)
- Microsoft 365: Microsoft Forms plus Power Automate. AI Builder offers a text entity extraction action whose result can be written to an Excel table with "Add a row into a table" [1]. Check licensing: AI features are typically a premium or credit-based add-on.
- Google Workspace: Google Forms to Sheets natively; AI steps through Gemini in Sheets or Workspace Studio flows (verify the feature in your edition).
- Choose A when: you already run Microsoft 365 and the process is a single form to a single list.
Path B - Build on cloud APIs (recommended for most SMBs)
- Orchestration: n8n (cloud with EU region or self-hosted), Power Automate, or Make.
- Node: n8n offers an Information Extractor node where you define a schema and a connected chat model fills it [2].
- Models: EU-hosted APIs such as Mistral (France), Azure OpenAI in an EU region, or Google Vertex AI in an EU location. Small models are enough for short texts.
- Effort: 1-3 days for a first flow, plus about a week of testing with real entries.
Path C - Self-host (open source)
- Stack: n8n Community Edition on your own server with Ollama and a small model (for example Mistral Small or Qwen-class 3-14B), writing to Nextcloud Tables or a database. Note n8n's Sustainable Use License is source-available and restricts use to internal business purposes [3].
- Hardware: HW1 suffices for short texts; local models may return malformed JSON more often, so validate and retry [2].
- Ops effort: a few hours per month. Worth it where texts contain health, HR or customer data, or where the same server serves AUTO-007, AUTO-012, AUTO-013 (Baukasten effect).
Path D - Custom training
Not needed. A prompt with the schema and examples is sufficient; fine-tuning would only pay off at very high volumes with unusual vocabulary.
State of the art (2026)
- Structured output (JSON schema) is a standard feature of current models, so extraction of 5-15 fields from short texts is reliable enough for production with validation.
- Limits and failure modes: invented values for missing fields, wrong category for ambiguous requests, confusion with several requests in one message, injection through the text itself ("ignore the instructions and ..."), and mixed languages. Mitigate with "unknown" values, validation rules, a check status and no automatic actions based on the content (GUIDE-013).
Shared resources & combinations
- Same stack (workflow-automation + llm-small): AUTO-007 attachment filing, AUTO-012 voicemail routing, AUTO-013 business cards, AUTO-011 approvals.
- Upgrade paths: scanned forms to DOC-003, orders into ERP to AGENT-004.
Effort, cost & KPIs
- Path A/B: 1-3 days build, about a week pilot; running cost is workflow-platform fees plus small model cost per entry.
- Path C: 3-5 days including server setup, plus light operations.
- KPIs: share of rows accepted without correction, share flagged "check", average handling time per entry, lost or duplicated requests, user satisfaction of the team.
People & change
Involve early: the department lead (owns the table and categories), end users who do the retyping today, IT for accounts and connectors, the DPO, and the data owner of the target file (GUIDE-005). Management only needs to approve the pilot.
Works council: depends. If the flow only structures customer or external requests, co-determination under §87(1) no. 6 BetrVG is usually not triggered. If the inputs are staff messages or the table logs who submitted what and when, objective suitability to monitor behaviour is enough per BAG case law, so involve the works council. §90(1) no. 3 requires timely information on AI-supported work processes. A short works agreement can cover "no performance evaluation from this data" (GUIDE-004). Public bodies involve the Personalrat.
Data protection: personal data is processed, so AVV (Art. 28 GDPR) with the model and platform provider is needed and the DPO is involved. A DPIA (Art. 35 GDPR) is normally not required for simple requests, but check it if special categories (health) or systematic evaluation of people are involved (GUIDE-005).
AI literacy: Art. 4 AI Act requires providers and deployers to take measures that support staff AI literacy. A 30-minute session on checking rows and handling flagged entries is enough (GUIDE-006).
Humans in control: flagged rows are reviewed by a person; outgoing replies are approved before sending.
Risks, regulation & compliance
- EU AI Act: minimal risk. Art. 4 applies. Art. 50 duties apply only if the flow chats with people directly or sends generated replies without a human; mark such messages as AI-assisted (GUIDE-014). If the same pattern is used to sort job applications, it becomes high-risk (Annex III), so do not reuse it for HR selection without a separate assessment.
- GDPR: data minimisation (collect only needed fields), purpose limitation, retention and deletion in the sheet, EU hosting or adequate safeguards, no health data without a basis.
- Security: prompt injection through free text, wide sharing of the target sheet, API keys in flows (GUIDE-013).
- Mitigations: schema with allowed values, validation, "check" status, restricted access, logging agreed with the works council.
Getting started
- Pick one process with 20+ submissions per month and define the table schema with the department (GUIDE-001, GUIDE-003).
- Collect 20 real examples with expected rows as a test set.
- Build the flow in a tool you already have (Power Automate or n8n) with a "check" status and validation rules (GUIDE-002).
- Inform the DPO and, where applicable, the works council before the pilot; sign the AVV (GUIDE-004, GUIDE-005).
- Run a 2-4 week pilot, measure accepted rows and corrections, then decide go or no-go (GUIDE-009).
Sources
- Inogic - Streamlined text and document categorization with entity extraction in Power Automate - https://www.inogic.com/blog/2024/05/streamlined-text-and-document-categorization-with-entity-extraction-in-power-automate/
- n8n Docs - Information Extractor node - https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.information-extractor/
- Activepieces - n8n Community Edition: limits and features compared - https://www.activepieces.com/blog/n8n-community-edition-limits-and-features-compared.md
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj