Catalog / Everyday Automation & AI Micro-Tools / AUTO-027

Software how-to help & on-the-job AI tutoring for office tools

Employees ask an AI assistant how to do something in Excel, Word, DATEV, the ERP or CAD and get step-by-step guidance, explanations of error messages and short exercises, reducing internal support requests.

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

Parts needed to run it yourselfAgents & automationDocument readingLanguage modelSearch & RAG

Every office has the same questions: "How do I do a pivot table?", "Why does DATEV show this message?", "Where is the export in our ERP?". Today they go to the colleague who "knows it", to the IT hotline or to a web search that returns outdated forum posts. An AI assistant that knows the tool, sees the screenshot and answers in the employee's language gives a step-by-step answer in seconds, explains errors and offers a short exercise. It is workplace self-help, not a training programme, and it takes load off the IT helpdesk.

Problem & value

  • Today: knowledge sits in a few heads. Colleagues interrupt each other, IT answers the same "how do I" tickets again and again, and many staff use only a fraction of the software they already pay for.
  • AI value: immediate, context-specific answers in the moment of need ("learning in the flow of work"), explanations of error messages, and simple practice tasks. Experts keep their focus; new hires and part-timers become productive sooner.
  • Evidence: vendor pilot results are mostly self-reported; measure your own baseline (internal how-to requests, time to solve).
  • Delimitation: formal learning programmes and tutoring are EDU-004; solving or routing real IT incidents is AGENT-016. Here the employee stays in control and asks for help.

Variants

  • Assistant inside the application: Copilot in Excel/Word/Outlook or Gemini in Google Workspace, which can answer "how do I...?" about the application in use.
  • Company-specific help bot: a chat over internal work instructions, ERP manuals and "how we do it here" pages (see KNOW-002, KNOW-004).
  • Screenshot helper and practice mode: explain a pasted error dialog; create a small exercise with sample data.
  • Private use: learning a tool with a chat assistant or a local app such as LM Studio.

How it works

  1. question (text or screenshot)
  2. context: tool, version, language (from the app or the user)
  3. retrieval: vendor documentation + internal manuals (optional)
  4. LLM answer: numbered steps, warnings, "where to click"
  5. source links + "was this correct?" feedback
  6. optional exercise with sample data
  • Key AI components: a general LLM with current product knowledge, optionally a vision model for screenshots, and retrieval over internal documents so that answers match your menus, templates and rules.
  • Human-in-the-loop: the employee performs the steps and sees immediately whether they work. Anything that changes data in bulk (delete, overwrite, post) must be tried on a copy first. Unsolved cases go to the colleague or IT.

Data preparation

Which data: mostly none for generic questions. For company-specific help: ERP and DATEV work instructions, internal process descriptions, FAQ and wiki pages, past helpdesk tickets (as a source of recurring questions), and templates.

Typical gaps: instructions for old software versions, long PDFs, contradictory versions, tickets with customer names or passwords, and knowledge that exists only in colleagues' heads (SharePoint, Nextcloud, wiki, file shares).

Preparation steps: 1. Collect the 20 most frequent how-to questions from the helpdesk tickets and from a short staff survey (half a day). 2. Pick 10-30 current documents per tool, mark outdated ones and note the software version (1-2 days). 3. Remove personal data and credentials from tickets and screenshots. 4. Write down unwritten rules as short instructions (PREP-016) and let experts review first answers. 5. Check access rights so the assistant only sees what every user may read (PREP-023).

Rough share of effort: low, typically 1-3 days for a first version; most value comes from the generic path without any own data. See GUIDE-003.

Implementation paths

Which path for whom

  • SMB - Path A: Copilot Chat or Copilot in Microsoft 365, or Gemini in Google Workspace, are mature, EU-hostable and already known to staff; the process is standard and needs no integration.
  • Enterprise - Path B: a company help assistant over ERP, SAP or CAD documentation needs integration with identity, document sources and a shared AI platform, so cloud APIs with own orchestration fit.
  • Developer - Path C: open-source chat front ends with a local model and retrieval run on HW2 and give full control for product builders.
  • Public - Path C: administrations use specialised technical procedures; procurement rules and confidentiality favour self-operated or sovereign German cloud hosting.
  • Microsoft 365: Microsoft 365 Copilot Chat and the Copilot add-on answer how-to questions about Word, Excel, Outlook and Teams; data stays inside the tenant with enterprise data protection, and Microsoft offers the EU Data Boundary [2][3].
  • Google Workspace: Gemini in Workspace apps, with Workspace data protection terms.
  • Other: ChatGPT or Claude business plans; DATEV, Lexware and many ERP vendors ship their own help assistants.
  • EU-hosted options: Mistral Le Chat (France), Aleph Alpha-based offers, IONOS AI Model Hub, STACKIT AI Model Serving (German cloud).
  • Choose A when: staff mainly use mainstream office software and licences already exist. Cost: licence per user per month; check current list prices before ordering.

Path B - Build on cloud APIs

  • Models: Mistral (EU), Azure OpenAI in an EU region, Claude or Gemini via EU regions of the hyperscalers, or IONOS/STACKIT model endpoints.
  • Stack: retrieval over internal manuals (Azure AI Search, OpenSearch, pgvector), Teams or web chat front end, single sign-on, source citations.
  • Choose B when: the main value is in the own ERP or CAD tool that no vendor assistant knows.

Path C - Self-host (open source)

  • Stack: Open WebUI with Ollama or vLLM, an open-weight model of 8-30B parameters (e.g. Qwen, Mistral, Llama or Gemma families), embeddings such as bge-m3, and a small vector store.
  • Hardware: HW2 (one 16-32 GB GPU) is enough for a team.
  • Choose C when: screenshots show customer or trade-secret data that may not leave the building, or the platform also serves AUTO-022 and KNOW-002.

Path D - Custom training

Not worthwhile for this use case. Retrieval over current documents beats fine-tuning because manuals and software versions change.

State of the art (2026)

  • Current LLMs answer common office-software questions well; vision-capable models read screenshots and error dialogs.
  • Limits: menu paths change between software versions and languages, so answers can describe buttons that do not exist ("hallucinated click paths"). Niche tools (DATEV special cases, custom ERP, CAD macros) are weakly covered without retrieval over your documents.

Shared resources & combinations

  • Same stack as AUTO-022 (custom assistants), AUTO-023 (prompt libraries), KNOW-002 (RAG over company documents) and KNOW-004 (technical documentation assistant); one internal chat platform can serve them all.
  • Combinations: AUTO-001 and AUTO-006 as first topics; AGENT-016 where unsolved questions become tickets.

Effort, cost & KPIs

  • Path A: hours to days (enable licence, short policy, a 60-minute kick-off).
  • Path B: 1-3 weeks for a pilot; cost drivers are API usage, search index and maintenance of documents.
  • Path C: about 1-2 weeks incl. hardware; cost drivers are GPU, operations and document upkeep.
  • KPIs: number of internal how-to tickets and interruptions per month, time to solve a typical task, share of answers rated helpful, share of answers with correct source, usage by team, onboarding time of new hires.

People & change

Involve early: management (budget, policy), department leads, end users (they define the real questions), IT (licences, helpdesk), the DPO and, where one exists, the works council (GUIDE-001, GUIDE-008).

Works council: depends. An assistant that logs who asks which questions can show who struggles with which tool; the BAG regards objective suitability for monitoring as enough under §87(1) no. 6 BetrVG, so Copilot-style tools are usually co-determined. §90(1) no. 3 requires timely information on AI-supported work processes; §80(3) allows an AI expert if needed. A short works agreement should exclude individual evaluation of question logs. Public bodies involve the Personalrat (GUIDE-004).

Data protection: prompts and screenshots may contain customer or employee data. Conclude an AVV (Art. 28 GDPR) with the provider, define "no personal data in screenshots" or blur it, and let the DPO check ; a DPIA is normally not needed unless logs are analysed per person (GUIDE-005, GUIDE-010).

AI literacy: Art. 4 AI Act requires providers and deployers to take measures that support the AI literacy of their staff. A 60-minute session on good questions and spotting invented menu paths is enough (GUIDE-006).

Communication and pilot: present the assistant as a patient helper that frees experts from repeat questions, not as a monitoring tool. Start with one team of 5-10 people and name an AI champion; run for 4 weeks and review the question list and feedback (GUIDE-008, GUIDE-009).

Humans in control: the employee runs every step.

Risks, regulation & compliance

  • EU AI Act: minimal risk; no Annex III category. Art. 4 literacy measures apply. Art. 50 transparency is satisfied by labelling the assistant as AI (GUIDE-014). Do not use logs to assess employees; that would move toward high-risk worker evaluation under Annex III (duties from 2 Dec 2027).
  • GDPR: prompts, screenshots and tickets can contain personal data; data minimisation, AVV, EU hosting where possible, short log retention.
  • Security and quality: secrets in screenshots, prompt injection via pasted documents, wrong click paths that damage data; use source links, a "try on a copy" rule and approved-tool lists (GUIDE-013, GUIDE-007).

Getting started

  1. Collect the 20 most frequent how-to questions from tickets and staff (half a day).
  2. Enable Copilot Chat or Gemini for a pilot team and agree on a one-page usage rule (GUIDE-007).
  3. Involve the works council and DPO before launch; sign the AVV (GUIDE-004, GUIDE-005).
  4. Run a 4-week pilot, track tickets, time to solve and feedback (GUIDE-009).
  5. If company-specific questions dominate, add retrieval over 10-30 curated work instructions (Path B or C).

Sources

  1. Microsoft Learn - Microsoft 365 Copilot overview - https://learn.microsoft.com/microsoft-365-copilot/microsoft-365-copilot-overview
  2. Microsoft Learn - Enterprise data protection in Microsoft 365 Copilot Chat - https://learn.microsoft.com/copilot/microsoft-365/enterprise-data-protection
  3. Microsoft - EU Data Boundary for the Microsoft Cloud - https://learn.microsoft.com/privacy/eudb/eu-data-boundary-learn
  4. Open WebUI documentation - https://docs.openwebui.com/
  5. Ollama - https://ollama.com/
  6. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj