Every company has its own language: "the Müller job", "KV" for cost estimate, article numbers where the third digit encodes the colour. People understand this; AI tools do not, and they mishear, misread or simply find nothing. A small, owned list of terms, abbreviations and synonyms, fed into search, transcription and assistants, is one of the cheapest ways to improve answer quality. It takes days, not months.
Problem & value
- Today: vocabulary lives in people's heads, old e-mails and scattered Excel sheets. A new colleague asks what "WE-Buchung" or "RMA" means; an assistant asks the same.
- Typical failures: transcripts garble product names; search finds nothing for "Kostenvoranschlag" because documents say "KV"; an assistant mixes up "Auftrag" (customer vs. production order).
- Value: better search hit rate, cleaner transcripts, fewer wrong assistant answers, faster onboarding. Qualitative expectation; measure it with a before/after test (see KPIs).
Variants
- Manufacturing/trade: part-number logic, variant codes, supplier short names, machine nicknames.
- Services: project and client code names, role titles.
How it works
- sources (price lists, ERP article master, handbooks, search logs, transcripts)
- candidate extraction (LLM proposes terms and abbreviations from documents)
- human review by department (meaning, synonyms, owner)
- terminology table (single source: Excel / SharePoint list / wiki)
- publish to consumers: search acronyms | synonym file | ASR custom words | assistant glossary
- feedback loop: failed searches and corrections add new terms
- LLM role: proposes "term, probable meaning, context" candidates and synonym groups; it does not decide meaning.
- Human-in-the-loop: a named owner per department approves each entry. Unreviewed AI-guessed meanings are the main risk, because a wrong definition then spreads into every consumer.
- Consumers: (1) Microsoft 365 Copilot Search / Microsoft Search acronyms and bookmarks (admins can add entries, stage them as draft, and import a CSV) [1]; (2) search engines with synonym files (Elasticsearch/OpenSearch synonym filters, Paperless-ngx tags and correspondents); (3) speech recognition with a vocabulary hint; (4) a glossary block in the system prompt of assistants (see PREP-013).
Data preparation
Which data: article master, price lists, handbooks, support tickets, search logs, transcripts and expert knowledge.
Typical gaps: the same abbreviation with different meanings per department, several spellings of a product name, customer-specific nicknames, outdated codes, and no owner.
Preparation steps: 1. Export article master from the ERP (PREP-005); sort by frequency. 2. Pull zero-hit searches and misrecognised transcript words; highest-value terms. 3. Let an LLM propose abbreviations and synonyms from 50-100 documents; export to Excel with the columns term, meaning, synonyms, department, owner, example sentence, status. 4. Hold one 60-90 minute review per department; resolve conflicting meanings by adding a department column rather than forcing one meaning. 5. Publish the approved list to the consumers above and name a maintainer (PREP-017).
Tools: Excel or SharePoint list as the single source, Microsoft 365 admin center (acronyms), Paperless-ngx or Nextcloud tags, n8n or Power Automate for sync.
Effort: a first list of 200-500 terms typically takes about 2-5 person-days spread over two weeks (own estimate, driven by meeting time of experts). Maintenance: roughly an hour a month. See GUIDE-003; link to PREP-010 (metadata), PREP-012 (KPI glossary) and PREP-014 (AI-ready authoring).
Implementation paths
Which path for whom
- SMB - A: the tools are already part of Microsoft 365, Google Workspace or the DMS; an Excel list plus built-in settings is enough, so buy nothing extra.
- Enterprise - A: commodity function; use the suite's acronym features or a wiki glossary.
- Developer - C: open-source search and ASR stacks give full control over synonym files and vocabulary biasing.
- Public - C: terminology often contains names of persons and cases; self-hosted or sovereign-cloud search and transcription fits procurement rules.
Path A - Buy (recommended for most SMBs)
- Microsoft 365: acronyms and bookmarks in Copilot Search / Microsoft Search admin settings, CSV import possible [1].
- DMS/Wiki: Paperless-ngx (open source, self-hosted), Nextcloud, Confluence, Notion, DocuWare, d.velop; store the glossary as a page and use tags for synonyms.
- Transcription: many tools offer a custom-vocabulary field; EU-hosted or local options include Nextcloud Assistant with local Whisper and Amberscript; verify hosting and AVV per vendor.
- Choose A when: you already use these suites and the list is below a few thousand terms.
Path B - Build on cloud APIs
- Speech: cloud ASR services offer phrase lists or custom vocabulary (e.g. Azure AI Speech phrase lists, Google Speech-to-Text adaptation) in EU regions.
- LLM glossary injection: pass the relevant glossary rows as context in each call, selected by embedding search; EU endpoints from Azure OpenAI (EU regions), Mistral or Aleph Alpha.
- Choose B when: you build your own assistant or search and need the list injected programmatically.
Path C - Self-host (open source)
- Transcription: Whisper (large-v3 or turbo variants, faster-whisper) with an initial prompt listing 5-10 key terms; community experience says the prompt mostly affects the first audio window and long term lists can hurt, so test on your own recordings [2]. Alternatives with hotword support exist.
- Search: OpenSearch/Elasticsearch synonym filters; Meilisearch and Typesense also support synonyms.
- Hardware: HW1 suffices for Whisper small/medium and small LLMs.
- Choose C when: transcripts or documents are sensitive (HR talks, customer calls) or the same platform serves several use cases.
Path D - Custom training
Only when vocabulary biasing fails: fine-tune Whisper on your own corrected recordings. Rarely justified for an SMB.
State of the art (2026)
- Assistants know common abbreviations but not private codes; injecting the glossary as context is the standard fix, not fine-tuning.
- Whisper-style ASR is strong on German, rare names stay error-prone; prompt hints plus glossary post-correction are common.
- Limits: homonyms need context; very long glossaries dilute prompts; lists age without an owner.
Shared resources & combinations
- Feeds PREP-013 (RAG knowledge base), PREP-014 (AI-ready authoring), PREP-010 (metadata) and PREP-012 (KPI glossary). One list, many consumers (search, transcription, assistants).
Effort, cost & KPIs
- Path A: 2-5 person-days for the first list, no extra licence cost.
- Path B/C: add 2-5 days for sync and prompt integration.
- KPIs: zero-hit searches before/after (same 50 queries), word error rate on 5 test recordings with/without vocabulary, share of entries reviewed in the last 12 months.
People & change
Involve early: the department leads (they own the meanings), long-serving staff and sales/service as sources, IT for publishing, and the data owners of article master data. The DPO is only needed if the list contains personal names (customer contacts, employees); keep such entries out.
Works council: normally no; the list contains no behaviour or performance data. Be careful if search logs or transcripts are analysed per person to find terms: that can fall under §87(1) no. 6 BetrVG, so use aggregated, anonymised queries (GUIDE-004). §90(1) no. 3 asks for timely information on AI-supported work processes in any case.
AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; a 30-minute session on how assistants use the glossary and how to report a missing term is enough (GUIDE-006).
Pilot: frame it as "teaching the tools our language"; start with one department; name an AI champion as maintainer (GUIDE-008, GUIDE-009).
Humans in control: every definition is approved by a named owner; AI only proposes.
Risks, regulation & compliance
- EU AI Act: minimal risk; only AI literacy (Art. 4) applies, as measures that support staff competence.
- GDPR: a pure terminology list is not personal data; avoid names of customers' contact persons and employees, or treat the list as personal data and sign an AVV (Art. 28 GDPR) with any cloud vendor that processes it. Transcripts used to find terms are personal data; use EU hosting or local Whisper.
- Confidentiality: code names for unreleased products or deals are trade secrets; check them against the AI policy before sending to external tools (GUIDE-007).
Getting started
- Pick one department and collect the 100 most-used abbreviations and product names from the ERP and handbooks (half a day).
- Take 20 recent failed searches or misheard words as the test set.
- Create the Excel/SharePoint table with owner and review date; let the department approve it in one meeting.
- Publish to one consumer first (Copilot Search acronyms or the Whisper prompt) and re-run the test set.
- Add a "report missing term" mailbox or form; review monthly (PREP-017).
Sources
- Microsoft - Microsoft 365 Copilot Search admin experience (acronyms, bookmarks) - https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-search-admin-experience
- Hugging Face forum - Adding custom vocabularies on Whisper - https://discuss.huggingface.co/t/adding-custom-vocabularies-on-whisper/29311
- OpenAI Whisper - GitHub repository - https://github.com/openai/whisper
- Paperless-ngx documentation - https://docs.paperless-ngx.com/
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj