Self-built custom assistants for recurring tasks (custom GPTs, Copilot agents, Gems)

Many office tasks repeat with the same rules: answer the standard supplier question, check an offer against the purchasing checklist, turn site notes into a service report. Today most people re-type the same prompt each time or paste the same files into a chat. A custom assistant stores the instructions and a few reference files once, so the task becomes one click and a colleague gets the same quality. Typical set-up time is an afternoon; the benefit is consistency and saved typing, not a new process.

Problem & value

• Today: employees keep personal prompt notes, copy rules from old mails and paste the same price list or checklist into a chat. Results differ by person, and knowledge stays in one head.
• AI value: a named assistant with fixed instructions, an output format and a small set of files. It handles first drafts and checks; the human decides.
• Benefit figures: no reliable independent figure exists for micro-assistants specifically. Measure your own baseline (minutes per task before and after) in the pilot instead of relying on vendor claims.

Variants

• Per-person assistant: private helper, no sharing, lowest risk.
• Team assistant: shared inside a department with one owner who maintains instructions and files.

How it works

1. task owner writes instructions + output format
2. uploads 2-10 reference files (checklist, price list, template)
3. platform indexes the files (retrieval) and stores the assistant
4. user pastes input (offer, mail, notes)
5. model answers following instructions and files
6. human reviews, edits, sends

• Key components: a general LLM, a system prompt, and simple file retrieval. No code is needed in the commercial builders.
• Human in the loop: the assistant only drafts. Sending, approving and booking stay with the person.

Data preparation

Which data: a short written task description, 2-10 clean reference files and 5-10 example inputs with the output you consider correct.

Where it lives: SharePoint or OneDrive, Google Drive, a Nextcloud share, or simply the employee's own documents.

Typical gaps: outdated price lists, conflicting versions of a checklist, scanned PDFs without text, files that contain personal or confidential data that should not be uploaded, and unclear ownership of the content.

Preparation steps: 1. Name one owner per assistant who keeps the files current. 2. Convert scans to text PDFs or Markdown (Paperless-ngx OCR or the scanner software) and remove old versions (see PREP use cases on document cleanup). 3. Split the source: keep only what the task needs; do not upload whole folders. 4. Strip or anonymise personal data in examples. 5. Write 5-10 test inputs with expected results to re-check after every change.

Effort: half a day to two days per assistant, mostly cleaning files. See GUIDE-003.

Implementation paths

Which path for whom

• SMB - Path A: the builders inside Microsoft 365, Google Workspace or a chat product already exist and the process is standard. Seats are usually already licensed.
• Enterprise - Path B: many assistants across departments need governance, logging and integration, which favours Copilot Studio or a shared internal platform on cloud APIs.
• Developer - Path C: Open WebUI with Ollama gives custom models and knowledge on own hardware without per-seat fees.
• Private - Path A: consumer or low-cost plans of the common chat tools are enough for personal helpers.
• Public - Path C: procurement rules and sovereignty favour a self-operated or German-cloud platform; certified products only with a proper AVV.

Path A - Buy (recommended for most SMBs)

• Microsoft 365 Copilot: Agent Builder and Copilot Studio create agents for Teams, SharePoint and Copilot. Microsoft's licensing guide (February 2026 edition; newer editions exist, e.g. July 2026) says employee-facing use is covered when users hold a Microsoft 365 Copilot licence; other use is billed via Copilot Credits [1]. Check the current terms before rollout.
• Google Workspace: Gems in the Gemini app. Google offers EU data-region controls for the Gemini app on certain editions [2]; verify that your edition and Gems are covered.
• Mistral (EU provider, Paris): the former Le Chat, renamed Vibe on 28 May 2026, offers agents on business plans; reports on which tier includes them differ, so confirm with the vendor [3].
• ChatGPT Business and Claude Team/Enterprise: custom GPTs and Projects; check EU data residency options and the AVV before use.
• Choose A when: the data is not highly sensitive, the tool is already licensed, and sharing within one team is enough.

Path B - Build on cloud APIs

• Platform: Copilot Studio, or n8n / Power Automate calling an EU-region model API (Azure OpenAI in EU regions, Mistral API, Google Vertex AI EU).
• Effort: 2-6 weeks for a governed platform with the first five assistants.

Path C - Self-host (open source)

• Stack: Open WebUI (custom models with attached knowledge) plus Ollama or vLLM; a 7-14B open-weight model such as Mistral or Qwen class is enough for checklists and drafting. n8n for triggers, Nextcloud for files.
• Hardware: HW1 (a modern CPU server or laptop NPU) runs small models for a few users; HW2 gives acceptable speed for a team.
• Choose C when: HR, trade-secret or regulated customer data is involved, or one platform serves many use cases (Baukasten effect with AUTO-023, KNOW-002).

Path D - Custom training

Not needed. Instructions and files cover this use case; fine-tuning only pays off at very high volume.

State of the art (2026)

• All major suites now ship a no-code assistant builder, and sharing inside the tenant is standard. Governance (who may publish, which sources) is the main differentiator, and Microsoft requires a separate Agent 365 licence for AI agent security capabilities for Copilot Studio agents from 1 July 2026 [4].
• File retrieval works well for short, well-structured documents and badly for long tables, scans and contradictory versions.
• Failure modes: invented clauses or prices, ignoring instructions on long inputs, outdated files, and prompt injection from pasted external mails. Always keep the human check.

Shared resources & combinations

• Shares the LLM and retrieval building blocks with KNOW-002 and the prompt content with AUTO-023 (prompt libraries).
• Combine with AUTO-018 (script help) and AUTO-027 (how-to tutoring); the same Open WebUI server can host all of them.

Effort, cost & KPIs

• Path A: 2-8 hours per assistant; cost is mostly existing seats, possibly credits.
• Path B: 2-6 weeks for a platform; API cost per use is usually small.
• Path C: 3-7 days to set up Open WebUI and Ollama; ongoing admin of a few hours per month.
• KPIs: minutes saved per task (measured), share of drafts accepted with small edits, number of active assistants and users, number of outdated assistants found in the quarterly review.

People & change

Involve early: management for the AI policy, department leads as owners, IT for accounts, sharing rules and the list of approved tools. The DPO checks which data may be uploaded and the AVV (Art. 28 GDPR); a DPIA (Art. 35) is normally not needed unless people are systematically evaluated (GUIDE-005, GUIDE-007).

Works council: depends. Assistants built on Microsoft 365 Copilot or similar tools log user activity, and the BAG holds that objective suitability for monitoring is enough for co-determination under §87(1) no. 6 BetrVG. §90(1) no. 3 requires timely information when work processes are planned with AI; the council may engage an AI expert under §80(3). Public bodies involve the Personalrat. Only relevant where a works council exists. A short works agreement on approved tools settles most questions (GUIDE-004).

AI literacy: under Art. 4 AI Act, employers must take measures that support the development of staff AI literacy. A two-hour workshop on writing instructions, testing and the do-not-upload list is enough here (GUIDE-006).

Communication and pilot: start with 3-5 curious employees as AI champions, one assistant each, and a shared catalogue so that good ones spread (GUIDE-008, GUIDE-009). Typical concerns: fear of surveillance (be transparent about logging), shadow AI (offer an approved tool instead of banning), quality worries (human review stays mandatory).

Risks, regulation & compliance

• EU AI Act: minimal risk for drafting and checking tasks. Do not build assistants that rank or select applicants or assess employees; that would move toward Annex III high-risk duties (apply from 2 Dec 2027). Art. 50 transparency applies if the assistant talks directly to external persons (GUIDE-014).
• GDPR: personal data in uploaded files or inputs needs a legal basis, an AVV and EU-hosting preference; avoid uploading customer lists.
• Confidentiality: trade secrets in public consumer tools; use business plans with no-training terms.
• Security: pasted external mail may contain injected instructions (GUIDE-013).

Getting started

1. Pick one repeating task with clear rules, for example the offer-versus-checklist check.
2. Collect 5-10 real inputs and the correct results; clean 2-3 reference files.
3. Build the assistant in the tool you already license; write the output format explicitly.
4. Test with the example inputs, then run a two-week pilot with three colleagues.
5. Register it in a shared catalogue with owner, files and review date; decide with the DPO and works council (if any) which tools are approved.

Sources

1. Microsoft - Copilot Studio Licensing Guide (February 2026) - https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/bade/documents/products-and-services/en-us/microsoft-365/1084694-Microsoft-Copilot-Studio-Licensing-Guide-February-2026-PUB.pdf
2. Google Workspace Updates - Data regions support for the Gemini app (June 2026) - https://workspaceupdates.googleblog.com/2026/06/gemini-app-data-regions-support.html
3. Mistral AI - Vibe (formerly Le Chat) - https://mistral.ai/products/vibe/
4. Microsoft - Defender XDR: transition of agent security to Agent 365 - https://learn.microsoft.com/en-us/defender-xdr/security-for-ai/transition-agent-security-to-agent-365
5. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj