Catalog / Everyday Automation & AI Micro-Tools / AUTO-002

Email thread catch-up & in-client reply suggestions (Outlook/Gmail)

Summarise long email threads inside the mail client, show what is still open or was agreed, and get editable reply suggestions - for every office worker who drowns in CC chains.

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

Parts needed to run it yourselfAgents & automationLanguage model

Anyone who returns from a meeting or holiday to a 40-message thread knows the problem: who decided what, and what is still open? AI built into the mail client reads the thread on request, produces a short summary with open points, and suggests a reply that the user edits and sends. It is interactive and user-triggered: nothing is sent automatically. For most SMBs this is a licence and a policy, not a project.

Problem & value

  • Today: staff scroll through quoted replies, forward threads with "FYI, see below", or ask colleagues to explain. Colleagues added late to a thread have no overview; replies take time because tone and facts must be reconstructed.
  • AI value: a summary in seconds, a list of open questions and agreements, and a first-draft reply. The value is saved reading time and fewer misunderstandings, not automation of correspondence.
  • Benefit figures: independent, sourced time savings for this feature are scarce; measure them in a pilot.

Variants

  • Microsoft 365: "Summary by Copilot" at the top of a thread in Outlook (new Outlook, classic Outlook, web), plus Copilot drafting and reply coaching [1][2].
  • Google Workspace: Gemini in Gmail summarises threads (the side panel and a "Summarize this email" button) and suggests drafts [3].
  • Open-source mail: Nextcloud Mail with the Nextcloud Assistant can summarise threads, with the language model running on your own server [5].
  • Public sector: same function, but sovereignty and procurement rules push towards self-hosted or sovereign-cloud mail.

How it works

user opens thread -> clicks Summarize / asks a question -> client sends the thread to the provider's LLM -> summary with references -> user asks for a reply -> draft -> user edits and sends

The language model only sees what the user can already see (the permission model of the mailbox). Microsoft shows numbered citations that jump to the source mail in the thread, which makes checking easy [1]. Human-in-the-loop points: the user decides whether to trust the summary, and every reply is edited and sent by the user.

Data preparation

Almost none. The data is the user's own mailbox. What does need attention: - Access rights: Copilot-style features respect existing permissions, so oversharing in shared mailboxes and SharePoint becomes visible. Clean up shared mailboxes and delegations before rollout (GUIDE-003). - Mailbox basics: Microsoft states that Copilot in Outlook works on the primary mailbox and cannot access encrypted emails [2]. Plan for that limitation with S/MIME-protected mail. - Sensitive folders: decide which mail (HR cases, legal, health data) users should not feed to AI at all, and write it into the AI policy (GUIDE-007). - Effort: 0.5-2 days for IT to check licences, settings and shared mailboxes; no PREP pipeline needed.

Implementation paths

Which path for whom

  • smb: A - the feature is part of the suite you already use, no integration, standard process.
  • enterprise: A - a commodity function; a separate build rarely pays off. Use B/C only if a central AI platform already exists.
  • public: C - sovereignty and procurement rules; Nextcloud Mail/Assistant or comparable self-hosted mail with a local model.

Path A - Buy (off-the-shelf)

  • Microsoft 365 Copilot (Business for up to 300 users): list price reported at about USD 21 per user per month on annual commitment, on top of a Business Basic/Standard/Premium licence; promotional prices changed during 2026, so check Microsoft's current price list before budgeting [4]. Microsoft also announced that Copilot chat can summarise emails for users without a Copilot licence, with its own privacy terms - check your tenant's rollout [2].
  • Google Workspace with Gemini: included in business plans according to current Google offers; verify the plan and the data-protection terms in the admin console. Do not use consumer or Labs/early-access programmes for business mail: Google's page states that data from the Labs programme may be reviewed by humans [3].
  • EU hosting: Microsoft processes Copilot data for EU tenants within the EU Data Boundary according to secondary sources, which I could not confirm against Microsoft documentation; verify in Microsoft Learn and your admin centre. Sign the vendor data-processing terms (AVV, Art. 28 GDPR).
  • EU-hosted alternatives: Nextcloud Mail with Assistant on a German host (e.g. IONOS, Hetzner, own server) [5]; Open-Xchange and similar suites offer AI add-ons - verify the model hosting per product.

Path B - Build on cloud APIs

Rarely sensible: a small add-in or n8n flow that reads a thread via Microsoft Graph or the Gmail API and calls an LLM (EU-region Azure OpenAI, Mistral, or a German provider). It only pays off if you need a custom summary format (for example, ticket or CRM fields) or want to avoid per-seat licences for a few users. Effort: 3-10 days per add-in including permissions review. Check AGENT-002 for the workflow pattern.

Path C - Self-host (open source)

Nextcloud Mail plus the Nextcloud Assistant can summarise threads; the Assistant can use a local model through an OpenAI-compatible endpoint (integration_openai), for example Ollama [5][6]. Test the Mail thread summary with your version; community reports show it does not always work out of the box [5]. Hardware: HW2 (one 16-32 GB GPU) for a 7-14B quantised model serving a team of 20-50 users; HW1 (CPU) works but is slow for long threads. Ops effort: about 2-5 days setup, then updates and monitoring. Choose C for public bodies, very sensitive mail, or when the same Nextcloud/LLM server also serves AUTO-004, AUTO-005 or DOC-014.

Path D - Custom training

Not needed. Tone and style adaptation works with prompts or templates (see AUTO-023).

State of the art (2026)

  • Thread summarisation and reply drafting are built into the major suites and are largely commodity. Differences are in citations, language coverage, and admin controls.
  • Limits and failure modes: summaries may omit a late correction, mix up who promised what, mis-handle long quoted chains, forwarded attachments, or irony. Reply suggestions can sound generic or promise things the user cannot commit to. Always read the thread's last two messages yourself before sending.
  • Prompt injection: an external email can contain hidden instructions aimed at the assistant. Keep the feature user-triggered and treat summaries of external mail with care (GUIDE-013).

Shared resources & combinations

  • Same licence serves AUTO-003 (meeting briefs), AUTO-004 (chat catch-up), AUTO-028 (absence handover) and AUTO-024 (daily planning).
  • Not the same as AGENT-001 (autonomous inbox triage and auto-responses) or LANG-002 (drafting new correspondence from scratch).
  • Self-hosted: one Ollama/Nextcloud server for several AUTO use cases (Baukasten effect).

Effort, cost & KPIs

  • Path A: 1-3 days for licence assignment, admin settings and policy; costs are per-seat licences (see above). A pilot of 5-10 people for 4 weeks.
  • Path B: 3-10 days per add-in plus API costs.
  • Path C: 2-5 days setup plus a GPU server; no per-seat costs.
  • KPIs: weekly active users of the feature, self-reported minutes saved per thread (survey), share of summaries judged correct in a spot check of 20 threads, number of reply drafts sent unedited (should stay low), policy incidents.

People & change

  • Involve early: management (budget, policy), IT and IT security (settings, permissions), the DPO (AVV, DPIA screening), the works council if one exists, and end users as pilot group (GUIDE-008, GUIDE-009).
  • Works council: yes. Microsoft 365 Copilot is named as typical example of software objectively suited to monitor behaviour or performance under §87(1) no. 6 BetrVG, because usage is logged; §90(1) no. 3 requires timely information about AI-supported work processes, and §80(3) lets the council engage an AI expert. A short works agreement (no individual performance evaluation, defined logging) usually settles it. Public bodies involve the Personalrat (GUIDE-004).
  • Data protection: a DPIA is usually not needed for a user-triggered summary tool, but the DPO should check the screening, the AVV (Art. 28 GDPR) and retention of prompts (GUIDE-005, GUIDE-010).
  • AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy. Offer a 60-minute session: what the tool does, how to verify, which mail must never be fed to AI (GUIDE-006).
  • Concerns: "my mail is read" (explain permissions and logging), fear of control (works agreement). Humans stay in control: nothing is sent without the user.

Risks, regulation & compliance

  • EU AI Act: minimal risk; no Annex III category. Art. 4 literacy support applies. If the drafted text is sent to third parties, the sender remains responsible for content (GUIDE-014).
  • GDPR: mail contains personal data of customers, applicants and colleagues; check legal basis (legitimate interest), AVV, data location and retention (GUIDE-005). Exclude special categories unless the DPO has cleared it.
  • Confidentiality: professional secrecy (lawyers, tax advisers, doctors) needs special review (§203 StGB), usually Path C.
  • Mitigations: admin disable for sensitive mailboxes, labels/sensitivity policies, user training, spot checks.

Getting started

  1. Write a one-page AI policy for mail (what is allowed, what never goes in) with the DPO (GUIDE-007).
  2. Check shared-mailbox permissions and switch the feature on for a pilot group of 5-10 users.
  3. Inform the works council and DPO before the pilot; agree on logging rules (GUIDE-004).
  4. Run 4 weeks, collect minutes saved and error examples, then decide go/no-go (GUIDE-009).
  5. If mail is highly sensitive, test Nextcloud Mail with a local model as an alternative.

Sources

  1. Microsoft Support - Summarize an email thread with Copilot in Outlook: https://support.microsoft.com/office/summarize-an-email-thread-with-copilot-a79873f2-396b-46dc-b852-7fe5947ab640 (accessed 2026-10)
  2. Microsoft Support - Frequently asked questions about Copilot in Outlook: https://support.microsoft.com/office/07420c70-099e-4552-8522-7d426712917b ; Message Center MC1124564: https://mc.merill.net/message/MC1124564 (accessed 2026-10)
  3. Google Workspace Updates - Gemini summary cards: https://workspaceupdates.googleblog.com/2025/05/gemini-summary-cards-gmail-app.html (accessed 2026-10)
  4. Microsoft 365 Blog - Microsoft 365 Copilot Business: https://www.microsoft.com/en-us/microsoft-365/blog/2025/12/02/microsoft-365-copilot-business-the-future-of-work-for-small-businesses/ ; Office Watch: https://office-watch.com/2025/microsoft-365-copilot-business-a-cheaper-ai-add-on-for-small-organizations/ (accessed 2026-10)
  5. Nextcloud - Improving productivity with a private AI Assistant: https://nextcloud.com/?p=232455 ; Nextcloud forum - Assistant in Mail: https://help.nextcloud.com/t/how-to-activate-the-nextcloud-assistant-at-mail/172314 (accessed 2026-10)
  6. Nextcloud admin manual - AI / Summary Bot (task processing providers): https://docs.nextcloud.com/server/stable/admin_manual/ai/app_summary_bot.html (accessed 2026-10)