This document is password protected.
This document is password protected.
A four-phase roadmap for EA Partner — from a working proof-of-concept to productive AI integration across the entire work environment.
The proof-of-concept is technically validated. We've shown that a modern AI client (Claude) can be connected to a Microsoft 365 tenant cleanly, securely, and with proper permission boundaries — including the niche workloads EA Partner relies on daily that Microsoft doesn't cover out of the box.
PoC result: Six Microsoft MCP servers (Mail, Calendar, Word, SharePoint, OneDrive, Identity) plus four custom MCPs we built (Teams chats, OneNote, Shared Mailbox, Group Mails) all run productively against our tenant. OAuth/PKCE, audit log, dry-run on write operations, prompt-injection protection.
The PoC ran against our tenant. It says: "This works technically." It does not say: "This works in EA Partner's reality." That step requires moving into your environment — with your mail, your documents, your workflows.
Our proposal: we split the engagement into four clearly delineated phases. Each phase delivers standalone value, has a fixed price range, and ends with a decision point. You can stop after any phase, continue, or reorder them.
The build-up is iterative: each phase builds on the last, each delivers real value on its own. Phase 0 (PoC) is complete and was free for EA Partner — the following phases are the productive implementation.
All MCPs from the PoC, freshly provisioned in your M365 tenant. Pilot user Thomas Etienne.
CHF 5,000 – 7,000Make the SMB archive searchable. The AI accesses historical projects contextually.
CHF 2,000 – 4,000«Cognitive Workspace» — dedicated web app. Business and private life in one interface with memory.
CHF 15,000 – 40,000+Your own infrastructure, maximum sovereignty. On-site setup — a joint pioneer step.
CHF 6,000 – 8,000Goal: within three weeks you're working in your daily Microsoft setup with an AI assistant that knows your real mail, calendar, OneNote and document context — properly authenticated, fully auditable, with no shadow IT.
Dedicated app in the EA Partner tenant. OAuth 2.0 with PKCE, single-tenant, public client. Clean permission structure, documented.
Mail, Calendar, Word, SharePoint, OneDrive, Identity. Provisioned, tested, usable in Claude Desktop or VS Code.
OneNote, Teams chats, Shared Mailbox, Group Mails — the workloads Microsoft doesn't cover itself. Code from the PoC, freshly configured.
Audit log on every write, dry-run by default on send/reply, prompt-injection markers on external content.
1:1 handover with you, cheat sheet for typical use cases, written setup documentation, troubleshooting guide.
Half-day session: we walk through your real workflows and identify where the assistant will create value first.
Phase 1 price: CHF 5,000 – 7,000. The range depends on how much custom-MCP adaptation your specific tenant configuration requires (e.g., group mail structures, shared-mailbox permissions). Fixed price after the setup workshop.
EA Partner sits on an SMB archive grown over years on a Synology — plans, correspondence, tenders, renovation files. Today the archive only helps when someone knows where to look. Phase 2 turns it into an active knowledge base, queryable in natural language.
Concretely: we get access to the Synology and build a connection that lets the AI search the SMB archive and return contextually relevant content. How that connection works in detail — direct SMB mount, Synology API, indexing via Active Backup, or a lightweight crawler — we decide after inspecting the actual environment.
The challenge: "grown over years" means a lot of data — and none of it is currently vectorised or prepared for an AI. Running everything through an embedding model in one go would be expensive and slow. So we approach it in stages.
Phase 2 opens the archive for the AI at the navigation level: filenames, paths, project numbers, dates — everything that can be unlocked without an embedding run. That resolves 60–70% of typical "where was that again…?" queries instantly.
The AI queries the Synology live through its own search. No vectorisation required. Finds what the Synology itself finds — filenames, full text on indexed formats.
We index paths, project numbers, modification dates. Quick filtering: "show me everything for Project X since 2024", even if the Synology can't do that itself.
What Phase 2 doesn't deliver — and why that's OK: True semantic understanding ("find correspondence with a similar tone to this one") requires vectorisation. That's a separate step with its own costs — we build it once we know whether you actually need it. Often it turns out after Phase 2 that navigation alone is enough.
Phase 2 price: CHF 2,000 – 4,000. Fixed price after the audit. Optional extension Phase 2b (selective vectorisation for semantic depth): approx. CHF 3,000 – 5,000 + ongoing embedding costs. We build 2b only once you concretely see the need after Phase 2 — not on speculation.
Phases 1 and 2 connect the AI to your existing Microsoft tools. Phase 3 builds something new: a dedicated web app that brings business and private life into one interface, maintains persistent memory and works proactively — rather than just responding to commands.
Not a second chat client. A dedicated interface with three pillars — connectors, memory, agents — built on an architecture we already have experience with (Next.js + our own "Brain" daemon, encrypted storage of sensitive tokens, vector search over personal knowledge).
All MCPs from Phase 1 plus private sources: personal mail account, personal calendar, notes & to-dos.
Persistent conversations, factual memory ("my favourite café in Lucerne", "Client X prefers PDF plans"), pgvector search across it all.
Proactive: morning briefing at 7 a.m., overnight mail triage, meeting prep with archive context, weekly summary on Friday.
Phase 3 price: CHF 15,000 – 40,000+ for exactly one concept. Lower bound = the concept you choose as a solid MVP, production-ready. Upper bound = the same concept with fine UX polish, multiple agent workflows and intensive iteration time. Precise scoping at the start of the phase — the range is depth, not breadth.
Three sketches — same functional core, three characters. We build exactly one of them, you choose the direction at the start of Phase 3. The other two are food for thought so you can make an informed decision.
How to read this: the three concepts are not add-ons. You get exactly one fully developed product — whichever you choose. The price range doesn't change because of it.
Three concepts · same functional core · final design language is defined with you in Phase 3
Before booking Phase 4, it's worth looking at the data-flow reality. In phases 1–3 we send requests by default to a frontier provider (Anthropic). If that feels too open, there are two escalation tiers.
Anthropic API. Content transported encrypted, processed under provider policy. Industry standard, cheapest option. Default in Phase 1–3.
Frontier provider with a contractual "no training on your data" commitment (Anthropic Enterprise / AWS Bedrock / Azure OpenAI). Higher token prices, small setup surcharge.
Local infrastructure, your hardware. Local open-source LLMs (still pioneer work) or hybrid mix. Maximum effort, maximum sovereignty.
Transparency note: at Tier 1, the LLM provider has technical access to the content of your requests — contractually safeguarded, but worth knowing. Token costs run directly through your API account (typically CHF 50–200/month for one active user, depending on usage) — they're not a subscription with us, but pay-per-use with Anthropic.
Phase 4 is also our first productive on-prem case — which is precisely why we run it as an R&D partnership at a reduced pioneer rate. Concretely: ~2 weeks on-site at your premises, joint setup on your hardware (e.g. NVIDIA DGX Spark or equivalent), evaluation of the right local model, installation in week 3. Local open-source LLMs are currently still a Goldilocks problem (small = weak, large = slow) — we may need to build a custom harness around it. The reduced price reflects this pioneer role: you get an on-prem AI earlier and cheaper than at our standard day rate, we gain our first validated on-prem case.
Phase 4 price: CHF 6,000 – 8,000. Includes setup, model evaluation, on-site installation. Fixed price after hardware inspection. Hardware procurement (Spark / workstation / GPU) not included — recommendation and procurement on request.
| Plan | Inclusion | Price |
|---|---|---|
| Basic Retainer | Updates, small adjustments, monitoring, max. 8h/month | CHF 2,000 / month |
| Active Retainer | Plus new agents, connectors, max. 12h/month, priority support | CHF 3,000 / month |
Both plans cancellable monthly. Unused hours carry over within the current quarter; they expire afterwards.
We work with a fixed cadence from "go". Phases 1 and 2 are small enough to wrap in under two months. Phase 3 is the substance — three to four months, with clear intermediate results every two weeks.
| Week | Phase | Milestone |
|---|---|---|
| W 1 | Phase 1 | Setup workshop, Entra app registered, first OAuth flows |
| W 2–3 | Phase 1 | All MCPs deployed, use-case workshop, handover |
| W 4 | Dogfooding | You work live with Phase 1; we're on standby for questions and quick fixes |
| W 5–6 | Phase 2 | Archive audit, architecture recommendation, custom MCP for the archive |
| W 7 | Decision gate | Which concept becomes Phase 3? Scope sharpening based on Phase 1+2 experience |
| W 8–10 | Phase 3 | Foundation: web app setup, Brain daemon, first connectors |
| W 11–13 | Phase 3 | Memory layer, vector search, first agent (morning briefing) |
| W 14–16 | Phase 3 | More agents, UI polish, fine-tuning the chosen concept |
| W 17–18 | Phase 3 | Hardening, handover, training |
Total: ~4 months for Phase 1–3. The order is not set in stone: you can e.g. do Phase 1 + 2, then jump straight to Phase 4 (On-Prem) and tackle Phase 3 (Personal Assistant) later. On request, Phase 1 and 2 can even run in parallel — that shortens the total duration by 2–3 weeks.
All prices in CHF, excluding VAT. Each phase is a separate engagement with its own sign-off — you never commit beyond the next phase. Microsoft licences and cloud costs (Anthropic API, optional vector DB hosting) are not included and run directly through your accounts.
| Phase | Content | Duration | Price |
|---|---|---|---|
| Phase 0 | Proof-of-Concept | completed | free |
| Phase 1 | Tenant integration & personal workspace | ~3 weeks | CHF 5,000 – 7,000 |
| Phase 2 | SMB archive connection & search | ~2 weeks | CHF 2,000 – 4,000 |
| Phase 3 | Personal Assistant — Cognitive Workspace | ~10–12 weeks | CHF 15,000 – 40,000+ |
| Trusted Cloud | Upgrade to no-train provider (Anthropic Enterprise / Bedrock / Azure) | one-time | CHF 1,500 – 3,000 |
| Phase 4 | On-Prem setup & connection (optional, R&D partnership) | ~3 weeks | CHF 6,000 – 8,000 |
| Retainer | Ongoing operations & development (optional) | monthly | CHF 2,000 – 3,000 / month |
Total Phase 1–3 (realistic range): CHF 22,000 – 51,000+. Phase 1 and 2 typically land in the lower to middle third of their range. Phase 3 is depth and is determined by your concept decision in week 7 — the upper bound is realistic with multiple agent workflows and full UX polish.
Now begins the part where ideas turn into real results.
Schedule the setup workshop