iConference AG
AI that lands

AI meets SharePoint: a pain to set up, a win to run.

We rolled out a company-wide AI layer at an SME: an AI assistant (Claude Opus 4.8, as of 06/2026), connected to Microsoft Graph via MCP, covering SharePoint, email and Teams. The short verdict first: setting SharePoint up was tedious. Running it is not. And on balance, the rollout clearly pays off.

What was tedious to set up

The build was a slog. SharePoint is a chore to handle: a sluggish interface and a logic of sites, libraries and permissions that doesn't explain itself. What saved the build was, of all things, the AI assistant itself — it helped structure, permission and tidy up, instead of leaving us alone with the Microsoft docs.

Why daily use convinces

In daily use the picture flips. That very Microsoft 365 storage works as a lightweight RAG: staff ask in plain language, the assistant fetches the right document and cites the source as it goes — no vector database, no pipeline of our own. The same goes for email and Teams. The knowledge already in the house becomes queryable.

Flow: employees ask a question in plain language, the AI assistant (Claude Opus) fetches the answer from SharePoint, email/Outlook and Teams via Microsoft Graph and MCP — and replies with the source. Lightweight RAG, no vector database.
The assistant fetches the answer from your existing M365 data — and cites the source.

When the effort pays off

Who it pays off for: anyone already running Microsoft 365 gets a viable AI layer on their own data with manageable effort — no new system, no data migration. The expensive route, moving data into a separate vector database first, isn't needed to get started.

To be honest: getting in takes patience, and SharePoint won't become pretty for it. But the hurdle sits in the setup, not in daily use — and it only comes up once.

Setup is the price. Running it is the payoff.

We guide Swiss SMEs along this path — human, competent, holistic.

Read next

29 July 2026

Private AI accounts at work: ban first, then introduce properly

Customer data in a private AI account: who is liable under the revised Swiss data protection act, how thin the case law still is, and the two steps an SME takes now.

Read →
22 July 2026

The most insidious trap in AI answers: they sound utterly convincing but are factually wrong

We tested four AI models on a fictitious contract file. The most dangerous result was not the error, but how convincingly the weakest model went wrong.

Read →
22 July 2026

Method: four AI models on an M&A contract file

Method and findings for the model test: four AI models against a fictitious, anonymised contract file, measured deterministically against a solution key.

Read →