SharePoint AI search in production: what we deployed at a construction company, what it cost in days, and what surprised us about adoption.

When SharePoint outgrows everyone who knows it

XTM Bygg is a construction company with a long track record of projects across Sweden and Norway. More than a decade of work in three countries has grown their Microsoft 365 environment to 865 SharePoint sites and over 400,000 documents and records – projects, contracts, certificates, invoices, HR forms.

There was exactly one way to find anything specific: know which site and folder it lives in. In practice that meant asking the colleague who knows. And when that colleague leaves, the knowledge leaves with them.

What we deployed

The SharePoint front page now has a search box you talk to the way you would talk to a colleague – in Latvian, English, Swedish or Norwegian – all four languages XTM works in:

“Where is the latest version of the certificate form?”
“Where is tool number 1858?”
“How many active projects do we have right now?”

The answer arrives in seconds, with links to the documents it came from. Questions about registers (projects, vehicles, tools, invoices) are answered by the server from live SharePoint lists, not by a language model guessing.

Three things without which no company should let a tool like this in:

Everyone sees only what SharePoint already allows them to see. Every candidate document is checked against the requesting user’s own permissions before it appears; when in doubt, it is dropped. Nothing new gets opened to anyone.
The system never acts on its own. It drafts an email reply or a meeting invite – a human edits and confirms.
The index refreshes itself – a new document becomes findable without anyone lifting a finger.

Numbers

Deployment, from access to front page	~8 working days (next client: ~3)
Indexed	865 sites, 400,000+ documents
Users	every employee with an M365 licence
Infrastructure	EU data centre of the client’s choice (here: Sweden)
Extra licences	none – runs on the existing M365 setup

The honest part: trust has to be earned

We expected the search box to be flooded with questions from day one. The opposite happened: people use it cautiously, like something foreign. The first queries are careful probes, not a working routine.

That is not a failure of the tool – it is a finding about AI adoption in general. A tool can work flawlessly and still meet employees who do not trust it yet, and management enthusiasm does not skip this stage. Trust is built by three things: ready-made example questions right in the box, answers that always carry a source link (verify it yourself), and an honest “I could not find this” instead of invented text. Every correctly answered question earns one user.

We measure and manage this stage rather than waiting for it to pass.

Who is behind it

Answergrove is a product of SIA KSJ – private AI search and work automation inside the client’s own Microsoft 365 environment. It grew out of five years of building M365 document flows and invoice approval chains for a construction group operating in three countries.

For the next client the first step is an audit with a roadmap: what your SharePoint environment can answer, what it costs and what it saves. Pricing: ksj.lv/en/answergrove-copilot-alternative

P.S.

XTM is hiring in Latvia for its Sweden and Norway projects – open roles: xtmbygg.com/vacancies