Copilot Studio at a factory: searching across 4,500 technical documents

An industrial components manufacturer with one production site in Poland had been accumulating technical documentation for more than 15 years. Around 4,500 files in total: ISO 9001 quality procedures, station-level safety instructions, product technology cards, material specifications, internal audits, CAPA reports. Scattered across SharePoint (partly indexed, partly not), the network drives, a NAS PDF archive and a few folders left behind by former quality engineers.

Production engineers had one recurring complaint: you simply can't find the current version of a procedure. A question like „what's the surface tolerance for product X, class Y, customer Z?” often ended with a call to a colleague on another shift — or even to a retired ex-employee. Time studies showed the average engineer was losing 18 minutes per procedure search, with 8 such searches a day — over 2.5 hours a day lost.

Two prior attempts at a manual cleanup had both stalled at 30%. Every department had its own naming logic, and trying to enforce a single standard always ended in a quality-department veto. Leadership needed a solution that worked ON top of the existing chaos, not one that required cleaning it up as a prerequisite.

Weeks 1–2: content audit and classification by business impact. Out of 4,500 files we identified 1,200 high-impact documents (current procedures, station-level instructions, customer specs). The remaining 3,300 went into a second indexing wave with lower ranking priority. The crucial guardrail: no changes to folder structure — documents stay where they are.

Build took 4 weeks. Azure AI Search as the indexing layer: connectors to SharePoint, NAS and the site's network drives; semantic ranker for natural-language Polish questions; Azure AD permissions respected (a shop-floor operator doesn't see office commercial documents). Microsoft Copilot Studio as the interface: agent available in Microsoft Teams and via web chat, with a manufacturer-tuned persona (technical, fact-based, always cites the source with a link to the document).

The last 2 weeks were a pilot with a 12-person group of production engineers with intensive tuning based on call recordings and satisfaction metrics. We caught cases like „question about product X returned documents for product X′ (one letter different)”. After two weeks of tuning, answer accuracy crossed 91%. Rollout to the entire factory took 5 days.

Six weeks after full rollout, production engineers reported reclaiming an average of 2 hours a day previously lost to searching. The clearest sign: the 8 daily recurring „what's the tolerance for X?” questions are now answered by the agent in 40 seconds instead of requiring interaction with another employee. The quality team logs an 80% drop in ad-hoc „where's the procedure?” emails.

An unexpected effect: the agent exposed inconsistencies in the procedures themselves. Asked about tolerance for class-A product, the agent honestly showed that two procedure versions existed — from 2019 and 2023 — with different tolerances and no clear marker of which one was current. That kicked off a key-procedure cleanup project — but driven by a concrete list of priorities the agent surfaced, rather than by a quality director's arbitrary call.

A quality audit conducted 3 months post-rollout flagged the Copilot Studio agent as a best practice and a replication candidate. Leadership is now considering the same approach for a second area: sales documentation and the commercial project archive, where office staff report identical chaos accumulated over years.