A thousand-page manual, one answer in 160 languages

The group produces in plants spread across several countries, and supply chain planning runs through a single shared platform. Demand, production capacity, stock, automation chains: it all lives in there, and it is all described in a technical manual of almost a thousand pages, written in English by the people who implemented the system.

On paper, every answer was already written. In practice, the manual was opened almost only by the people who wrote it. A planner on the plant floor who needs to understand why a dashboard is not updating has no time to hunt for the right chapter in a thousand-page PDF, possibly in a language that is not their own. They call the colleague who remembers. And in every country that colleague ends up becoming the point everything passes through.

The cost showed most with new starters. Weeks of shadowing before they could work on their own, with knowledge passed on by word of mouth from one person to the next, changing a little at every step.

Then there was a second audience with different needs. The people who configure and maintain the platform do not want the simple explanation: they want the exact parameter and the maintenance procedure. The same tool had to serve both, without forcing anyone to read the answer written for the other.

• Answers in more than 160 languages, for the people who use the system and the people who maintain it
• Every answer says which part of the manual it comes from
• The documents stay on SharePoint, with the permissions they already had

The manual has to be prepared first

The first job was not writing the agent's instructions. It was understanding how Copilot Studio reads documents. Each file has a limit on how much text the agent takes into account, and beyond a certain size the search does not reach the end. Tables, which are everywhere in a technical manual, are read badly inside SharePoint. Loading the whole PDF would have produced an agent that was brilliant on the first hundred pages and inventive on the other nine hundred.

So the manual was split by topic rather than by page: one file for each subject that stands on its own, from demand planning to system maintenance. When a question comes in, the agent retrieves the right piece from its knowledge source and cites where it took it from.

An English base, answers in any language

The knowledge base stayed in English, the language the manual is born and updated in. The agent understands the question in whatever language it arrives and answers in the same one. When the manual changes, one text gets updated.

Rules first, answers second

The agent has precise instructions: answer only on the basis of the manual, say where every answer comes from, and admit when the information is not there instead of building a plausible one. The people reading those answers make operational decisions on them, and an agent that is confident about something it does not know does more harm than no agent at all.

Tested to make it fail

Before opening it to people we put it through a set of questions written on purpose: easy, procedural, ambiguous, out of scope, about things the manual does not say. Every wrong answer became a fix, and the fixes went into an improvement report shared with the client. The agent is published in Teams, where people already work, and the rollout starts in one or two countries: users rate each answer, and only then does it widen.

New starters get an answer straight away, in their own language, instead of waiting for the expert colleague to be free. That is where the biggest saving shows: over 70% of the time that used to go into training.

The expert colleague, meanwhile, stops being the one person an entire country has to go through. And the manual, which used to be opened only in emergencies, becomes what it was meant to be from the start: the source everyone draws on, updated in one place.