Answers from your documents, with the source cited.
A general purpose chat model doesn't know your procedures. An assistant that reads your documents and states which document it drew each answer from is a genuinely useful working tool.
From input to result
- Inputs: PDF and Word files, Procedures and SOPs, Meeting minutes, Voice input
- Application core: Content index, Semantic search, Language model, Source citation
- Results: Answer with citation, Summary, Draft record, Task list
When it makes sense
- The same question keeps coming up, and the answer is buried in a document.
- New colleagues spend their first month asking more than working.
- Procedures exist, but no one finds them in time.
- Replies to customers get written from scratch every time.
What the first scope includes
- Selecting materials and agreeing what goes into the index
- Ingesting and indexing documents (PDF, Word, minutes)
- Retrieval augmented answers with a mandatory source citation
- Roles and permissions by document group
- Logging questions and answers for quality review
- A simple interface for the team, with no special onboarding needed
Why a general purpose model isn't enough
Language models are good at form and poor at facts they don't know. If you ask a general model what your complaints procedure is, it will answer confidently and wrongly. That's why a company assistant must not guess, it has to answer from actual documents and state which ones.
Technically, this means your documents are split into meaningful chunks and stored in an index. When a question comes in, the system first finds the most relevant chunks, and the model then composes an answer from them. Alongside the answer, it shows the document and the location the content came from. If the material doesn't contain an answer, the assistant says so rather than inventing one.
This distinction matters for anything where a mistake costs something: work instructions, contractual deadlines, safety procedures and responses to customers.
Where it pays off most
The assistant is most useful where knowledge is written down but scattered, and where the same question comes up several times a week.
- Internal instructions and procedures no one finds in time
- Onboarding a new colleague without constantly interrupting the team
- Searching contracts, quotes and technical documentation
- Drafting a reply to a customer from existing materials
- Summarising a long set of minutes into tasks with owners
Access, privacy and control
The assistant must not show more than a person would be allowed to see. Permissions are therefore set at the material level: HR documents are visible to HR, technical instructions to the whole team. Questions and answers are logged, so it's possible to check what the system answered and on what basis.
When personal data is involved, it must be decided in advance which documents are even allowed into the index and where the processing takes place. This isn't a formality, it's a condition for the tool to be usable in industries with sensitive documentation.
A pattern from practice: a voice note or a short field entry becomes the basis for a structured report that a person reviews and confirms before it is submitted. The same principle applies everywhere, the model prepares a draft, responsibility stays with the person.
What the assistant isn't
It is not a substitute for well organised documentation. If procedures are outdated or contradict each other, the assistant will simply expose that faster. Often the first tangible result of a project is cleaning up the material itself.
Nor is it a decision making system. It does not confirm, send anything externally, or close deals without a person. It suggests, and it cites its source.
How it works
- We clarify the process: Which questions repeat, and which materials actually answer them.
- Clickable prototype: A trial on a limited set of documents, to check answer quality.
- Working first version: An index of the whole agreed material set, with permissions and logging.
- Handover: Instructions for adding new documents and an agreement on regular updates.
Frequently asked questions
- Do our documents go into a public model for training? No. Material is kept in your own index and used only to compose an answer. Processing terms are agreed in writing beforehand.
- What if the assistant can't find an answer? It says the answer isn't in the material. That's a deliberate design choice, better an empty answer than a made up one.
- How many documents are needed to get started? A few dozen documents from the area where questions repeat most is enough for a meaningful trial.
- Can it connect to our existing storage? SharePoint or a similar solution is technically a common candidate as a document source. Whether the connection is feasible I check beforehand, based on your licence and access.