Home / Solutions / AI agents and automations
Private AI assistant over company documents and ERP data, read-only
What I hear
We want to ask questions of our own documents and ERP data without sending them to a public AI service, and without the assistant being able to change anything.
A private assistant indexes your documents and a read-only view of your ERP data on your own hardware or private network. A person asks a question, the system finds the relevant passages and records they are allowed to see, and a language model writes an answer that cites them. Nothing is sent to an outside service unless you choose a hosted model.
How does retrieval work?
Documents are split into passages, converted into numeric vectors by an embedding model and stored in an index along with the source file, page and permissions. A question is converted the same way, the closest passages are fetched, and those passages (not the whole library) go to the model with an instruction to answer only from them and to say when they do not contain the answer. For ERP data, I prefer exact queries over fuzzy search: a fixed set of read-only questions such as stock level or open invoices for a customer, run as database queries, with the results handed to the model to phrase.
How are permissions mirrored?
This is the part most demos skip. Each indexed passage carries the access rules of its source: which folders, groups or users can open it. At question time the assistant filters by the asking user’s identity before retrieval, so a restricted passage is never in the model’s input. The rules are read from your existing directory or file system and refreshed on a schedule, so when someone loses access, the assistant follows. I test with real accounts at different levels and attempt to extract restricted content on purpose.
Local model or hosted model?
| Local | Hosted | |
|---|---|---|
| Data leaves your network | No | Yes, to the provider |
| Answer quality on hard questions | Generally lower | Generally higher |
| Hardware | You buy and run it | Provider’s |
| Ongoing cost | Power, upkeep | Usage fees |
Hosted providers publish data-handling terms, and some businesses are comfortable with those. Others cannot send documents out at all. Either way, the retrieval and permission design is the same, and the model can be swapped later.
What should you expect on quality?
Good at: finding the relevant policy, contract clause or procedure, summarising a document, answering with a citation you can click. Weaker at: questions needing many documents combined, arithmetic across tables, scanned documents with poor text recognition, and anything where the documents contradict each other. It can still be wrong, which is why every answer shows its sources and the interface makes checking them one click. I do not quote benchmark figures, because the only number that matters is how it does on your own questions, so we build a test set of real ones first.
What hardware?
It scales with model size and how many people ask at once. A small model can run on one capable workstation, larger ones need a GPU server. I measure candidate models on your documents and then recommend, rather than naming a machine in advance.
When is off-the-shelf better?
If your documents already live in a platform that has a built-in assistant and you accept its data terms, use that first. Private builds make sense when data cannot leave, or when ERP data and documents need to be answered together with your permissions.
What I need from you
- A sample of documents and 30 to 50 real questions people ask.
- A read-only database account or export for the ERP.
- Someone who knows the access rules and can test with me.
What it replaces
searching shared drives by hand and asking the one colleague who knows where things are
How pricing works
One fixed price, agreed before I start. Optional monthly maintenance after handover.