Document AI that answers from your documents and reads your scans into data
Document AI that answers from your own documents and data with the source shown for every answer, and that reads scans, invoices and forms into structured data for a person to check.
The answer is in a policy, a contract or a manual, and two versions often disagree. Staff search, or ask a colleague. General chat tools do not know your documents. Much information still arrives as scans, photos and PDFs that someone retypes.
Every answer shows the passage it came from
The assistant searches your own documents and data (retrieval, or RAG). It answers in the language of the question and shows the passages it used. When no source supports an answer, it says so and names the person who owns the topic. For documents that need reading, we build extraction. OCR and vision models turn scans, invoices and forms into fields. Fields the model is unsure of go to a person to check.
Systems it touches
- SharePoint and file shares
- Policies, contracts and manuals
- Scans, invoices, forms and handwritten answers
- Structured data from your databases
Example
Two expense documents that disagree
A made-up case that shows the mechanism. It describes no client.
- An employee asks whether a taxi to the airport can be claimed.
- The assistant searches only the documents that employee may open. It finds the travel policy and an older expenses guide.
- The travel policy allows the claim with a receipt. The older guide says nothing about taxis.
- The assistant answers from the travel policy, quotes the clause and links to both documents.
- It flags the gap to the HR policy owner, who updates the older guide.
- The question, the answer and both sources are logged, and the question joins the test set.
It works in the systems you already have
We choose models by how well they answer your questions and read your documents, including your own terms and your staff’s handwriting. For sensitive material, the search, the extraction and the model can all run inside your systems or on our private GPUs.
Where your data staysPermissions limit what it can touch
Where a person approves
- Document owners check answers on their topics before launch
- If no source supports an answer, it gives none rather than guess
- Extracted fields it is unsure of go to a person before they are used
- A reader only gets answers from documents they are already allowed to open.
- Every answer links to its sources, so a reader can check it against the document.
- The test questions and sample documents run again on every change to documents, instructions or model.
Everything we build is handed to your team
Who owns the code, and how handover works- A search index of your documentsKept inside your own systems, updated from the source.
- The assistant, or the extraction pipelineCode in your repository, running in your accounts.
- Tests built from your own staff’s questionsReal questions and sample documents in your repository, run again on every change.
How we would start
Discovery, then a trial on one set of documents
We choose one set of documents with a clear owner. We collect real questions and samples from the people who use them, and build the assistant or the extraction against those tests.
How we work togetherQuestions
How do you know the answers are right?
We test them. The test questions are written the way your staff write them. Every answer shows its source. We choose a model only once it passes the tests.
What happens when documents change?
The search index updates from the source, and the test questions run again. A policy change cannot quietly break the answers that depend on it.
Can it grade written answers or read handwriting?
Yes. The assistant scores each written answer against your rubric and gives a reason for every score. Handwritten answers are captured as images and read by the OCR model you choose, and unclear readings are flagged for a person. A teacher or reviewer confirms each grade before it is used.
Send us the documents your staff search most
Add the questions people ask about them. We will explain how we would test an assistant against them before anyone relies on it.
Start a project