Product · first application
In developmentRead, check, search and cite, with a person in the loop.
Doc Intel takes in your documents, proposes what is in them, has a person confirm it, and answers questions with the exact passage each answer came from.
- Mutual NDA, Riverside InstrumentsField · Governing lawState of Delawarep.6 §11.4Conf. 0.61Low confidence
- Notice of award, sample R01Field · Budget period end2027-06-30p.1 §IIIConf. 0.54Low confidence
- Lease amendment no. 2, Building CClassification · Tier3Restrictedp.1 headerConf. 0.93Always reviewed: tier placement
Ingest and split
In developmentBring documents as they are.
Contracts, legal agreements, grant records, forms and notes: text, tables, diagrams and charts. When one file holds several documents, Doc Intel splits it into separate ones so each can be read, filed and permissioned on its own.
- Mutual NDApp. 1–7
- Master services agreementpp. 8–16
- Signature pagepp. 17–18
Sample data
Templates and extraction
In developmentDescribe a document type once.
Set up a template for each recurring type, such as your NDA structure or a funder’s form. Matching documents are recognized, routed to the right repository and given a default access setting. Each proposed field shows its confidence and its source.
| Field | Value | Source · confidence |
|---|---|---|
| Effective date | 2025-03-01 | p.1 preamble 0.97 |
| Notice period | 30 days | p.4 §7.2 0.95 |
| Confidentiality survives | 3 years | p.5 §9.1 0.92 |
| Governing law | State of Delaware | To review 0.61 |
Human review
In developmentMachine output is a proposal, not a fact.
Low-confidence values go to a review queue. A person accepts, corrects or rejects each one, and the decision is recorded. Some situations go to review whatever the confidence.
What sends a value to a person
- Confidence below your threshold. You decide where the line sits.
- Placing a document in a hidden tier. Always reviewed, regardless of confidence.
Repositories and tiers
In developmentNeed to know, down to a single document.
Documents live in repositories with their own access, and any document can carry its own. On top, your organization defines clearance tiers with your own names and levels. Search enforces them passage by passage, not just at the door.
Tiers defined by your organization
Example labels- 1InternalEveryone in the legal team
- 2CounselAttorneys and named paralegals
- 3RestrictedNamed people only. Invisible to everyone else, including in search.
Search with citations
In developmentAsk in words. Match exact figures too.
Search combines meaning with exact keyword matching, so a case number or a dollar figure actually matches. Every answer cites the passage it came from and is marked “not verified”, with a provenance line, until a person checks it.
Model selection and benchmarking are under way.
Two agreements set that cap: the Riverside master services agreement1MSA p.9 §14.2 and lease amendment no. 22Lease p.3 §5.
Sample data
Reproducible answers
In developmentSame access, same question, same answer.
Two people with the same access who ask the same question get the same answer, and so does one person asking twice. We engineer and test this as a guarantee. Someone with less access gets an answer built only from what they can open, with nothing in it that hints at the rest.
30 days for the NDA, 60 days for the MSA.
30 days for the NDA, 60 days for the MSA.
The vendor summary lists a 30-day notice period for Riverside.
Cross-document findings
In designContradictions and outliers, across a repository.
Finds similar documents, clauses that contradict each other, and values that stand out. Each person sees a finding only as far as their access allows.
Notice period differs between two Riverside agreements
Audit trail
In designA record of who saw and changed what.
Viewing and sensitive actions are recorded in a tamper-evident trail shared by every application. Your organization sets which tiers are audited. Doc Intel already emits these events; the shared audit service is being built.
- 10:42:07J. Rivera opened Lease amendment no. 2 3Restricted
- 10:39:51J. Rivera accepted Tier placement, proposed at 0.93
- 10:31:12A. Chen corrected Governing law on Mutual NDA
Design concept · sample data
Where it runs
Your document text never goes to a public AI API.
Every model that reads document content is an open-weight model hosted on Adduce’s own infrastructure on Microsoft Azure, with each customer isolated from the others. Read Trust and security.
Get in touch
Tell us which documents you’d bring.
Five screens are built and running on sample data. Tell us about your document types and we will tell you plainly what exists and what doesn’t.