Document Intelligence
Questions answered across every contract, not one document at a time.
What is document intelligence?
Document intelligence means answering questions across a whole collection of documents at once, such as which agreements contain a particular obligation, rather than extracting fields from a single document.
Reading one document is a solved problem. The valuable and harder question is collective: which of our two hundred contracts contain a change of control clause, what do our policies collectively say about overtime, which agreements expire before June.
Those questions currently get answered by somebody reading, which means they mostly do not get asked. The cost of asking is high enough that people work from assumption instead.
The difficulty underneath is that most organisations do not actually know what documents they hold. An answer covering the documents you gave us is not an answer covering your obligations, and that distinction is where this work either succeeds or quietly misleads.
Why coverage comes before answers
Duara builds document intelligence in Kenya over contract sets, policy libraries and correspondence archives.
We state coverage before we state answers. An organisation that receives a confident answer over sixty percent of its contracts has been misled in the most dangerous way available, because the answer looks complete and the gap is invisible.
Who this is for
- Questions about your documents go unasked
- Contract obligations are not centrally known
- Policy positions are scattered across documents
- Somebody needs to know what is in the archive
Industries we serve
- Legal and professional services
- Property and construction
- Financial services
- Insurance
- Large corporates
- Public sector suppliers
Problems this solves
These are collective questions that individual document reading cannot answer at a reasonable cost.
- Collective questions never asked. Finding which agreements contain a clause requires reading all of them, so the question is answered by assumption instead.
- Obligations nobody has aggregated. What the organisation has actually committed to across all its contracts exists nowhere as a single view.
- Policies that disagree with each other. Several documents cover the same subject and nobody has read them together to see whether they are consistent.
- Unknown archive contents. Nobody can say what documents exist, which means no answer over them can honestly claim completeness.
What we deliver
Questions answered across the whole collection with coverage stated explicitly, obligations aggregated, contradictions surfaced, and an honest account of what the collection does not include.
Questions answered across the whole collection
Coverage stated before any answer is given
Obligations aggregated into a single view
Contradictions between documents surfaced
Dates and deadlines extracted across the set
A list of what could not be read or found
Features
- Coverage stated first. An answer over sixty percent of your contracts presented confidently is the most dangerous output available, because the gap is invisible.
- Collective rather than individual. The value is in questions across the whole set, since reading one document was never the expensive part.
- Contradictions surfaced. Where two policies disagree we report both rather than resolving it, because choosing would be making policy.
- Unreadable documents listed. Poor scans and damaged files are reported rather than silently omitted from the answer.
- Sources cited per answer. Which documents produced an answer, so it can be checked rather than trusted.
Benefits
- Questions get asked at all. Lowering the cost of a collective question changes which questions an organisation bothers to ask.
- Obligations become visible. A single view of what you have committed to is frequently the first time it has existed.
- Contradictions found before they matter. Two policies disagreeing is better discovered internally than by a party relying on one of them.
- Honest completeness. Knowing what the answer does not cover is what makes it usable in a decision.
Technologies we use
- Corpus assembly. Establishing what documents exist and which are in scope, before any question is answered.
- Cross document querying. Answering questions across a whole collection rather than one file at a time.
- Obligation extraction. Pulling commitments, dates and conditions into an aggregated view.
- Contradiction detection. Identifying where documents in the set disagree with each other.
- Coverage reporting. Stating what was read, what was unreadable and what was not included.
Coverage is the whole risk. A system answering over the documents it was given, without stating what that excludes, produces answers that look authoritative about a corpus it never saw.
How the process works, step by step
Document intelligence usually takes five to nine weeks, and the first phase is establishing what documents you actually have.
-
Step 1
Free document review
We establish what collection exists and, more importantly, what is likely missing from it.
-
Step 2
Corpus assembled
Documents are gathered and their condition assessed, which frequently reveals gaps nobody knew about.
-
Step 3
Coverage established
What proportion of the real collection is readable is stated before any answer is produced.
-
Step 4
Questions answered
Collective queries run across the set, with sources cited and coverage attached.
-
Step 5
Gaps reported
What could not be read or found is listed, since that is what makes an answer usable.
Quality assurance and how we work
The commitments attached to answering questions people will rely on.
- Coverage is stated before answers. A confident answer over part of your contracts is the most dangerous output available, because it looks complete and the gap is invisible.
- Unreadable documents are listed. Silently omitting a poor scan from an answer is how a gap becomes invisible, so they are reported explicitly.
- Contradictions are reported, not resolved. Where two policies disagree, choosing between them would be making policy, which is your decision rather than ours.
- Every answer cites its documents. An answer over a corpus cannot be verified unless you can see which files produced it.
- We do not interpret legal effect. Reporting that a clause exists is retrieval. Saying what it means for you is advice, and that stays with your lawyers.
Frequently asked questions about document intelligence
How is this different from reading a document?
Reading one was never the expensive part. The value is collective questions, such as which of two hundred contracts contain a particular clause.
What is the main risk?
Coverage. An answer over sixty percent of your contracts presented confidently is dangerous precisely because it looks complete and the gap is invisible.
Do you state coverage?
Before any answer, always. What proportion of the real collection was readable is the context that makes an answer usable.
What if documents cannot be read?
They are listed rather than silently omitted. Quiet omission is exactly how a gap becomes invisible.
What if two agreements say different things?
Both are reported. Choosing between them would be making policy, which is your decision rather than ours.
Does it tell us what a clause means?
No. Reporting that a clause exists is retrieval. Saying what it means for you is advice and it stays with your lawyers.
Can it find contract expiry dates?
Yes, across the whole set, which is frequently the first question clients ask and the one they cannot currently answer.
Do we know what documents we hold?
Most organisations do not, and establishing that is the first phase rather than an assumption we make.
How soon can we query the collection?
Five to nine weeks, with the first phase spent establishing what collection actually exists.
What happens with poor quality scans in the set?
Yes, with quality affecting reliability, and anything too poor to read confidently appears in the coverage report rather than being guessed at.
Can it handle correspondence archives?
Yes, and email archives frequently contain commitments that never reached a contract, which is why they are worth including.
Does it work on Swahili documents?
Yes, though most commercial documentation we see is in English, and we test on your material rather than assuming.
Can we ask questions ourselves?
Yes, and lowering the cost of asking is most of the point. Questions that were too expensive to answer start getting asked.
Is this the same as document processing?
No. That extracts fields from individual documents. This answers questions across a whole collection.
Can it monitor new documents?
Yes, so a new agreement joins the collection and future answers include it rather than going stale.
Will it find obligations we forgot about?
Frequently, and that is usually the most valuable single output of the first run.
Does it replace legal review?
No. It narrows what needs reading and cites what it found. The review itself stays with your advisers.
Can it work across systems?
Yes, where documents live in several places, though gathering them is part of the corpus work rather than an afterthought.
Where do our documents sit during this?
In your own systems. Processing runs against your infrastructure rather than requiring an export to us.
What does document intelligence cost?
Collection size and document condition determine it, quoted after the document review.
Glossary of terms
Document intelligence terms in plain language.
- Corpus
- The full collection of documents a question is answered across.
- Coverage
- What proportion of the real collection was actually readable and included.
- Cross document query
- A question answered across many documents at once rather than one at a time.
- Obligation
- A commitment contained in a document, worth aggregating across a contract set.
- Contradiction
- Two documents in the set saying different things about the same subject.
- Citation
- The specific documents that produced a given answer.
Related services in AI Solutions
Talk to us about document intelligence
Ask which of your contracts expire in the next six months. If nobody can answer without reading them, that is the gap.
Start your project