Product / Bulk Review
Thousands of contracts. One table you can actually read.
Ask the same questions of every document at once. Each one becomes a row, each question a column, and every answer carries the clause it was read from and its own confidence. You review the exceptions, not the corpus.
| Document | Change of control? | Terminate on reorganisation? | Intra-group carve-out? | Consent required |
|---|---|---|---|---|
| Supply Agreement, Kaveri Metals | Yes cl. 14.2conf 0.97 | Yes, on 30 days’ notice cl. 14.3conf 0.94 | No cl. 14.2conf 0.91 | Prior written, counterparty cl. 22.1conf 0.95 |
| Licence Agreement, Orbis Software | Yes cl. 9.1conf 0.96 | No cl. 9.4conf 0.93 | Yes, affiliates cl. 9.1(b)conf 0.95 | Notice only cl. 9.2conf 0.92 |
| Logistics MSA, Deccan Freight | Yes cl. 18.1conf 0.95 | Unclear: cl. 18.1 and Sch. 3 conflict cl. 18.1 · Sch. 3conf 0.41 · exception | Silent Not in document | Prior written cl. 25.3conf 0.90 |
| Lease Deed, Unit 4B Okhla | No Not in document | No cl. 11conf 0.89 | — | Lessor, for assignment cl. 12.1conf 0.93 |
The problem
Nobody has read the contracts
Not all of them. Not recently. A legacy repository is read in the moment a question forces it (a transaction, a dispute, an audit) and then only the part the question touches.
So the answers you have are inherited. The renewal dates came from a spreadsheet somebody maintained until they left. The entity attribution was right when the group last reorganised. The liability positions are known for the forty agreements that mattered in the last deal and unknown for the rest.
The usual quote is several weeks of junior time, and it buys you a point-in-time answer that starts ageing the day it is delivered. Most teams decide the question was not that important.
How it works
You write the questions. In your own words.
There is no schema to design. Describe the job the way you would brief an associate (find anything that blocks the restructuring: change of control, assignment, and any consent we would need) and Orion proposes the columns, each with a line saying why it is there.
You edit that list, add your own, remove what you do not need, and approve it. Then it runs.
Orion · proposed columns
| Proposed column | Why |
|---|---|
| Does the agreement contain a change-of-control clause? | The brief turns on what happens on a reorganisation |
| May the counterparty terminate on a group reorganisation? | The termination right is the operative risk, not the clause itself |
| Is there an intra-group carve-out? | Transfers between group entities may be exempt; this removes false positives |
| Whose consent is required, and in what form? | Consent mechanics decide the timetable |
The economics
Read the exceptions, not the corpus
The point of a review is not that a machine reads every document. It is that a lawyer no longer has to.
Most answers in a contract estate are unambiguous: a date is a date, a named party is a named party. Those settle on the first pass and nobody opens the file. What is left is the set where the instrument is genuinely unclear, contradicts itself, or disagrees with your own records. That is the set worth a lawyer’s afternoon, and it is a fraction of the whole.
Exceptions arrive grouped by the reason they were raised, so most clear in batches rather than one at a time. If a hundred and forty documents share one entity ambiguity, you resolve the ambiguity once.
Cover sheet names “Meridian Holdings”; signature block names “Meridian Infra Pvt Ltd”. Choose once, apply to all.
Resolve for 140Termination right in the body differs from the schedule.
Renewal date in your spreadsheet differs from the signed copy.
Flagged for a person rather than guessed at.
The three commitments
Work that holds up when it is questioned
Accuracy is a claim about averages. What matters in practice is whether one specific answer can be defended, months later, to someone who is not on your side. Three things have to be true for that.
01
Every answer names its passage.
Not the document: the clause. Click a cell and you land on the words it was read from, in the instrument itself. Where the document is silent, the system says so and marks the field derived rather than read, instead of inferring something plausible.
02
Every action names its author.
Accepting an extraction is a professional act, so it is recorded as one. The audit captures views as well as actions, which is what makes an ethical wall provable rather than asserted.
03
Every review keeps its method.
The questions are shown, edited and approved before anything runs, and the approved schedule is stored with the result. When somebody asks what was checked, the answer is a record rather than a recollection.
Your spreadsheet
Your register is a control, not a source
Most teams have a spreadsheet that is supposed to describe the contracts. Most of those spreadsheets have drifted: a value updated in one place and not the other, a counterparty renamed after an acquisition, a date entered from the covering email rather than the instrument.
Bulk Review does not trust it and does not ignore it. It reads both, and where the register and the instrument disagree, that disagreement becomes a finding with the clause attached, rather than a silent decision about which to believe.
This is usually the first thing a review surfaces, and often the most useful. Every downstream number (the total value, the renewal schedule, the board paper) has been built on the spreadsheet. Until the disagreements are settled, none of those figures is safe to quote outside the building.
Licence Agreement, Orbis Software · renewal date
Your register
31 Mar 2027
Contracts register.xlsx · row 212
The instrument
31 Dec 2026
cl. 3.2 as amended by Amendment 1, cl. 2
Entity attribution
A fixed vocabulary, not a guess
Where a group has several arms, attribution is the question the whole exercise usually exists to answer, and it is the one most easily got wrong, because the entity on the cover sheet is not always the entity in the signature block.
You give Bulk Review the list of entities that exist. It may assign a document to one of them or to none. It may not invent a company. Anything that cannot be placed becomes an exception for a person to decide, which is a far better outcome than a plausible answer nobody checks.
Asking the table
Questions about the whole corpus, answered instantly
Once the extraction is done, the register itself can be asked questions, and because it is reading the extracted table rather than re-opening the documents, the answers are immediate and cost nothing.
Answers act on the table. Show me those 63 filters the grid. Add that as a column runs the question across every document. The thread is saved against the review, so a colleague can read how a conclusion was reached instead of asking.
Orion
412 expire within twelve months, and 188 renew automatically, but 63 is the number that matters. In 63 the notice period is longer than the time remaining, so the window to prevent renewal has already closed. Those are not a diary item any more; they are a decision about whether to accept another term.
For in-house teams
A number you can put in front of the board
The questions that reach you are rarely about one contract. What is our exposure if this counterparty fails. Which agreements does the restructuring touch. What renews before March, and what have we already lost the right to stop.
Each of those is a question about the whole estate, and each is currently answered by somebody reading until they run out of time. Bulk Review answers them across every document, and (this is the part that matters when the question came from a board) every figure traces back to the page it came from.
The renewals case is usually the one that pays for itself. Not the agreements expiring, which a calendar can tell you, but the ones where the notice period is already longer than the time remaining. Those have stopped being a diary entry and become a commercial decision, and most teams find them after the fact.
For law firms
Take on the work you have been declining
Volume review is work most firms quote reluctantly. It is hard to price, it consumes juniors, and the margin is poor because the cost is hours.
When the cost is the extraction rather than the reading, the economics invert. You can quote on volume work that was previously uneconomic, and you can take a corpus that would have needed a team and put one lawyer on the exceptions.
The work is shared properly. Several lawyers on one review, each action recorded against the person who took it, presence visible so two people are not settling the same rows. And because the approved schedule is kept with the result, your method is documented, which matters when the other side asks how disclosure was conducted.
What it is not
What we are not claiming
It does not replace the reading that matters.
The exceptions still need a lawyer, and the judgement calls are still judgement calls.
It does not pretend to certainty it does not have.
A field the instrument leaves silent is marked as silent. A date derived from a term rather than read from the page says so.
It does not hide what it could not process.
Files that failed are listed rather than dropped, so the record of what was not reviewed is as complete as the record of what was.
FAQ
Questions about Bulk Review
How long does a review take?
It depends on the corpus and the number of questions. You see the estimate before you approve the schedule, and nothing runs until you do.
Can I trust the answers?
You do not have to. Every answer carries the passage it was read from and its own confidence, scored per cell rather than per document. A contract can name its parties beyond doubt and still leave its entity ambiguous.
What if the extraction is wrong?
You amend it, and the amendment is recorded against you with the original kept. The audit shows the change, who made it, and when.
Can we add a question after the run?
Yes. A new column runs across the whole corpus and takes minutes, not another review.
Does it work on scanned documents?
Scans go through OCR on intake. Anything below the confidence threshold is flagged for a person rather than guessed at.
What about privilege?
Privileged documents are tagged at intake and governed by the same matter-level access as the rest of the Vault.
Where do the documents come from?
A review runs over a corpus. The corpus comes from the Vault.
Deduplicated on intake, tagged by matter, and scoped so nobody sees what they should not.
Keep exploring
Bulk Review
Bring a real corpus
A demonstration on our data proves very little. Bring a sample of your own documents and we will run it in front of you, including the parts it gets wrong, because that is the part you need to see before you buy anything.