Disagreement is preserved, never resolved for you
It would be easy to have a final agent pick a winner and produce a cleaner report. We do not, because the moment we do, you lose the ability to see what the other side will argue.
§ About Us
Built on one conviction: you need the bad news first
Every other decision in this product follows from that. If we ever ship a feature that makes an output look tidier by hiding what it does not know, we have made the product worse.
§ 01 · The problem we started from
Confident prose is the wrong output for a contested matter
Put a case file into almost any AI tool and you get back something fluent. It reads well. It sounds settled. And it gives you no way of telling which of its conclusions were obvious and which were finely balanced, because the model has already averaged that away before it started writing.
For most tasks that is fine. For a contested matter it is precisely backwards. Litigation is not decided where everyone agrees. It is decided at the point where two honest readings of the same document diverge — and that point is exactly what a single confident summary erases.
So we built the opposite. Independent agents, each holding one posture, reasoning separately on the same record. Where they concur, that is recorded. Where they split, that is recorded too, with both lines of reasoning and the material each one relied upon. The split is not a failure of the system. It is the output.
§ 02 · What follows from it
Four design rules we do not break
Each one costs us something. Each one is there because the alternative would make the product less useful to someone who has to defend the output.
Every conclusion carries its source
If an assertion cannot be tied back to a document, a page and a passage, it does not belong in the report. Unsourced conclusions are how practitioners get into trouble.
Gaps are declared, not filled
Where a document is pleaded but not produced, or an assertion is undated, the report says so. It does not infer what was probably there. An honest blank beats a plausible invention.
The adversarial agents are genuinely adversarial
Opposing Counsel is instructed to defeat your case, not to note a mild concern. A tool that flatters your theory is worse than no tool, because it sends you into court reassured.
§ 03 · Who operates it
A named company, at a real address
You are about to put privileged material into someone's software. You are entitled to know exactly whose.
Registered entity
Rank First Technologies Private Limited
Litora AI is owned and operated by Rank First Technologies Private Limited, a private limited company incorporated in India under the Companies Act, 2013.
- Corporate Identity Number U58201PB2026PTC068900
- Registered office at S.A.S. Nagar (Mohali), Punjab 160055, India
- Governed by the laws of India
- Billing in pounds sterling
What we are not
Not a law firm
We are a software company. We do not practise law, we do not represent anyone, and we hold no authorisation from any bar council, law society or regulator — nor do we need one, because we do not do the thing that would require it.
Nothing this platform produces is legal advice, and using it creates no professional relationship between you and us. The disclaimer sets that out in full, and we would rather you read it before you buy than after.
Where to reach us
Real people, named routes
- General — contact@litoraai.com
- Billing and refunds — billing@litoraai.com
- Privacy and data requests — privacy@litoraai.com
- Security and vulnerability reports — security@litoraai.com
- Grievance Officer — grievance@litoraai.com
§ 04 · How we build
Opinions we hold about AI in legal practice
Stated plainly, so you can disagree with them before you spend money rather than after.
Models are wrong sometimes, and saying so is not a weakness
Every vendor knows this. Most bury it in a footnote. We put it on the home page, because a practitioner who believes the output is reliable is more dangerous than one who knows it is not.
Preparation can be automated. Judgement cannot
Reading four hundred pages and building a chronology is work a machine should do. Deciding what to argue is work you should do. We are deliberately built to stop at that line.
Case material is not training data
Your documents are processed to answer your instructions and nothing else. They are not used to train models, ours or anyone else's, and our provider agreement forbids it.
Charge for the matter, not the month
A subscription bills you in the months you are quiet and gets in the way in the months you are busy. Credits do neither. You pay for the matters you actually run.
We would rather charge honestly for the first matter
Every analysis costs us real money the moment it starts. A free tier would mean either a crippled product or a cost recovered somewhere less visible. Ninety-nine pounds, stated up front.
Jurisdiction-neutral by design
The platform reasons from the record you supply rather than from a fixed body of law. That makes it useful across forums — and makes verifying the law against the current source your job.
§ 05 · What we will not build
The features we have been asked for and declined
A roadmap is defined as much by what is left off it.
5.1
A one-click “final answer”
Asked for often. The bench gives a reasoned view at the end of the argument, and that view names what it could not decide on the record it was given. A single line with no reasoning attached would be adopted without being read, which is the opposite of what this is for.
5.2
A percentage on its own
We do give a figure — advocates asked for one, repeatedly, and refusing outright was precious rather than useful. But a bare number would be repeated to clients within a week as though it were a measurement, so it never appears alone. It carries a confidence level computed from how much of the file was actually readable, the reasoning behind it, and what would move it. When confidence is low the figure is pulled toward the middle and the screen says, in terms, that it should not be repeated to a client as it stands.
5.3
Automatic filing or service
Nothing this platform produces should leave your desk without a qualified professional having read it. Building a button that skips that step would be building the thing we most want to prevent.
5.4
Cross-customer insight
“Learn from every matter on the platform” is a good pitch and an unacceptable use of privileged material. Your matters are read for your matters only.
5.5
Consumer self-litigation
This is a tool for people with professional training and professional obligations. Marketing it to litigants in person would produce confident filings nobody competent had checked.
5.6
Citations presented as checked
We could show case names without the unverified label and the product would look considerably more finished. Nothing here reads a law report, so that label stays until something does. A fabricated citation carried into a court is the one failure we would not be able to make good.
§ 06 · What we got wrong
This ran sixteen agents. It now runs three.
The first version instructed sixteen specialists across every file and reported where they divided. It was the thing we were proudest of, and it was wrong.
That the split was the finding
If sixteen readings of one record diverged, the divergence would show you what the other side would argue. That reasoning is sound — when the readings are genuinely independent.
They were not independent
Sixteen agents were one model reading one record from angles that overlapped heavily. They mostly agreed, and agreement between instances of the same model is not corroboration — it is one piece of reasoning arriving sixteen times, looking like a bench that concurred.
Two advocates and a bench
The structural work — timeline, contradictions, procedure, provisions — happens once, properly, in the assessment. Then two advocates argue, which is a genuine disagreement because they are instructed to opposite ends. The bench presses both and gives reasons.
We are saying this on our own About page rather than quietly changing the number on the pricing table, because a firm deciding whether to trust this product is entitled to know how we behave when we find out we were wrong.
Disagree with any of this?
We would rather hear it. If something on this page is wrong, or a design rule we hold looks mistaken from inside a practice, tell us and a person will read it.
Litora AI is operated by Rank First Technologies Private Limited, a private limited company incorporated in India under the Companies Act, 2013 (CIN U58201PB2026PTC068900). Not a law firm. Not legal advice.