Blog
What should we give an AI tool before asking it about a matter?
Give it the matter's papers in full rather than your own summary, a short note of what the matter is and what stage it is at, and one specific question. Vague inputs produce vague answers.
Alesis · · 5 min read
Three things: the matter's papers in full, a short statement of what the matter is and where it has got to, and a specific question you would be willing to put to a colleague. Most disappointing answers from an AI tool come from thin inputs rather than from the tool itself. If you feed it a two-line summary and ask an open question, you will get something that reads well and tells you nothing.
Start with the papers, not your summary
The instinct is to summarise the matter first, then ask. Resist it. Your summary carries your existing view of the case, including the bits you have stopped noticing, and an answer built on it will simply agree with you.
Give the tool the underlying documents: the correspondence, the contract, the pleadings, the attendance notes, the expert material, the client's own bundle of papers. Anything you would hand a new fee earner taking the file over is the right starting point.
There is a second reason. A summary cannot be cited. If an answer rests on the documents themselves, you can check it against a page. If it rests on your paraphrase, the only thing you can check is whether the tool repeated you accurately, which is not worth knowing.
Where a file is large, resist the urge to pick out the twenty documents you think matter. That decision is itself part of the work, and it is the part most likely to be wrong on an inherited or long-running file.
Say what the matter is and where it has got to
A short orientation makes a real difference. Two or three sentences is usually enough:
- what kind of matter it is, and which side you act for
- the stage it has reached, for example pre-action, post-issue, post-exchange, or nearing completion
- which jurisdiction's law and procedure applies
- anything unusual about the client or the retainer that shapes what is useful
Jurisdiction matters more than people expect. England and Wales, Scotland and Northern Ireland differ, and an answer that quietly assumes the wrong one will look perfectly plausible.
Stage matters because it changes what is worth saying. A question about a document's weaknesses gets a different useful answer before issue than it does a fortnight before trial.
Mind the gaps in what you upload
Files are rarely complete or tidy. Before you rely on anything, know what is missing and what is unreadable.
Common problems:
- scanned documents where the scan quality defeats text recognition, so pages are effectively blank
- attachments that were never saved to the file, only referenced in an email
- duplicates and near-duplicates, including several versions of the same draft with no clear labelling
- documents that exist only in a case management system and not in the folder you exported
- material from the client sitting in a mailbox rather than on the file
A tool that silently skips a page it cannot read is worse than one that says so. When you first set a matter up, it is worth asking directly what the tool could not read, and checking that number against your own sense of the file.
Duplicates deserve a word of their own. If three versions of a contract are on the file and only one was executed, an answer may be drawn from the wrong one. Label the executed version, or say which it is.
Ask one answerable question at a time
The biggest single improvement most fee earners can make is to narrow the question. Compare:
- "What do you think of this case?"
- "What does the file say about when the defendant first knew about the defect, and which pages support that?"
The first produces an essay. The second produces something you can verify in ten minutes and either use or discard.
Good questions to put to a matter file tend to share features. They ask what the documents say rather than what the answer should be. They ask for the source of each point. They are limited to one issue. And they invite the tool to say that the papers are silent, rather than pushing it to produce something.
For a larger piece of work, ask for the steps first. Seeing how a tool proposes to approach a review, before it does it, is cheaper than reading a finished answer and trying to work out what it did.
Be deliberate about what you do not put in
A few things are worth a moment's thought before they go anywhere near a tool.
Keep matters separate. Material from one client should not sit in a conversation about another, even where the same fee earner runs both, and even where it would be convenient.
Check the position on material you hold subject to an undertaking, an order restricting use, or a confidentiality club. The fact that a document is on your file does not always mean you are free to process it however you like.
And check what your engagement terms and your data protection documentation actually say, so that what you do matches what you have told clients and recorded internally.
Where Alesis fits
Alesis is an AI assistant for UK law firms, made by L25 Limited, used through the web browser, with one conversation for a matter. It answers questions about that matter from the matter's own papers and names the page each answer came from; if the papers do not say, it says so. Documents are read page by page, so citations point at pages, and any page it could not read is flagged rather than skipped. For a bigger question it sets out the steps first and waits for the fee earner's go-ahead.