Bringing Legal AI into the practice

Standing Lawyers has adopted Claude, Anthropic's AI assistant, as the practice's enterprise legal AI.

This note explains the decision properly: what the assistant does inside the practice, what it is not allowed to do, where client confidentiality sits, and what you should ask any law firm that tells you it uses AI.

Standing Lawyers has adopted Claude, Anthropic's AI assistant, as the practice's enterprise legal AI. The announcement fits in one sentence. The questions it raises do not, and clients are entitled to ask them. So rather than wait to be asked, this note answers them in the order they usually come.

It would be easy to write this the way technology adoptions are usually written about, with the word transformation somewhere in the first paragraph and a stock photograph of a robot beside a gavel. I am going to do the opposite. What follows is a plain description of what has changed inside the practice, what has deliberately not changed, and how the duties a lawyer owes a client survive contact with a tool this capable. The decision took longer than the setup did, and most of that time went into the questions in the second half of this note.

What "adopted" actually means

Start with what it does not mean, because the words "legal AI" now cover everything from a search box to a fantasy.

There is no chatbot on this website. Nothing you send the practice is answered by a machine. No document leaves the practice written by software that no lawyer has read. If you email us at midnight, the reply you get the next morning was written or read, in full, by a lawyer, the same as before.

What has changed is inside the practice. Claude is now the working assistant on the desk: the tool the lawyer turns to first when a document needs to be read, a position needs to be researched, or a draft needs to exist by the afternoon. It sits where the research database, the precedent bank and the late-night first draft used to sit. It is a tool in the practice, not the practice.

The deployment is an enterprise one, on commercial terms, in a workspace the practice controls. That distinction matters more than it sounds, and the section on confidentiality below explains why.

Where the assistant fits

Most legal work has two layers. There is a reading layer: absorbing the forty-page agreement, finding the six clauses that matter, checking them against the last version, locating the statute and the circulars under it. And there is a judgment layer: deciding what the six clauses mean for this client, what to push back on, what to concede, what to write in the email that goes with the mark-up. The reading layer eats most of the hours. The judgment layer is what clients are actually paying for.

Claude works in the reading layer. Three ordinary examples, because the ordinary ones are the point.

A vendor agreement arrives at six in the evening, forty pages, counterparty's paper, and the client wants to sign this week. The old first step was a highlighter and two hours. The new first step is that Claude reads it in one pass and produces the map: where liability is capped and at what, which indemnities run one way, what auto-renews and on what notice, which clauses depart from the positions we normally hold. Then the lawyer does what the lawyer always did, which is read the agreement. But now the reading starts from a map instead of a blank margin, the two hours become one, and the hour that remains is spent on the four clauses that deserve it.

A compliance question comes in: what does a company of this size, in this state, in this sector, actually have to do under a statute that has three sets of rules and a stack of circulars underneath it. Claude assembles the skeleton of the answer and points to the provisions. The lawyer then reads the provisions themselves, in the bare Act and the notified rules, before a word of advice goes out. The assistant shortens the search. It is never the authority. The authority is the text of the law, and there is no version of this practice in which advice cites something the lawyer has not personally read in the original.

A first draft needs to exist: an employment agreement for an unusual role, a policy that has to fit a hundred-person company rather than the ten-thousand-person company most templates secretly assume. Claude produces the first draft from the practice's own starting positions. The lawyer rewrites it. Anyone who has supervised a junior knows exactly what this relationship is: the draft arrives faster and more complete than a trainee's would, and it still goes out under the senior's pen or not at all.

Notice what all three examples have in common. The assistant's output is never the product. It is the scaffolding the product gets built on, and the builder has not changed.

Where it does not fit

Some of this practice's work never touches the assistant, and the line is not drawn by sentiment.

Judgment does not delegate. What to fight over and what to let go, whether a claim is worth bringing at all, when a negotiation has reached the point where the relationship matters more than the clause, whether the honest advice is "do not do this deal": these are not reading problems, and no amount of fluency in a machine makes them one. A model can list the considerations. It cannot own the call, and it will not be asked to.

Anything client-facing goes out only after a lawyer has read every word and adopted it as their own. Not skimmed. Read. The rule inside the practice is a signature rule: if your name would go under it, your eyes go over it, entirely, against the source documents. The convenience of the tool is precisely why the rule has to be mechanical rather than a matter of discretion.

And nothing is filed before a court or a regulator on the strength of the assistant's memory. Every citation is verified in the reported original before it is relied on. The reason for stating this so flatly is in the second-last section of this note.

The confidentiality question, answered properly

This is the question a client should ask first, and the one most firms answer vaguest. Here is the position, in the order the law puts it.

A lawyer's duty of confidence does not soften because a tool is useful. Communications with a legal adviser have been protected in India since the Evidence Act of 1872, and that protection carries forward today in section 132 of the Bharatiya Sakshya Adhiniyam, 2023. The professional duty runs alongside it, and the DPDP Act's obligations around personal data sit on top of both. None of that moved an inch when this practice adopted an AI assistant. If a tool could not be used inside those duties, the tool would go, not the duties.

So the adoption was conditional, and the conditions were not negotiable. Client material is handled on commercial terms under which it is not used to train the model. The workspace is the practice's own, access to it is limited to the practice, and what goes into it is decided by the lawyer handling the matter, not by convenience. Matter material goes in because the matter needs it there, and for no other reason.

There is also a plainer, older safeguard that predates every product feature: restraint. The assistant sees what it needs to do the job in front of it. A clause that needs review does not require the client's entire corporate history for company. This is the same discipline a careful lawyer has always applied before sending anything to anyone, including to opposing counsel, including to a courier. The habit did not have to be invented for AI. It just had to be kept.

What this means for you as a client is simple to state: the duty of confidence you rely on is exactly where it always was, with the lawyer, and the lawyer has arranged the tools to fit under it rather than the other way around.

Why Claude

The honest answer is that the reasons were boring, which is what you want reasons to be when the subject is your clients' documents.

The practical requirements came first. Legal documents are long, and an assistant that cannot hold a whole agreement, its annexures and the previous version in its head at once produces analysis with seams in it. Claude handles documents at the length legal work actually comes in. The drafting and analysis are strong enough to be worth supervising, which is the correct bar: a tool too weak to bother checking is a toy, and a tool you are tempted not to check is a hazard.

The vendor mattered too. Anthropic publishes serious work on how these systems fail and how to make them fail less, offers enterprise terms that a confidentiality-bound profession can actually sign, and builds the controls a small practice needs without a procurement department to negotiate them. That combination is rarer than it should be.

No tool was ever going to be chosen because of a benchmark table. This one was chosen because it could be fitted inside the duties described above without bending them.

What it changes in the shape of a week

A small practice's real constraint is not knowledge. It is the queue.

Work arrives in bursts. Three clients have an urgent document on the same Tuesday, and until recently the only honest options were to triage, to stay up, or to make someone wait who should not have to. The first pass that used to consume the evening now starts in minutes, which flattens the burst. The queue gets shorter not because anyone reads less carefully but because the mechanical part of careful reading has stopped being the bottleneck.

Where does the recovered time go? Into the layer that cannot be accelerated: thinking about the matter, the phone call where the real question finally surfaces, the third read of the clause that decides the deal, the note that explains a risk in words a board will act on. If adopting an assistant let a practice do sloppier work faster, it would be a bad trade at any price. The trade on offer is the opposite one: the same care, applied sooner, with more of the week left for the part of the work that justifies engaging a lawyer at all.

Clients should notice exactly one thing: drafts and answers arrive sooner, and they read the way they always did. If you ever notice a second thing, I want to hear about it.

The failure everyone has read about

AI's characteristic failure in law is well documented and worth describing without cosmetics, because a firm that will not name it should not be trusted to manage it.

Language models are fluent even when they are wrong. Asked for authority, an unsupervised model can produce case names that sound real, citations formatted perfectly, holdings summarised confidently, for judgments that do not exist. Courts in more than one jurisdiction have now dealt with filings built on such phantoms, and the lawyers who signed them discovered that pointing at the software is not a defence. The signature, it turns out, still means what it always meant.

That failure mode is why the practice's rules are mechanical. Nothing is cited that has not been read in the reported original. Statutory positions are checked in the bare Act and the current rules, not in a summary of them. The assistant is treated at all times the way a prudent senior treats a brilliant junior's overnight memo: gratefully, and with a red pen. The gratitude does not extend to filing it unread.

There is a second, quieter risk: subtle wrongness, the analysis that is ninety per cent right in a profession where the missing ten per cent is usually where the liability lives. The defence against that is not technological at all. It is the discipline of reading against source, which is the same discipline that catches a junior's plausible mistake, or one's own at 1 a.m. The tools change. The red pen does not.

Eight questions to ask any firm that uses AI

If your lawyers use AI, or you are choosing lawyers who do, these questions will tell you most of what you need to know. This practice's answers are in brackets, so you can hold it to them.

1. Does anything reach me that a lawyer has not fully read? (No. Every word is read and adopted by a lawyer before it leaves.)

2. Which company's system is it, and on what terms? (Anthropic's Claude, on enterprise commercial terms, in a workspace the practice controls.)

3. Is our material used to train the model? (No. That was a condition of adoption, not an afterthought.)

4. Who can see what goes into it? (The practice. Access is not shared, and what goes in is decided matter by matter.)

5. Are citations and statutory references verified in the originals? (Always. Nothing is cited from the assistant's memory.)

6. What will you not use it for? (Judgment calls, strategy, court filings taken on trust, and anything client-facing without full review.)

7. Who is accountable if something it helped produce is wrong? (The lawyer. The signature rule exists so that this answer can never change.)

8. Will you tell me how this affects my matter if I ask? (Yes, in plain terms, the way this note is written.)

A firm that answers these eight readily has thought about the tool. A firm that cannot answer the third or the seventh has adopted something it does not yet understand, and the gap will be yours to absorb.

What stays the same

The profession has absorbed new tools before. The typewriter did not draft, the word processor did not negotiate, and the online research database did not make judgment optional; each one moved hours from the mechanical layer to the thinking layer, and the lawyers who used them well simply thought more per week than the lawyers who did not. This adoption is that event again, larger.

What stays the same is everything a client actually relies on. The duty of confidence, in full. The verification of every authority against its original. The judgment about what your business should do, made by a person who can be questioned across a table and who answers for the advice. And the signature, which has survived every tool the profession has ever adopted, because it is not a formality. It is the point.

The work is faster now. The word is still ours.