How a mid-size law firm reads every contract with a Claude legal assistant

The review work that eats a law firm’s junior hours

A mid-size law firm sells judgement, but a lot of its day goes to reading. Supply agreements, leases, services contracts and shareholder deeds arrive in volume, and someone has to read each one in full before a partner can advise on it. That first pass usually lands on a junior solicitor or a paralegal, who works through the document clause by clause, checking indemnities, liability caps, termination rights and governing law against the firm’s house position. It is necessary, it is slow, and it is expensive.

The trap we see is not the time alone, it is the inconsistency. Two juniors reading the same contract on different days flag different things. A clause that one catches, another reads past, because the check lives in a person’s head rather than in a shared standard. On a busy week the review gets compressed, and that is exactly when an unusual indemnity or a buried liability cap slips through. A single missed clause can shift risk onto a client or unpick a deal, and the cost of that lands long after the bill.

Research has the same shape. A lawyer needs the firm’s own precedent on a point, knows it exists somewhere in a decade of matters, and spends real time finding it. The knowledge is in the firm. It is just not at hand when the contract is open and the clock is running. The work has a high bar too. Whatever helps must respect client legal privilege and confidentiality, sit within the professional conduct duties of the Legal Profession Uniform Law, and handle personal information under the Privacy Act 1988.

Why Claude, and what grounds it

The aim is a first reader that never tires and never skips a clause, with a lawyer reviewing everything it produces. We headline these builds on Claude, Anthropic’s large language model, because the core task is long-document reading and careful reasoning, which is where it is genuinely strong. A commercial contract runs to many pages of cross-referenced defined terms, and Claude holds that context across the whole document and reasons clause by clause without overstating what it found. For legal work, a model that reasons carefully and flags uncertainty beats one that sounds confident and guesses.

The model alone is not enough, so retrieval grounds it. Rather than rely on what the model learnt in training, the assistant retrieves from the firm’s own clause library and precedent bank through retrieval-augmented generation (RAG). When it flags a liability cap, it points to the firm’s house position on liability caps and the matter where that wording was last negotiated. Every flag and summary cites its source, so the lawyer verifies in seconds rather than taking the assistant’s word. Update the precedent and the assistant’s answers move with it.

It sits inside Microsoft 365, where the firm’s lawyers already work, so a review starts from a document in the system rather than a separate tool. The boundary is firm. The assistant retrieves only from approved firm material, client data is not used to train the model, and access follows the firm’s existing matter permissions, which keeps confidential and privileged material controlled. A lawyer reviews and signs off on every output. The assistant accelerates the first pass and surfaces clauses and precedents. It does not advise, and it does not replace the solicitor who does.

Building it, and where it got hard

We built it in phases. First a narrow scope on the firm’s highest-volume contract type, with a curated clause library and a clear set of checks. Then grounded retrieval against the precedent bank, then the Microsoft 365 integration so it met lawyers in their normal workflow, and only then a wider set of contract types. The model was rarely the hard part. The friction lived in trust, and one example stands in for the rest.

Early in testing the assistant confidently mis-summarised an unusual indemnity clause. The contract carried a non-standard carve-out, and the summary smoothed it into the ordinary version a reader expects, stating something the clause did not say. In most settings that is a minor slip. In legal work it is unacceptable, because a lawyer relying on a wrong summary of an indemnity is the exact failure the tool was meant to prevent. The instinct to write a cleverer prompt was the wrong instinct.

The fix was grounding and humility, not persuasion. We tied every claim the assistant makes to its source, with a pinpoint citation back to the clause it came from, so a summary cannot drift away from the words on the page. We tuned it to say “I cannot find this” rather than invent a tidy answer when the clause is unusual or the precedent is missing, because a flagged gap is useful and a confident invention is dangerous. And we made the lawyer review step mandatory, so nothing reaches a client without a qualified person checking it against the source. The unusual clause that started the problem now gets flagged as unusual, which is exactly what a junior should escalate.

What changed

In a representative deployment the assistant moved a first-pass review of a standard commercial contract from a matter of hours to a matter of minutes, with the lawyer then verifying the flagged clauses rather than reading the document cold. The consistency mattered as much as the speed. Because every contract was checked against the same clause library, the flags stopped depending on which junior happened to pick it up or how late in the week it landed. Each flag and summary pointed back to a specific clause and a precedent in the firm’s own library, so a lawyer could confirm the source in seconds instead of hunting for it.

These figures are illustrative. They describe the pattern we see rather than a published result for a named firm. The shape is the point. The repetitive first read comes off the junior’s desk, the clause checks become consistent, the firm’s own precedents arrive at the moment the contract is open, and the lawyer spends the recovered hours on advice and negotiation. The assistant reads. The lawyer decides.

Where this fits

A Claude legal assistant is one application of our AI Agents service, grounded with retrieval and sitting inside Microsoft 365, for the realities of an Australian professional services firm. It is a contained, high-value place to start, because the work is repetitive, the firm already owns the precedents that ground it, and a lawyer stays in the loop throughout. If first-pass contract review is eating your junior hours, the place to start is to map your highest-volume contract types and the clause checks that matter most, then decide where a lawyer must verify before anything reaches a client.