Contract review with AI: what it catches and what it misses

By Michael Nadalin, Founder, Market Lead · 18 September 2026 · 9 min read

A 62 page supply agreement arrives at 4pm with a note asking for comments by 9am. You have two other matters open and the client has already signed the heads of agreement. This is the exact moment people reach for an AI tool, and it is the exact moment the decision about what to trust it with actually matters.

The honest answer is not that AI contract review works or that it does not. It works very well on one category of clause and fails in a specific, predictable way on another. The failure is not a blank screen or an error message. It is a confident, well written paragraph that reads like a competent junior wrote it and is wrong in a way you will only catch if you read the clause yourself.

What follows is the split as it actually falls: the clause types where the machine earns its place in the workflow, the ones where it quietly creates risk, and the Australian layer that no general purpose model carries.

The clauses it reads well

The pattern is simple. AI is strong wherever the answer lives inside the text and requires comparison rather than judgment. Defined terms are the clearest case. A language model will find the term defined in clause 1.1, used in a different sense in clause 14, and never used at all after clause 3, faster and more completely than a human reading at 11pm. The same applies to cross-references, numbering breaks, and a schedule that contradicts the operative provision it is meant to support.

Mechanical commercial terms are the second strong category. Payment timing, invoicing triggers, notice periods, addresses for service, term length, renewal mechanics, termination triggers and their notice requirements. These are extraction tasks. Ask for them in a table across five supplier agreements and you will get in two minutes what used to take an afternoon.

The third is comparison against a known position. If your firm has a playbook, the tool is genuinely good at telling you which of your standard positions are missing or altered. That is a diff, not an opinion, and it is where the time saving is real.

The clauses it gets wrong while sounding right

Every clause that requires holding three provisions in mind at once is where the trouble starts. An indemnity means nothing in isolation. It means something in combination with the liability cap, the consequential loss exclusion, the insurance clause, the carve-outs from the cap, and whatever the parties agreed in a side letter that is not in the document you uploaded.

A language model will summarise the indemnity accurately. It will tell you the indemnity is broad. What it will not reliably do is notice that the cap is expressed to apply to all claims including indemnity claims, that the carve-out for third party IP claims sits outside the cap, and that the insurance clause requires a level of cover that does not match the uncapped exposure. That is the review. The summary is not the review.

The second failure mode matters more because it is invisible. Silence reads as a clean bill of health. When the tool returns nothing on a clause, that is not a finding that the clause is fine. It is the absence of a finding, and the two look identical on screen at 11pm.

Why risk allocation clauses break it

It helps to set out what a solicitor actually does when reviewing a liability regime, because the breakdown point becomes obvious once the steps are separate.

The tool performs the first two steps well and often better than a person working quickly. It collapses at step three and has no access at all to steps four and five. That is not a quality problem that a better model fixes next year. Steps four and five require the statute that overrides the drafting and the commercial context that was never in the document.

The Australian layer that is not in the text

A large share of what makes an Australian contract review useful is knowledge that sits outside the four corners of the document. General purpose models are heavily weighted toward American drafting, and the tells are consistent. Limitation clauses without the carve-out for rights that cannot lawfully be excluded. Liquidated damages provisions drafted against a penalties doctrine that is not ours. Pay when paid mechanics in a construction subcontract, which security of payment legislation does not allow to work the way the drafting suggests.

Then there is the category of clause that is perfectly drafted and still worthless. Retention of title is the clearest example. The clause can say everything it needs to say about title remaining with the supplier until payment. If the security interest was never registered on the Personal Property Securities Register, your client stands in the queue with the unsecured creditors when the customer goes into administration. No amount of reading the clause reveals that, because the problem is not in the clause.

Proportionate liability is the other one worth flagging. The regimes are not uniform across the states and territories, and whether parties can contract out of them is not uniform either. A clause that is effective under one governing law can be ineffective under another. The governing law clause is therefore not boilerplate, and a tool reading the document in isolation has no reason to treat it as anything else.

A workflow that keeps you defensible

The useful framing is not whether to use AI on contract review. It is which pass it belongs to. Use it to do the mechanical sweep before you read, so that your own reading starts from a map rather than page one. Never use it as the last thing that happens before advice goes out.

The practical test for any clause is this: could a careful reader answer the question using only the words on the page. If yes, the tool is useful. If the answer requires another document, a statute, or knowledge of what the client is actually trying to achieve, you are the only one in the room who can answer it.

What to say when someone asks if you used AI

Say what you did. The defensible position is not that no software touched the file. It is that a solicitor read the clauses that carry the risk, checked the law that applies to them, and can point to the source for every proposition in the advice. That position is unaffected by whether a tool produced the first draft of the defined terms table.

Several Australian courts have now issued practice notes and guidelines on generative AI use, and the direction of travel is consistent: disclosure where it matters, verification always, and no reliance on unverified output. The same standard makes sense in transactional work even where no court is involved, because the professional obligation it reflects was already there.

The practical requirement is that the tool lets you check it. Any output that asserts a clause says something should let you click through to the clause, and any output that asserts the law says something should let you click through to the legislation or the judgment. That verifiability is the design principle behind Legal Brain, and it is worth insisting on from any tool you put in front of a client matter. If you cannot get from the answer back to the source in one step, you are not reviewing the contract. You are reading a summary of it and hoping.

Frequently asked questions

Can I use AI to review contracts in Australia?

Yes, and the professional obligations are the same as they have always been. You remain responsible for the advice regardless of what produced the first draft, so the tool sits in the first pass rather than the last. The practical limits are confidentiality of client material and your duty to verify anything stated as law or as the content of a clause.

Which contract clauses should I not rely on AI to review?

Anything that allocates risk across multiple provisions: indemnities, liability caps, carve-outs from those caps, consequential loss exclusions and insurance obligations. These require reading several clauses against each other and against the statute that may override them. A model will summarise each clause accurately and still miss the interaction that matters.

Is it safe to upload a client contract to a general AI chatbot?

Check the terms before you do it. The questions are whether the document is retained, whether it is used to train the model, and where it is stored. Many consumer grade tools answer those questions in ways that are difficult to reconcile with your confidentiality obligations, which is the reason legal specific tools exist.

Do I have to tell my client I used AI to review their contract?

There is no single national rule, so check your jurisdiction's professional conduct guidance and your own engagement terms. The safer practice is to address AI use in your costs agreement up front rather than deal with it after the fact. Where a court is involved, follow the relevant practice note, because several Australian courts have specific requirements.