The verification burden: when checking AI costs more than doing it

By Michael Nadalin, Founder, Market Lead · 17 August 2026 · 7 min read

A four page memo lands on your desk at 4pm. It is well structured, it reads with confidence, and it cites six authorities for a proposition you were not certain about that morning. You have two options. Send it, or check it. Checking it means pulling every judgment, reading the paragraphs the memo leans on, and satisfying yourself that the proposition holds. By the time you finish, it is 6pm and you have effectively done the research yourself.

That gap is the verification burden, and it is the part of the AI conversation that most vendor demos skip. The demo shows the draft appearing in eleven seconds. It does not show the ninety minutes on the other side, because the ninety minutes is where the value either survives or does not.

This is not an argument that AI has no place in an Australian practice. It is an argument that the honest unit of measurement is draft time plus verification time, not draft time. Once you measure that way, some tasks are obviously worth automating and some are obviously not, and the line between them is sharper than the hype suggests.

Where the verification time actually goes

Practitioners who have run this properly describe the same arc. The first few minutes feel like a win. Then the check begins, and the check is not a skim. It is a reconstruction of reasoning you did not do, from sources you did not choose, following a trail you cannot see.

The critical moment usually arrives around the halfway mark, when you realise the fastest way to be confident is to search the point yourself. At that stage the draft has contributed a structure and a starting vocabulary. Useful, but not ninety minutes of useful.

Why checking is harder than writing

There is an asymmetry at the heart of this. Producing fluent, plausible legal prose is cheap. Establishing that fluent, plausible legal prose is correct is expensive, and the cost does not fall as the output improves. It rises, because better output is harder to doubt.

The failure mode everyone talks about, a case that does not exist, is actually the easy one. It dies in a search. The expensive failure is a real Australian judgment, correctly cited, quoted accurately, and applied to a proposition it does not support. Nothing on the face of the document flags it. Only reading the judgment does.

Not every task carries the same verification cost

The burden is not uniform, and treating it as uniform is why firms end up with a blanket ban or a blanket permission, both of which are wrong. What drives the cost is whether you can bound the check.

If the source material is closed and sitting in front of you, verification is bounded and cheap. If the task requires the model to have searched the whole of Australian law and told you what matters, verification is unbounded, because you are being asked to check for absences. You cannot spot a controlling authority that was never mentioned by reading the document that failed to mention it.

The only test that matters: can you check it faster than you can do it?

Before using AI on a task, run it through a short sequence. Not a policy document, a habit. It takes about fifteen seconds and it will keep you away from the tasks where the maths does not work.

The last question is the one firms skip. If a partner has to verify a junior's AI assisted memo at partner rates, the firm has moved work up the chargeout scale, not down it. That is a worse outcome than having the junior do the research the long way, and it shows up in write offs rather than in any AI report.

A verification protocol a firm can actually run

Most firm AI policies are one page of prohibitions that nobody reads twice. What works better is a short protocol attached to the workflow, because the risk is not that people use AI, it is that they use it and cannot later demonstrate what they checked.

Courts here have already moved on this. The NSW Supreme Court and the Supreme Court of Victoria have issued practice guidance on generative AI use in litigation, and the direction of travel is clear across other Australian jurisdictions. Your duty to the court is personal. It does not transfer to a vendor, and it is not discharged by a disclaimer in a terms of service document.

What to require of a tool before it earns a place

If verification is the real cost, the useful question about any legal AI product is not how good its writing is. It is how much of your checking it removes. Those are different things, and a tool can be excellent at the first while doing nothing for the second.

Three requirements do most of the work. First, every proposition must link to the actual source text, not to a citation string, so you can read the paragraph in one click rather than starting a fresh search. Second, the tool must draw from a defined, current corpus of Australian legislation and case law, and tell you what is in it, because a system that cannot state its boundaries cannot tell you what it missed. Third, it must say when it does not know. A tool that answers everything confidently has handed the entire verification burden back to you and charged you for the privilege.

That is the standard Legal Brain is built against, and it is the standard worth applying to every product in this category, ours included. Ask any vendor to show you the checking pass, not the drafting demo. If they cannot, you now know where the ninety minutes will land.

Frequently asked questions

Does AI actually save time on legal research?

It depends entirely on the task. On bounded work such as summarising a document you already hold or extracting details from a known bundle, the saving is real because the check is quick. On open research into an unsettled point, verification usually costs as much as doing the research yourself, because you cannot detect what the model left out without searching the field independently.

Is checking that the cited cases exist enough?

No. A citation check only tells you the judgment is real. The more common and more dangerous error is a genuine Australian case, cited accurately, applied to a proposition it does not support. Catching that requires opening the judgment and reading the reasoning around the passage relied on.

Can Australian solicitors use generative AI in court work?

Use is not prohibited outright, but several Australian courts including the Supreme Court of NSW and the Supreme Court of Victoria have issued practice guidance restricting how generative AI may be used in preparing material for proceedings. Check the direction applying to your jurisdiction and matter before you start, and remember your duty to the court sits with you personally regardless of what tool produced the text.

How should a firm measure whether an AI tool is worth it?

Measure draft time plus verification time against the time to do the task without the tool, and record who did the verifying. If a partner is checking work a junior would otherwise have completed, the firm has shifted the task up the chargeout scale rather than saving anything. Track it for a month on real matters, not on a demo.