If you work anywhere near employment law, human resources, or professional services in New Zealand, a headline is going to make the rounds in the next few weeks, either by landing in your inbox or popping up in your LinkedIn feed. It will tell you that most legal professionals now think the cost of checking AI outputs has overtaken the benefit of using AI at all.
It will be shared by someone who has been waiting three years to say I told you so. I have sympathy for the impulse. There is satisfaction in watching an over-promised technology come back down to earth, and if you have spent a career being told your judgement was about to be automated into obsolescence, you have earned a small moment of vindication.
The problem is that the headline does not describe the study it claims to report.
What the research actually found
The underlying work is the Morae AI in Legal Report 2026, published on 8 July this year under the title The Intelligence Gap: Why AI in Legal Isn't Delivering and What Needs to Change. It surveyed 850 senior legal professionals across the United States, the United Kingdom, Australia, and the Middle East, split evenly between in-house legal departments and private practice. Fieldwork was carried out by Coleman Parkes in February and March 2026.
The numbers are not comfortable reading. Only 33 per cent of respondents said they trust the results of AI-assisted legal work. Some 48 per cent named poor accuracy and hallucinations as the most significant constraint on effective use. Half reported high concern about liability if AI-assisted work produces errors, and 62 per cent were extremely or very concerned about the reputational damage that would follow.
And yes, 67 per cent were concerned that the cost of human verification and oversight might outweigh the efficiency benefits.
That last sentence is where the trouble starts.
The first problem is a change of tense
Morae's own account says that 67 per cent are concerned that the cost of verification might outweigh the benefits. It is a worry about a possibility.
By the time it reached the trade press, that concern had become a finding that respondents believe “the burden of human verification outweighs the benefits AI was supposed to deliver” (Global Legal Post, 14 July 2026).
A concern has been converted into a conclusion. An anxiety about the future has been reported as a measurement of the present. This is not fabrication, and I doubt anyone did it deliberately. It is the ordinary compression that happens when a survey of 850 people is reduced to a headline of nine words. But the compression removed exactly the qualifier that was doing the work.
Nobody in that survey measured anything. They were asked how they feel about a cost they do not track, weighed against a benefit they do not track. That is still a legitimate and useful thing to ask senior people. It is not evidence that AI has stopped paying for itself.
The second problem is the omission
Here is the part that ought to be on the front of every summary of this report, and is on almost none of them.
Morae's own conclusion is that the problem is not the technology.
The media release carries it in the subheading: the challenges stem from flawed implementation, not from the tools. The report page says the fix lies in implementation, data, and governance rather than in better software. Morae's chairman and chief executive, Shahzad Bashir, puts it in seven words: “The root of the problem isn't the technology itself.”
The study that is about to be forwarded to you as proof that AI does not work is a study concluding that AI is being deployed badly by organisations that have not put their own house in order first. Those are not the same claim. They are close to opposite claims.
And the report's own data supports its authors, not its headline:
- 80 per cent agree that effective AI use depends on high-quality, well-governed data.
- 77 per cent say poor information quality undermines outcomes regardless of how good the tool is.
- Only 26 per cent of legal leaders are confident their organisation's information is ready for AI to work with.
- Nearly one in three organisations has no formal governance framework for AI-generated outputs at all.
- 71 per cent have formal records retention and governance policies — but only 30 per cent are confident those policies are actually applied.
- Only 30 per cent are confident their information is properly classified: 32 per cent for privileged material, 27 per cent for confidential, 26 per cent for sensitive.
Read that list again without the words “AI” anywhere near it. Fragmented systems. No reliable file inventory. Policies on paper that nobody follows. Material that has never been correctly classified as privileged or confidential.
Those are records management failures. They were failures in 2015. They will be failures in 2035. What generative AI has done is switch on a light in a room nobody had looked at properly for twenty years, and the room's condition is now blamed on the light.
What the verification tax actually is
This is the whole argument. The survey asks whether verification costs more than AI saves. It does not ask what verification cost before.
And verification always cost something. Reading a junior's memorandum before it went out. Checking that the precedent pulled off the shared drive was not the 2019 version. Sitting with someone less experienced to explain, again, why the authority they had found did not say what they thought it said. Whether you did that work yourself or someone above you did it, the answer is the same. Quality control has never been free.
It simply never appeared anywhere. It was absorbed into supervision, training, hierarchy, and the general overhead of running a professional practice. It had no line item, so it had no visible cost, so it felt like nothing at all.
What 67 per cent of the respondents are describing is not a new expense. It is the first time in their careers that quality control has arrived with a price tag attached.
That is uncomfortable. It is not the same thing as being new.
And the same survey settles the point in the respondents' own voices: 89 per cent of them said AI-generated legal work should be checked by a human before it is used. Eighty-nine. The same people, in the same questionnaire, who said verification might cost more than it saves also said it cannot be skipped. The position of this profession is that verification is expensive and that verification is non-negotiable.
That is not a profession giving up on AI. That is a profession discovering that it has to build, deliberately and at cost, a supervisory function it used to get for nothing.