The Point of Utility: Why Legal AI Adoption is Determined by Data Availability, Not Intelligence

Much of the discussion around legal AI centres on one question:

When will AI replace lawyers?

I increasingly think this is the wrong question.

The more interesting question is:

When does AI become the default way of performing a legal task?

Those are not the same thing.

For much of the last two-three years, AI has largely been an enhancement to existing workflows. Lawyers research with ChatGPT instead of Google. They draft emails faster. They summarise documents. Useful improvements, certainly - but hardly a revolution.

Yet something has changed.

The latest generation of AI systems are becoming truly agentic. Rather than solving isolated problems, they are beginning to execute complete workflows. This has led many to believe that the final barrier to widespread automation is simply better models or more system integrations.

I think there is a more useful way of viewing it.

The Point of Utility

Every legal task has a threshold where a lawyer subconsciously decides:

"It is now easier to complete this task with AI than without it."

I call this the Point of Utility (POU).

Importantly, POU has nothing to do with perfection. AI does not need to outperform lawyers. It simply needs to reduce enough friction that using it is easier than reverting to legacy tools. Once a task crosses this threshold, adoption becomes voluntary rather than mandated. That distinction matters enormously.

Why Some Tasks Reach POU Before Others

M&A due diligence has arguably crossed POU.

Modern legal AI platforms can review thousands of documents, identify risk and generate structured reports to a standard that many lawyers now actively prefer.

By contrast, highly localised corporate restructuring advice often has not.

Not because today's models lack intelligence, but because they lack context. The difference is subtle but profound. The bottleneck is increasingly no longer the model. It is everything surrounding the model.

The Hidden Infrastructure Layer

This is where I believe the market is still undervaluing legal AI. Everyone is focused on frontier models. Far fewer people are focused on the infrastructure that determines whether those models are actually usable.

Examples include:

• firm-specific institutional knowledge

• structured precedents

• document connectivity

• workflow mapping

• metadata

• evaluation datasets

• historical drafting behaviour

• decision provenance

  
  

Individually, these appear mundane. Collectively, they determine whether a legal workflow reaches POU. In other words, they are not simply data assets.

They are utility assets.

A New Asset Class

Assuming this framework is correct, an interesting investment thesis follows.

As legal AI matures, the most valuable assets may not be better language models. Those will increasingly become commoditised.

Instead, value will accrue to the scarce assets that move legal workflows across the Point of Utility.

The firms and platforms that own structured institutional knowledge, workflow intelligence, evaluation infrastructure and contextual legal data will possess an increasingly valuable layer sitting beneath every AI system.

The market is still largely pricing these assets as operational overhead. I suspect they will eventually be recognised as strategic infrastructure.

What Comes Next

This raises a more interesting question than whether AI will replace lawyers.

Which assets most efficiently move legal work across the Point of Utility? The answer is unlikely to be another language model. It is more likely to be the infrastructure that enables those models to understand context, navigate workflows and earn lawyers' trust.

That is where I suspect much of the next decade's value will be created.