Ghost in the machine

AI is moving out of experimentation and into the day-to-day machinery of legal practice. Ger Perdisatt looks at what firms are finding and why the next constraint may have surprisingly little to do with AI models themselves

The new High Court Practice Direction HC142 on the responsible use of generative AI in court documents came into effect on 1 September. Its starting point is worth noting: the court accepts that AI may improve efficiency, reduce costs, and widen access to justice, but its use changes none of the obligations owed to the court. 

Whoever puts the document in remains responsible for it, whether or not a machine helped to write it. 

That puts the most important piece on the table early. Human judgement stays primary. AI can inform, assist, draft, or automate around it, but professional responsibility doesn’t migrate to the machine. 

So, once you accept that, a wider question opens up: where can AI, and technology more generally, actually improve how a practice works? 

A year ago, most conversations I had with firms started in the same place: what can it do? Now it’s more likely to be: why is it really good at one task and laughably poor at the next? Why can one team find immediate value while another can’t see a worthwhile use for it at all? 

The instinct is to blame the technology. Better model. Better product. More AI. That isn’t what we’re seeing. The questions have changed.

Deus ex machina 

When we developed the AI resources for the Law Society’s Practice Essentials earlier this year, the pace of change was already obvious. Any attempt to prescribe the right product, or a fixed set of use-cases, was going to age like milk. 

The Society was prescient in making Practice Essentials practical rather than predictive: resources to help firms make good decisions as the technology moves, rather than pretending to know what the market will look like in a year (nobody does). 

The AI resources cover safe-use patterns, shadow-AI discovery, vendor assessment, and workflow readiness – and the toolkit is built to keep evolving. 

Several months on, that approach looks better rather than worse. Models have improved, products have multiplied, AI is turning up inside software that firms already own, and users – solicitors, legal executives, administrators, office managers and others – have gotten considerably more capable. 

But adoption isn’t moving in a neat line from curious to mature. In the same firm, one process may already be heavily automated while another practice area is only starting to map what’s possible.

A litigation team might have work suited to AI-assisted drafting, while the accounts team can strip hours out of billing through automation that barely involves generative AI at all. One firm we work with has practice areas deliberately moving at different speeds on a shared technical and governance foundation. 

So the useful unit of analysis isn’t really the firm in the aggregate. It’s the specific work to be done – department by department. Is this actually an AI problem?

Cogito, ergo sum 

Once firms start implementing rather than demonstrating, they find that a lot of what gets called an ‘AI use-case’ is several distinct problems bundled together.

Take a typical request: can we use AI to automate this process? Maybe. AI might be right for reading an incoming document, classifying it, comparing versions, or preparing a draft.

But if the next step is to calculate a known figure, populate a field, assemble a standard document, or move information reliably between systems, conventional software is often simpler and safer (and you’ve probably already bought it).

There then is the final step where a solicitor has to exercise judgement. 

Those are three different kinds of work: (i) tasks that can run deterministically; (ii) tasks where AI assists but a person approves; and (iii) tasks where the substantive decision stays human.

One of the clearest signs that a firm is getting better at AI is that it stops trying to use AI for everything. 

Most of the public energy around legal AI goes towards the legal work itself, understandably. The big legal-tech AI platforms are showing sophisticated capability in research, contract review, document analysis and drafting.

But a practice does a lot more than research an issue or draft a clause. Files have to be opened. Information gets extracted, checked, and re-entered. Documents are compared and assembled. Reports, schedules, and bills are produced. Deadlines are chased. Meeting participants is organised.

The same information gets moved between systems that were never designed to talk to each other. And this work sits across solicitors, legal executives, paralegals, secretaries, administrators, finance teams, and office managers.

There’s a lot of low-hanging fruit in that machinery. It’s less exciting than asking an AI to draft an ingenious argument – but saving ten minutes out of a process that runs hundreds of times a month lands on a practice much faster.

Which is why the requests we hear now are so specific: ‘Open the matter and put the information in the right places’; ‘Prepare a chronology’; ‘Compare these versions’; ‘Pull together the billing information’; ‘Find the previous work’; ‘Help someone new to the area understand how similar matters were handled’.

Those are far better approaches than answering the question: ‘What can AI do?’

Most don’t need some future super-intelligent model. Today’s AI technology is already capable of most of what’s on that list. What it really lacks is reliable access to the context, systems, and accumulated know-how needed to do the job properly.

Ex nihilo nihil fit

This is where AI is exposing a much older problem. Firms aren’t short of information. They have document-management systems, email, case-management and billing systems, shared drives, and years of closed matters.

But information isn’t knowledge. A file may hold the final advice without recording why that route was chosen. A precedent bank may hold five versions of a clause without noting that one came from the other side, another was an exceptional compromise, and a third is what the practice group actually uses now.

Firms have always filled those knowledge gaps with people. The partner remembers why a position was taken. The legal executive knows which template everyone really uses. The accounts team understands an unusual billing arrangement. The administrator knows why a step that looks redundant isn’t.

That’s real organisational knowledge. The problem is that its meaning, authority, and context are rarely captured in a form that technology can reliably use.

AI didn’t create that problem. It made the cost of it visible. A better AI model won’t fix that.

There’s a major model release every few weeks, and the improvements are real. But suppose tomorrow’s AI brain is twice as capable. Will it know which precedent represents the firm’s current position? Why an unusual concession was made? That the document marked ‘final’ was superseded after a phone call nobody wrote down?

No. A better brain does more with the context available to it. It can’t reliably conjure context that was never captured. That’s why some of the more interesting AI work now starts a long way from drafting.

It starts with how the work happens, which systems hold the relevant information, what’s being repeated needlessly, what knowledge is missing, and which sources should be trusted.

On another engagement, we’ve turned decades of an organisation’s own material into a connected, searchable system that runs privately.

The AI brain is deliberately replaceable; as better ones arrive, it gets swapped. The durable asset is the organisation’s own knowledge, made usable by whatever comes next.

Acta, non verba

Diagnose before you prescribe. That’s also the logic behind Practice Essentials. The value of a workflow-readiness diagnostic isn’t that it tells a firm which product to buy. It forces a more basic question first: ‘Is this piece of work actually ready?’

The same goes for understanding what’s already in use, where it’s appropriate, and what to examine before buying anything else. Those are reasonably stable questions in a very unstable market.

When something doesn’t work, the diagnosis might be the AI. But the issue may be that nobody agrees how the process should run. The relevant knowledge is scattered across closed matters and people’s heads.

A legacy system seems to offer no sensible way for modern technology to interact with it. On one current engagement, that means building around closed systems where value can be delivered now, while parking other work until a new core system is in.

Or someone may simply have seen an AI problem where a small piece of conventional automation would do the job better. The answer follows from the diagnosis – not the other way round.

Equo ne credite

The first phase of generative AI was mostly about handing people remarkable new capability and encouraging them to poke at it. That was necessary. Curiosity builds familiarity (and, in a few cases, familiarity has bred contempt).

The next phase is harder. It’s about turning the occasional individual win into something a practice can use deliberately, across legal and operational work, without confusing automation with AI, or AI with professional judgement.

A more capable practice might end up using AI less in some processes because it has learned where straightforward automation is safer and simpler.

It might get more out of improving file opening, billing, or matter administration than another drafting tool. It might get better results from a smaller, cheaper model because its own knowledge is better organised.

And it will leave certain decisions with people because, as HC142 makes explicit for court documents, responsibility doesn’t disappear just because a machine contributed.

Alea iacta est

The models will keep changing, faster than any sensible practice can track, release by release. That’s exactly why Practice Essentials has to keep evolving alongside them.

The more durable question is this: are we becoming the kind of practice that can take advantage of whatever comes next?

That means understanding the work first: where technology helps, where certainty matters, what the firm needs to make accessible, and where human judgement belongs.

That’s the journey that Practice Essentials is built to support: not just from curiosity to governance, but also from curiosity to capability.

Ger Perdissat is CEO and founder of AI Acuity.

Explore the full suite of Practice Essentials on the Law Society website. For queries, suggestions or feedback, contact the Solicitor Services Department at solicitorservices@lawsociety.ie.