A lot has happened since (at times the news cycle feels hourly), so we thought it was worth reflecting on what has changed. And, more importantly, what those changes mean for law firms making technology decisions today.
Three developments have reshaped the strategic landscape:
So, what does this mean in practice? Let’s take each in turn.
Client use of AI has accelerated dramatically. FTI’s most recent General Counsel Report found that 87% of in-house legal teams now use GenAI (nearly double the figure of a year earlier). Strikingly, that adoption has in many cases outpaced their external advisers'.
The practical effect is felt across the whole market. Whether you are advising an individual on their estate planning or a multinational company on a complex merger analysis, increasingly, clients arrive not with an instruction but with something already in hand. “I’ve put this draft together - can you review it?” or “We’ve run this analysis, what do you think?”
On the surface, that sounds like less work for the law firm, and that’s what the client expects. In practice, any lawyer with supervisory responsibility knows the opposite is often true. Reviewing a document of unknown provenance - where you don’t know what has been assumed, what has been omitted, or how it was produced – can be far more demanding than doing the work yourself. Quite often, you are better off starting again.
In the short term, it puts a premium on being able to articulate the value you add. If a client can generate a plausible looking first draft themselves, the differentiator is no longer the production of text. It is the judgement, supervision and accountability wrapped around it. Firms will increasingly need to communicate that value to clients explicitly.
In the longer term, we expect this to settle down. As both clients and firms grow more comfortable, we think GenAI will be used in more measured ways. For example, rather than using it to produce the whole end product, deploying it to handle the parts it is well suited to: marshalling facts, organising information and surfacing the right starting point.
Which brings us back to Principle 2 from our 10 Principles: GenAI is a tool, not a product. The value isn’t in the model; it is in deploying it thoughtfully, end-to-end, in ways that solve a real problem.
The second big shift is where AI can now operate.
Word is where most legal work happens and, until recently, the major AI platforms didn’t perform well there. They were clunky, they mangled formatting, and while they were fine for rewriting a paragraph of website copy, they were generally not credible for making tracked-change amendments to a clause.
That has changed in recent months. The launch of legal-specific plugins in enterprise tools like Microsoft Copilot and Claude has moved the dial. It is now plausible to ask Copilot to insert a party’s details into a template document, or to ask Claude to apply a formatting change to a clause to bring it in line with your house style.
This is especially powerful for review and negotiation workflows. If you have a playbook position, you can use these tools to evaluate an incoming draft quickly and produce a first-pass mark-up to work from.
But this comes with an important caveat: using AI to draft puts the lawyer in exactly the position we described above. The output has to be reviewed. We have started to think of this as the “AI review tax”. Every AI-assisted task carries a review overhead, and the value the tool delivers has to comfortably exceed that tax. Otherwise you have simply moved the work (often, ironically, from junior to senior resource), not reduced it.
We think AI in Word will become a fixture - this is not a passing capability. But AI punishes laziness, and the dynamic to watch closely is training and supervision. Two things follow.
First, favour tools that are transparent and easy to supervise; tools that make it easy for the user to do the right thing, and that make clear where content has come from (see Principle 7). The harder it is to see what the AI did, the heavier the review tax.
Second, invest in training. A couple of maxims we have found useful:
Neither is complicated, but together they make the difference between AI that sharpens a lawyer’s judgement and AI that quietly erodes it.
The third development is about using AI to build tech.
“Vibecoding” is the practice of building software by describing what you want in natural language and letting an AI generate the code - iterating by feel, rather than writing much (or any) of it yourself. It has made building software dramatically more accessible to people who would never previously have called themselves developers.
Fuelled by platforms like Lovable, OpenAI’s Codex and Claude Code, vibecoding has taken off. We helped judge the LMA Edge Hackathon earlier in the year, and it was striking how much teams could prototype in a short space of time. More broadly, we’re starting to see ambitious attempts to replicate the kind of functionality you would associate with established SaaS products (see, for example, MikeOSS).
The catch is the familiar engineering lament: “it works on my machine”. Prototyping something impressive is one thing; productionising it (making it secure, reliable, maintainable, and safe to put in front of clients) is quite another. And handing a vibecoded app to a software developer is much like a lay client asking a lawyer to review their AI-generated legal advice.
For that reason, we suspect the durable impact of vibecoding will look less like a wholesale revolution and more like the “low-code / no-code” transformation that preceded it. Flexible platforms for building tools that come with sensible guardrails (like Airtable or Power Automate) will help teams solve their own problems. Existing SaaS products will extend their capability by building integrations for these platforms. Rather than disappearing, they will become tools that are used by both AI and human users entirely.
What is our role in the evolving landscape?
We remain genuinely bullish about AI - that hasn’t changed, and our development continues at pace. But we are equally clear-eyed about what it takes to make AI deliver a return across an entire workflow, not just in a demo. The AI review tax is real, and the only way to come out ahead is to build on solid foundations.
For us, those foundations are dependable, rules-based automations and trusted templates. This is why Principle 1 remains our north star: precedents crafted by experienced knowledge lawyers are still the gold standard. AI’s role is to augment that. It should help with the truly novel, deal-specific drafting not contemplated in precedents, rather than reinvent templates.
That distinction is precisely what shapes Clarilis AI Draft. Our automations have long enabled lawyers to produce a first draft that is 90% complete in minutes. But no automation, however thorough, can account for every nuance, exception, or deal-specific scenario. AI Draft is built to close that final 10% gap: whether you need bespoke rent-review mechanics for a new lease, or a complex earn-out schedule for an SPA, AI Draft generates review-ready drafting tailored to your practice area and reflecting the deep experience of our knowledge lawyers.
Critically, AI-generated content is visually distinguished from automated content within the document, so you always know exactly what you are reviewing and why. The provenance is built into the process, not bolted on as an afterthought.
It is also what distinguishes our approach from horizontal AI assistants and general-purpose drafting tools. A chat interface bolted onto a model can produce something that looks like a draft. What it cannot do on its own is guarantee where the content came from, or make the output easy to supervise. Those are precisely the things that determine whether AI saves a firm time or simply adds to the review pile.
We’d love to discuss how this thinking applies to your firm, and how Clarilis AI Draft fits alongside your wider AI strategy. Click here to learn more.