Most law firms I speak to are in the same position. They know AI is happening. A few people in the practice have tried ChatGPT. Maybe someone did a half-day workshop last year. But nothing has stuck, nothing has changed how the firm actually operates, and the gap between “we should do something about AI” and “we have done something about AI” feels wider than ever.
Meanwhile, something is happening in the background that most UK firms haven’t fully registered. A small number of practices, mostly in the US, but increasingly in the UK, have stopped treating AI as a novelty and started treating it as core infrastructure. Not a bolt-on tool. Not a chatbot in the corner. A fundamental part of how the firm produces work.
The results are striking enough that they deserve attention. Not because every firm should rush to adopt AI tomorrow, but because the economics of legal practice are being quietly rewritten, and the firms that understand what’s actually possible will make better decisions about when and how to act.
The Shift Nobody Expected
The conventional wisdom about AI in legal was always that it would be a tool for large firms with large budgets. Enterprise software, six-figure contracts, innovation committees, and pilot programmes. The big firms would adopt first, and smaller firms would follow once the technology became cheaper.
That is not what happened.
What actually happened is that the most advanced AI implementations in legal are emerging from small practices. Two-person firms. Boutiques. Practitioners who decided to work directly with the underlying AI models rather than waiting for a legal-specific product to wrap them in a polished interface.
One example has become the reference case for the entire profession. Zack Shapiro, a Yale Law graduate who clerked in the federal courts and trained at Davis Polk, built Raines LLP as a two-person practice. The firm represents over 200 startups and investors across corporate, venture, and regulatory work. The article he wrote about how the firm operates on Claude AI was read over 7.5 million times, including by an estimated 1.5 million lawyers.
The reason it resonated so deeply is not because it described some futuristic vision. It described something that is possible right now, today, with tools that already exist. And the implications for how legal practices operate are profound.
What “Running on Claude” Actually Means
When people hear “AI in a law firm,” they tend to imagine one of two things: either a chatbot that lawyers type questions into, or some kind of automated document assembly tool. The reality of a properly configured Claude implementation is neither. It is closer to having an extremely capable junior colleague who has read every document your firm has ever produced, understands your house style, knows your templates, and can work at extraordinary speed, but who always needs your judgment on the decisions that matter.
The architecture works in layers. Let me walk through what each one looks like in practice.
Layer 1: Personalisation
Every user in a Claude Team Plan gets their own configuration. Not a generic account. A setup that reflects how they specifically work. Their communication style. Their preferred document formats. Their tone of voice. The way they structure advice letters versus the way they structure board minutes. Claude learns these preferences and applies them consistently across everything it produces for that person.
This matters more than it sounds. One of the main reasons lawyers abandon AI tools is that the outputs feel generic. They read like they were written by a different firm. When the AI is configured to your personal style, the gap between “AI-generated draft” and “my first draft” shrinks dramatically. You spend less time rewriting and more time refining.
Layer 2: Shared Projects
A Project in Claude is a persistent workspace loaded with your firm’s content. Think of it as giving Claude access to the institutional knowledge that currently lives in people’s heads or scattered across SharePoint folders.
A litigation Project might contain your standard pleading templates, your preferred authorities for common arguments, your house style guide, and examples of successful submissions. A corporate Project might contain your precedent library of shareholder agreements, your standard due diligence checklist, and your preferred NVCA document set.
When a lawyer works within a Project, Claude draws on all of that context automatically. The output reflects your firm’s standards, not generic AI defaults. This is the difference between asking a stranger to draft something and asking a colleague who has worked at the firm for five years.
Layer 3: Skills
This is where things get genuinely powerful. A Skill is a reusable instruction set that encodes a specific workflow. It tells Claude not just what to do, but how you want it done: your analytical framework, your preferred format, your quality standards, your voice.
Example: Contract Review Skill
A counterparty sends back a redlined agreement. Forty pages of changes across representations, indemnification, IP, and closing conditions. You upload the document and invoke your contract review Skill.
Claude organises every change by severity. It flags where the counterparty has shifted risk. It identifies tensions between modified provisions. It checks for standard provisions that should be present but aren’t. It produces a structured summary with specific counter-language for each material issue.
The cross-document analysis that would have taken an associate the better part of a day happens in minutes. The senior lawyer then applies judgment: which fights matter given this particular deal, which concessions are strategic, which issues are posturing. The AI handles the production. The lawyer handles the decisions.
Example: Redlining Skill
Claude doesn’t just analyse documents. It can produce tracked changes. It opens Word documents at the XML level, applies changes attributed to the lawyer’s name, preserves every formatting detail, and outputs a clean .docx with real tracked changes that opposing counsel can open in Microsoft Word and review normally.
From receiving a markup to having a response package ready to send: under an hour, of which roughly 30 minutes is the lawyer’s own thinking time.
Example: Client Advice Skill
A client asks a question that requires research across multiple areas. The lawyer activates a research Skill that structures the analysis: identify the relevant statutory framework, map the case law, note any recent developments, and produce a draft advice letter in the firm’s standard format with the lawyer’s personal tone. The letter goes out looking like it took two days. It took two hours, and most of that was the lawyer thinking about the answer rather than producing the document.
The key insight from practitioners who have built these workflows is that the Skills encode judgment, not just process. Each Skill is refined across dozens of matters until it captures not just “how to review a contract” but “how this lawyer reviews a contract, with their analytical framework, their quality standards, and their preferred approach to flagging risk.” That accumulated judgment compounds. The fiftieth matter starts from a higher baseline than the first.
Layer 4: M365 Integration
Claude Team Plan connects natively to Microsoft 365: Outlook, SharePoint, OneDrive, Teams. This means Claude can read your existing content, search your document estate, summarise email threads, and work with the files that are already in your systems. You don’t need to copy and paste content into a separate tool. Claude operates within your existing infrastructure.
For a firm already running on Microsoft, this eliminates the most common barrier to adoption: the friction of switching between tools.
Enjoyed this? Join the newsletter.
One email a week. What I'm building, learning, and what's actually working. No fluff.
Free. Unsubscribe anytime.
What This Means for Regional UK Firms
The examples above come largely from the US market, where adoption is further ahead. But the dynamics apply directly to UK regional practices, and in some ways the opportunity is even more significant.
Large City firms have innovation budgets, technology partners, and dedicated teams evaluating AI products. Regional firms typically do not. This has created an assumption that regional firms will be last to adopt. But the new generation of AI tools, Claude in particular, has inverted that logic. The setup cost is negligible. A Claude Team Plan costs $20 per user per month. The implementation does not require an IT department or a six-figure software contract. It requires someone who understands the firm’s workflows and can configure the tool properly around how the practice actually operates.
The firms I’m speaking to in the UK are all saying the same thing: we know we need to do something, there are too many options, we don’t know where to start. That uncertainty is not a sign of being behind. It’s a sign of being thoughtful. The firms that rush to adopt the first legal AI product they see will likely end up with an expensive tool that nobody uses. The firms that take the time to understand what proper implementation looks like, and then move decisively, will have a genuine advantage.
The Economics
Let’s talk about what this actually does to the numbers.
The cost structure is almost trivially small. Seven users on a Claude Team Plan costs roughly £1,330 per year. That is less than the monthly cost of a junior paralegal. The return on that investment depends entirely on how well the tool is implemented.
A properly configured Claude setup does not replace lawyers. It changes what lawyers spend their time on. The production work, first drafts, document review, correspondence, reporting, accelerates dramatically. The judgment work, advising clients, making strategic decisions, building relationships, gets more time allocated to it because the production work is no longer consuming the day.
For a firm billing on time, this creates a philosophical question about value pricing that is beyond the scope of this article. But for a firm that is capacity-constrained, too much work, not enough people, the answer is simpler: the same team can handle significantly more matters without adding headcount, and the quality of the output improves because senior lawyers spend less time on production and more time on the work that actually benefits from their experience.
The Adoption Gap
The reason I’m writing this is not to sell AI. It’s because there is a structural gap between what is possible and what most firms are doing, and that gap is growing.
Anthropic, the company behind Claude, now holds 73% of the enterprise AI market, up from 40% just three months ago. Their enterprise platform is purpose-built for organisations that need accuracy, data privacy, and workflow integration. They do not train on your data. They offer SSO, centralised admin controls, and usage analytics. The security and governance questions that legitimately concerned firms two years ago have largely been resolved.
Meanwhile, venture-backed companies are pouring capital into this space. The infrastructure for AI adoption in legal is being built rapidly, with specialist firms raising significant funding to deploy AI adoption programmes into law firms. This is no longer a niche experiment. It is becoming a market.
The barrier is no longer technology, cost, or security. It is implementation. The firms that have someone to properly configure the tool, load the right content, build the right Skills, train the team, and sustain adoption, get transformative results. The firms that buy seats and hope for the best get expensive novelty that fades within 90 days.
What Proper Implementation Looks Like
If you’re a managing partner or practice head reading this and thinking “we should probably do something,” here is what “doing something properly” actually involves:
First, understand the workflows. Before anyone touches a tool, someone needs to map the firm’s actual workflows. Not the idealised version, the real one. What do your fee earners produce most often? What takes the most time? Where does quality vary? What would change if the first draft of anything took 15 minutes instead of half a day?
Second, configure the tool around the firm. A generic Claude account is useful. A Claude account loaded with your precedents, your house style, your templates, your brand guidelines, and your preferred analytical frameworks is transformative. The configuration is what turns a general-purpose AI into something that feels like it works at your firm.
Third, build Skills for the work you do most. Identify the five or ten most common tasks across the practice. Build a Skill for each one. Refine them over time. These become the firm’s reusable AI infrastructure: a library of encoded expertise that compounds with every matter.
Fourth, train the people, not just the tool. The biggest determinant of whether AI adoption sticks is not the technology. It is whether the people in the firm change their habits. This means hands-on coaching, working through real tasks together, and making sure every person has a setup they will actually use tomorrow. A group training session is a starting point. Individual configuration by function is where the value lives.
Fifth, sustain it. The firms that get the most value from AI are the ones that keep building. New use cases emerge every week. New Skills get created as the team’s confidence grows. An internal champion who understands how to maintain and expand the setup is essential for long-term adoption.
The Question That Matters
The question facing regional UK law firms right now is not “should we adopt AI?” That question has been answered by the market. The question is whether you implement it properly, configured around how your firm actually works, with the depth of setup that produces real results, or whether you buy seats, attend a workshop, and end up in the same place six months from now.
I work with professional services firms to do the first version. A focused enablement sprint that configures Claude around your practice, builds the Skills your team needs, loads your content, and trains your people so that by the end of the engagement, every person has a working system they will actually use.
Not a product. Not a workshop. Not a slide deck about what AI could theoretically do. A working implementation, configured to your firm, ready to use.