Hey, happy Sunday.
This week:
Law firms are starting to put real numbers behind their AI budgets
Legal AI launches, funding and hiring continue to accelerate
Claude’s new watermark could change how AI-generated content is identified
PARTNER CONTENT
Granola is an AI notepad that turns your meetings into clear, searchable notes without adding another bot to the call. For lawyers, that means less time writing up calls and more confidence that key client instructions, actions and decisions aren’t lost.
Try one month of Granola for free.
There is no clear consensus as to how much a law firm should allocate to AI, but the most successful are nonetheless starting to set clear trends. Kirkland’s commitment of $500m over the next three to four years to developing its own AI platform, starting with around $100m in 2026 implies a healthy chunk of that figure will go towards AI, compared to its $10.6bn of revenue last year, that means this year’s planned investment is equivalent to just under 1% of annual revenue. Goodwin Procter is landing in almost exactly the same place. The firm has earmarked around $25m per annum spend on buying/downstream AI tools, versus revenues of over $2.7bn (or 1%).
A lot of lawyers are talking about “how much spend should be allocated”. The challenge for law firms is that, when it comes to AI, headcount and software are no longer zero sum; if an application accelerates workflow, thereby reducing the number of lawyers required to perform a certain task, the correct metric to evaluate it against is the value of headcount / working hours saved, rather than software costs.
Goodwin’s target for its AI initiatives is illustrative of this. By aiming to release 1m hours of lawyer and business professional time per annum – around 1.5 hours per working day – it is in effect arguing that its $25m spend on AI technologies should be assessed not against decreases in software spending, but increases in value of services derived from the reallocation of thousands of lawyer hours towards more lucrative or differentiated areas of work. A similar argument applies to the $50m allocated to its legal analytics platform, which targets a reduction of 100,000 hours of lawyer and business professional time per year.
This, in turn, throws light on the build-versus-buy debate. While Kirkland represents the vanguard of the build school – with over 180 technology professionals anticipated to work on its platform – it is nonetheless notable that it will license a range of third party AI tools for its platform. For most law firms, a hybrid approach is likely to be the most viable, given the challenge of attracting the necessary technical expertise to support in house AI capabilities. Harvey has raised over $1bn and was valued at $11bn in March, while Legora is understood to be in advanced talks to raise a round of financing at a valuation of over $10bn. This will enable these firms to invest heavily in their engineering talent and in product development, which will allow them to fund their expansion at a scale far beyond what most traditional law firms would be able to achieve.
George’s thoughts: My sense is that building out AI capabilities in house may well become another lever differentiating the most successful law firms from everyone else. Most firms are likely to focus on downstream applications and make their own niche in building in-house workflows and data assets while buying best-in-class underlying technology platforms.
That said, any firm that wishes to have a genuine chance at product development should consider investing more heavily in its own technical talent. The economics of hiring suggest that they should be willing to pay engineers comparable to what other technology firms in the broader tech industry are offering for similar roles. Harvey has already started to do this, with its legal engineering roles – for highly qualified candidates – now starting at $220,000-$320,000 plus equity.
Spellbook launches an AI-powered document editor
DeepJudge unveils a new protocol for handing work between AI agents
Anthropic moves to watermark AI-generated content
Ivo expands into contract lifecycle management with ‘Collaborate’
Aloi raises $7m to build AI for ‘organisational judgment’
Legora reportedly targets a $10bn valuation
Forward Deployed Legal Engineer at Epiq (London)
Senior AI Engineer - Legal AI at 9Fin (London)
AI Native Engineer at Legalist (New York)
Anthropic is going to start to watermark text generated by its Claude models - but not in the way you might expect. Claude’s next-word suggestions will have its randomness tweaked so that, over large texts, there will appear to be statistically meaningful deviations from randomness that only Claude would produce. Nothing will be added to the text itself, and no special characters inserted. Anthropic will use Google DeepMind’s SynthID-Text, and make an API available to calculate probabilities that a given text was generated by Claude.
The most straightforward reason is down to regulation - specifically, Article 50 of the EU AI Act, which entered into force on 2 August 2026. This obliges providers of generative AI to make their outputs - text, images, audio, and video - “detectable and attributable to an AI system to the extent technically feasible.” Anthropic has also signed the EU’s Code of Practice on Transparency of AI-Generated Content, pledging to watermark content “initially on a global level,” as it “cannot be regionally limited at this time.” The true reason likely has less to do with bans on AI-generated content, and more to do with establishing a baseline for provenance research, so that platforms, publishers, and other entities have more reliable methods for differentiating between human and AI content.
Whether or not that will happen is a separate issue. Anthropic itself notes that detection becomes more difficult “for shorter texts, fact-like content, and text that has been substantially revised.” Detection becomes “challenged entirely” when the text has undergone extensive rewriting - and a watermark only serves to indicate that Claude was used to generate it, not that it is authentic, factual, accurate, or otherwise good in any particular respect. A more significant effect may be seen at the level of individual platforms. If large social media platforms and search engines begin employing similar detection mechanisms, texts identified as AI-generated may be de-platformed or ranked lower than their human counterparts, making large-scale dissemination more difficult.
Claude’s watermark, in other words, will not stop AI slop - but it will make AI-generated content significantly easier to detect, and consequently to control.
If you have suggestions for improving Best Practice, or simply want to introduce yourself, then hit reply to this email. I read every response!
See you next week,
George, Editor of Best Practice
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.