The most important sentence in legal AI this year was only half a boast
When Kirkland & Ellis committed $500mn (£394m) to a proprietary AI platform in May, its chair Jon Ballis told the Financial Times that widely available tools were “raising the floor for everyone”, but that his firm had to go further because “we don’t get hired for the floor”. The line was read as a statement about Kirkland. It is more useful read as a statement about everybody else: the floor is rising, whether or not you have half a billion dollars to stand above it. Around a fifth of the world’s large corporate law firms are already experimenting with building or adapting legal AI of their own, on an estimate from Brian Tang, who is executive director of LITE Lab@HKU, an interdisciplinary programme at Hong Kong University. The other 80% are not standing still. They are standing on a floor that keeps moving up under them.
Strategic Insight: The build-versus-buy framing flatters the firms doing the building and paralyses the ones that cannot. For most UK practices the operative question is narrower and more answerable: which parts of our advantage are already inside the software we rent, and which parts are still ours?
Why commoditised tooling is a pricing problem before it is a technology problem
Generative AI reached near-universal use in the profession before most firms had settled what it was for. Adoption among legal professionals surveyed by LexisNexis has reached 94% while worry about fabricated output climbed to 83%, which is a fair description of a market that bought first and reasoned afterwards. The tools work. Contract drafting, first-pass research, document review and summary are handled competently by software any competitor can license by the seat.
That competence is the problem. When a capability is purchasable, it stops being a reason to instruct one firm over another and becomes a reason for the client to ask why the bill has not moved. Our news desk covered the announcements and the sums involved when the FT reported them. The strategic residue is simpler than the spending: bought tools have become the price of entry, and a price of entry cannot also be a differentiator.
| Data point | Figure | What it means for a firm outside the top tier |
|---|---|---|
| Kirkland & Ellis platform commitment | $500mn (£394m) | Sets a spending benchmark no UK mid-market firm can or should match |
| Goodwin Procter annual AI budget | ~$25mn (£20m) a year | Even the “modest” build budget exceeds many firms’ entire IT spend |
| Large corporate firms building or adapting tools | ~20% worldwide | Four in five are competing on something other than proprietary tooling |
| Global legal tech market, 2026 to 2033 | ~$31bn rising to ~$70bn | Most of that growth is money spent buying, not building |
Those market figures, which the Financial Times attributes to the research company Grand View Research, describe a buying market. The build stories are being reported precisely because they are the exception.
What “building your own AI” actually turns out to mean
Read the detail and almost none of these firms are building models. Freshfields runs software from Harvey and Legora, the two vendors most often named as market leaders, and separately builds, partnering with Anthropic and with Google Cloud, and running a technology lab of 50 people in Berlin. Its agreement with Anthropic buys proximity to the developers and early sight of what is coming, not a model of its own. Sebastian Lach, who is co-chief executive of Eltemate, the technology entity of Hogan Lovells Cadwalader, and a partner at that firm, puts the substance of it plainly: the technology cannot be bought off the shelf because “the knowhow needs to be injected as part of the tool”.
That last clause is the whole argument, and it is not about software engineering. The scarce input is codified expertise: how this firm structures a particular deal, which clauses it will and will not accept, what its best partner checks before signing off, why a precedent was drafted the way it was. Buying a platform does not create that. Building one does not create it either. It has to already exist in a retrievable form, and in most firms it does not.
Critical Context: Every firm already owns the raw material for differentiation. It sits in matter files, precedent banks, closed transactions and the heads of eight or nine people. What it lacks is not a model. It is the structure that would let a model reach it.
The variance is in the user, not the product
Max Junestrand, chief executive of Legora, makes an observation that should reorder budgets. He contrasts the average buyer of Legora with the most sophisticated one, and the distance between them is considerable: “You can be extremely advanced and effective by really leveraging all of the different bells and whistles in the tool, versus being a novice.” Gabe Pereyra, who co-founded Harvey and is its president, notes that most firms accept some infrastructure will always be bought in, though enquiries about differentiation rose after Kirkland’s announcement.
Both are interested parties. Both are also describing something visible in any firm that has run a legal AI licence for a year: the distribution of value across seat-holders is wildly uneven, and nobody is measuring it. A firm paying for 400 seats and getting expert-level use from 30 of them has a return problem that no amount of proprietary development will fix, because the same 370 people will underuse the bespoke tool too.
Hidden Cost: Under-instrumented licences are the largest unbooked write-off in legal AI. The spend appears in the technology budget; the waste appears nowhere, because nobody reports utilisation depth to the board.
Who actually has to do this work
The firms treating this as a procurement decision are giving it to the wrong people. Codifying knowhow is knowledge management, quality assurance and supervision, run by lawyers with practice credibility, and it competes directly with chargeable time. That is the real constraint, and it explains why so many firms would rather sign a licence than do the harder thing.
The supervision burden is already established: our analysis of the hallucination problem in legal practice set out why verification at source consumes exactly the time the tool was bought to save. A separate warning that City firms risk a crisis of judgment as juniors accept AI output uncritically describes the other end of the same failure. Both get worse, not better, when the tooling becomes more capable and less visible.
| Stakeholder | Primary impact | What they need | How to measure |
|---|---|---|---|
| Managing partner / board | Spend rising with no defensible differentiation story | A stated position on what the firm will never buy in | Share of AI budget tied to a named practice advantage |
| Practice group heads | Client pressure on price for now-commodity work | Authority to reprice commoditised tasks deliberately | Realisation rate on AI-assisted matter types |
| Knowledge lawyers / PSLs | Become the differentiation function, usually without resourcing | Ring-fenced non-chargeable time and a retrieval standard | Volume and reuse rate of codified precedent |
| IT and information security | Confidentiality exposure widens with every integration | Veto authority and a client-consent position | Approved integrations vs. shadow tool detections |
| Junior lawyers | The training tasks are the automated tasks | Deliberate exposure to work the tool now does | Supervised verification hours per trainee |
🎯 Success Factor: The firms that get value from this decade will be the ones whose knowledge function had budget and status before the technology arrived. Where it was a cost centre run by two part-timers, no platform will compensate.
A route that does not require a platform budget
💡 Implementation Framework: Codify before you customise
Phase 1: Establish the baseline (Weeks 1 to 6)
- Measure depth of use, not seat count, per practice group on existing licences
- Identify the three matter types where the firm genuinely outperforms the market
- Audit whether the knowhow behind those three is retrievable or personal
Phase 2: Codify the differentiators (Months 2 to 6)
- Structure precedent, playbooks and drafting standards for those three areas so a tool can consume them
- Configure retrieval and prompt libraries against that material inside the software you already license
- Set one named owner per area and give the role protected non-chargeable time
Phase 3: Decide what, if anything, to build (Month 6 onward)
- Reassess build proposals against a codified baseline rather than a vendor demonstration
- Price the confidentiality, security and maintenance load, not just development
- Treat any build as a wrapper on codified knowhow, never as a substitute for it
Priority actions by where you are starting
If AI use is informal or ungoverned. Stop the tool comparison. Find out which of your practice areas clients instruct you for rather than a cheaper competitor, and establish whether the reasoning behind that work exists anywhere other than in individual memory. Set a single rule while you do it: anything AI-assisted that leaves the firm is verified at source by a named person.
If licences are deployed and used. Instrument them. Depth of use per group, per matter type, against the outcome you bought the tool for. Then invest the difference between average and expert use before you consider a development budget, because Retrieval-Augmented Generation over your own well-structured material typically returns more than fine-tuning over material that was never structured.
If a build case is already on the table. Ask what proprietary input the tool will consume on day one. If the answer is a plan to gather it later, the project is a knowledge management programme wearing a software budget, and it will be cheaper and more honest run as one.
Resource Reality: Phase 1 costs roughly four to six weeks of one senior knowledge lawyer’s time plus reporting access from your vendor. Phase 2 is the expensive part, but it is expensive in partner attention rather than capital, which is why it gets deferred indefinitely in firms that have not made it someone’s job.
Four things this decision breaks that nobody budgets for
The knowhow you meant to codify walks out
The material that would differentiate a bought tool is concentrated in a small number of senior practitioners, and it is not written down anywhere a system can read. Those people are also the busiest and the most likely to be recruited by a firm that has already worked this out.
Mitigation Strategy: Treat codification as a retention and succession issue rather than a technology project. Start with the practice areas most exposed to a single departure, and make contribution to the precedent estate an explicit part of partner appraisal rather than a discretionary favour.
Outsourcing the build does not outsource the liability
Brian Tang expects a service-provider market to emerge for firms that want bespoke tooling without becoming software developers, particularly smaller ones. That market will be useful. It will also not absorb the confidentiality and security obligations, which stay with the regulated firm no matter who wrote the code.
Mitigation Strategy: Assess any provider on how it handles client confidentiality and incident response before assessing its product. Establish the firm’s position on client consent for third-party processing before signing, not during a matter, and keep the SRA and the wider regulatory picture in the room while you do it.
The client repricing conversation arrives first
Clients are running the same experiments. Alan Mason, Freshfields’ global managing partner, notes that clients are also still working out which tool suits which task, and that his firm can deliver advice inside whichever platform a client prefers. A client fluent enough to hold a preference is fluent enough to ask what the technology has done to your fee.
Mitigation Strategy: Decide deliberately which task types you will reprice and which you will defend, before a client forces the question in a panel review. Repricing commodity work from a position of choice reads as confidence; doing it under pressure reads as a discount.
Training pipelines break silently
The tasks that built judgment in a second-year solicitor are precisely the tasks that are now automated. The damage appears three to five years later, in a cohort that can operate the tools competently and cannot tell when the output is wrong.
Mitigation Strategy: Reserve a defined proportion of automatable work for supervised human completion and treat it as training investment with a named cost, not as inefficiency to be eliminated. Measure verification hours per trainee alongside utilisation.
Reality Check: None of this produces a board-level announcement. Codifying three practice areas takes six to twelve months, generates no press release, and is the only part of this that a firm without a platform budget can actually own.
What a firm below the magic circle is really deciding
The bifurcation the FT describes is real, but the two options it presents are not the two options most UK firms face. Nobody outside a handful of practices is choosing between a $500mn platform and off-the-shelf software. They are choosing between letting bought tools define their capability, which converges every firm on the same floor, and doing the unglamorous work of making their own expertise machine-readable so that the same bought tools produce something a competitor’s licence cannot.
Three factors separate the firms that will manage it:
- A defensible answer to “what do we do that a licence cannot”. Named practice areas, not a values statement. If nobody can name three, the differentiation problem predates AI.
- A knowledge function with authority. Codification competes with chargeable hours and loses every time unless someone senior has made it un-loseable.
- Measurement of depth, not deployment. Seat count is a procurement metric. Expert-level use per practice group is a strategic one.
Strategic Insight: The build-or-buy question has an unhelpful shape because it puts the variable in the software. The variable is the input. A firm with codified knowhow gets differentiated output from a bought tool; a firm without it gets an expensive commodity either way.
Your next steps
Immediate (this week):
- Pull depth-of-use reporting from your existing legal AI vendor, by practice group
- Name the three matter types where the firm genuinely outperforms
- Confirm who owns knowledge management and what protected time they have
This quarter:
- Audit whether the reasoning behind those three areas is retrievable or personal
- Set a verification standard for AI-assisted work leaving the firm, with named accountability
- Agree a board position on which commoditised task types will be repriced
This year:
- Structure precedent and playbooks in the differentiating areas for machine retrieval
- Reassess any build proposal against that codified baseline
- Establish a client-consent and confidentiality position covering third-party AI processing
Source: Law firms seek bespoke differences in legal AI, reported by Nick Huber for the Financial Times, 3 September 2026.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the outsourcing market Brian Tang anticipates actually reaches UK regional and mid-market firms, or whether the gap between the building 20% and everyone else simply widens. Read more analysis at Insights, or get in touch.