Two research reports landed on the same day with opposite messages. Deloitte's "The AI Imperative" projects that AI agents could take on roughly 30% of in-house legal work within three to five years, while a new Litera survey finds legal departments adopting those tools far faster than they are building the governance to control them. Big Law made two contrasting bets in the same week: Reed Smith sent its partners back to school with a Cornell AI leadership programme, and Kirkland & Ellis signed an exclusive litigation-AI partnership with Syllo on top of its $500M in-house build. And Spellbook launched Autonomous Contract Management, pushing routine contract work from AI-assisted to AI-run. All of it sits against a loud week in the wider AI market: Anthropic's Fable 5 returned globally after US export controls were lifted, and OpenAI was reported to be weighing a delay of its IPO to 2027 as Anthropic's valuation overtook it — the models and capital that legal AI ultimately runs on. Read together, the week describes a market repricing capability, governance, talent, and the platforms beneath them all at once.
June 30 produced the clearest before-and-after picture the legal AI market has had all year — not because the two reports agreed, but because they didn't. Deloitte's "The AI Imperative" measures how much work is about to move to machines; Litera's survey measures how little of the control layer that move requires is actually in place. Read side by side, they describe a capability curve and a governance curve pulling apart.
Deloitte's report, based on a survey of more than 100 legal-department leaders across the Americas, APAC, EMEA, and the UK and Ireland, puts a number on the direction of travel: AI agents could take on roughly 30% of in-house legal work within three to five years. The knock-on effect lands on law firms. Eighty-five per cent of respondents expect AI to change how firms bill to a moderate, large, or very large extent, and the share of firm work billed by the hour is projected to fall from 72% today to 44% within two to three years — with some general counsel already targeting 20-40% cuts to outside-counsel spend over the same window.
Litera's survey is the counterweight. Legal leaders report real operational gains — 64% say AI is improving efficiency and 62% say it is automating workflows — but the governance scaffolding is missing. Seventy-five per cent name data privacy and security as their top AI-related risk, and more than two-thirds say they have not updated key contract clauses in the past twelve months to reflect the exposure AI introduces. Litera's framing is that in-house teams are "winning on risk instinct, losing on governance infrastructure": the people sense the danger, but the policies, clauses, and controls that would contain it haven't been built.
These two reports are the same story told from opposite ends. Deloitte describes the prize — a third of in-house work moving off the traditional cost base — and Litera describes why so many teams can't safely claim it yet: the guardrails are missing. The gap between "we adopted a tool" and "we have a governed way to run it at volume" is exactly where commodity work stalls. Inexpensive work keeps getting done by expensive resources not because anyone chose that, but because the alternative — high-volume NDA, procurement, and redline work run by supervised agents that know your templates, terms, and escalation rules, with a human reviewing every output before it leaves the system — is the piece Litera's respondents haven't built. The 30% is reachable. The governance layer is the toll on the road there.
Deloitte's 30% is an industry average; your department's real number is higher or lower, and you can't know which without doing the categorisation work. Have you mapped which matter types in your own queue are the agent-ready share — by volume, spend, and escalation rate — and paired that map with the governance controls each category needs before it can run? Or are adoption and governance still moving on separate tracks, the way Litera says they are across the market?
Big Law's AI strategy has usually been discussed as a single race. This week it split visibly into two theories of the firm. Kirkland & Ellis bet on proprietary systems, taking an exclusive partnership with a litigation-AI vendor; Reed Smith bet on people, sending its partners to a custom AI programme at Cornell. Both are responses to the same pressure Deloitte just quantified — and they are not mutually exclusive, which is the point.
Kirkland's June 29 partnership with Syllo gives the firm the exclusive right, for a period, to build proprietary tools within and around Syllo's litigation-AI platform, chosen after a multiyear evaluation of the market. It sits on top of the $500M in-house AI programme the firm disclosed a month earlier, staffed by more than 180 AI engineers and data scientists and roughly 100 equity partners embedded in "AI Pods." Reed Smith, announcing the same day, went the other way: a custom curriculum with Cornell, taught at Cornell Tech this September, moving partners from foundational AI literacy through applied strategy, ethics, and governance.
The contrast is real but the two bets converge on the same requirement. A proprietary system is only as good as the partners who know when to trust it and when to override it; a trained partner is only as productive as the tooling they have to direct. The firms that pull ahead will be the ones that do both — and that treat the supervising human and the underlying system as a single capability, not two competing budget lines.
The same fork faces every in-house team, at a smaller scale. When you invest in legal AI this year, are you funding the tool, the training, or the operating model that connects them — and have you decided which of those you should own versus source from a provider that arrives with both the system and the supervision already in place? Big Law can afford to run both bets in parallel. Most in-house teams have to choose where their scarce time actually goes.
Spellbook spent four years as a drafting copilot inside Microsoft Word. Autonomous Contract Management, launched to early-access customers on June 30, changes where the work starts: contracts are pulled in, triaged, reviewed, and redlined against house standards before a lawyer opens the file, then tracked through signature and renewal. CEO Scott Stevenson frames it as "infrastructure for agreements," with a regulatory-change monitor, Spellbook Radar, due in Q4.
The reach is not trivial. Spellbook serves more than 4,500 legal teams across 80 countries, including Dropbox's in-house team and the law firm Kennedys, and is backed by Khosla Ventures at a $350M valuation. ACM is the company's biggest expansion since it launched in 2022 — and its clearest statement that end-to-end contract work is becoming an agent-native category rather than an assisted one.
What doesn't change is who configures it. ACM is licensed software the customer deploys, which means the customer's own team builds the playbooks, encodes the escalation rules, and monitors for quality drift as the system runs against live counterparties. That is the same governance work Litera's survey says most teams haven't done — and an autonomous tool raises the stakes on getting it right, because the tool acts before a human sees the file, not after.
Spellbook's launch confirms the diagnosis this briefing keeps returning to: the contract queue is inexpensive, high-volume work that has been sitting on expensive desks because the infrastructure to route it elsewhere didn't exist. An autonomous product is one answer — but it moves the burden rather than removing it, asking your team to become the systems integrator that builds and maintains the playbooks and escalation logic. Flank's answer to the same commodity-contract problem is a managed service: supervised agents that already know your templates, terms, and escalation rules, with a human review gate before anything leaves the system — so your team supervises outcomes instead of operating infrastructure. The two approaches solve the same routing problem from opposite sides of the org chart.
Every tool in this briefing is built on someone else's frontier model and funded by someone else's capital. Two stories outside the legal beat set the terms for the ones inside it: a frontier model went dark and came back by government order, and the largest AI IPO of the cycle slipped a year while the pecking order flipped.
Neither story is about legal work directly, and that is exactly why they belong here. A legal team evaluating an AI vendor is implicitly making a bet on the model that vendor is built on and the balance sheet keeping that vendor alive. When a model can be pulled offline by export controls for nearly three weeks, "which foundation model does this run on, and what happens if it becomes unavailable?" stops being a technical curiosity and becomes a continuity clause. When the capital market reorders which labs lead, the same question applies to the vendor itself.
You diligence a vendor's security and its features. Do you also diligence the layer beneath it — which foundation model it depends on, what your fallback is if that model is restricted or deprecated, and whether the vendor has the funding to still be here in three years? The Fable 5 outage and the OpenAI IPO delay are the same lesson from two directions: the AI you rely on is only as dependable as the model and the money underneath it.
A pair of surveys, two law-firm strategy moves, a product launch, a model pulled offline and restored, and a delayed mega-IPO don't obviously belong together. This week they describe one market at a single moment: capability, cost structure, talent, governance, and the models and capital underneath them all repricing at once — and moving at different speeds.
The capability curve is steepening. Deloitte's 30% and Spellbook's autonomous system are the same signal at two altitudes — an industry-wide projection and one vendor's concrete bet that end-to-end contract work can run with the human moved to the end of the process rather than the middle. The cost structure is following: outside-counsel spend targeted down 20-40%, hourly billing projected from 72% to 44%. And Big Law is hedging on both fronts at once, Kirkland building proprietary systems while Reed Smith trains the partners who will have to supervise them.
Litera is the curve that hasn't kept up. Adoption is real, the operational gains are real, but the governance infrastructure — updated clauses, data controls, documented human review — is missing in most departments. That gap is not a reason to slow down; it is the specific work that turns a promising pilot into a system a legal team can safely run at volume. The teams that close it first are the ones that will actually capture Deloitte's 30% rather than just reading about it.
Put the week together and the shape is clear: the size of the opportunity is no longer in dispute, the tools to capture it are multiplying, and the only thing consistently missing is the governed operating model that lets a team run commodity work at scale without a lawyer touching every file. That is the whole thesis this briefing has tracked all year — inexpensive work is still done by expensive resources wherever the routing infrastructure to redirect it hasn't been built. The teams that move first don't wait for the industry average to describe their own queue back to them; they outsource that work to supervised agents deliberately — agents that know their templates, terms, and escalation rules, with a human reviewing every output before it leaves the system.