Hi there,

I’m Martin, CTO at Flank.

Below you can read a rather personal reflection on how agents have actually changed my team.

Subscribe now

Coding became a team sport

I am a social scientist. My PhD was in computational social science, where I built agent-based models way before anyone heard of AI agents. I used them to research how thousands of simulated workers following a few simple rules turn small differences in what people know and can afford into large differences in where they end up. In my free time, I still work on personality and passion at work. It is a varied background, and until recently none of it looked like preparation for anything specific.

I now serve as CTO of Flank. I took over the engineering team at the start of the year, and the first big job was a strategy for how we would use AI coding agents. I expected the challenges to be technical. Most of them were social and organisational.

The first thing the agents did

The first thing the agents did was flatten us. The gap between what a junior and a senior engineer on the team can ship in a week has never been smaller. That matches what the studies of support desks and consultancies found: the people who gained most from AI were the ones who had the least to start with. I believe it. I also think it is half of the picture, because the studies of learners find the opposite once the tool is taken away. Students with a chatbot did better on their homework and worse on the exam. Engineers given an assistant for a library they had never used finished a bit faster and understood noticeably less, especially when something broke.

The second consequence surprised me more. Coding became a team sport, far more than it was when everyone typed their own code. You need the fast movers, the people who are not afraid to experiment and not afraid to tear it all out and write it again. Someone has to think about reliability, and about what the shipped system does at three in the morning. And you need the slow ones, the deliberate ones, the people who feel they own the product and read it as if they will be the one fixing it. One person with agents can build a product now. It still takes all of these people to grow it, maintain it and make it work for an enterprise customer.

The slowest engineer

The slow ones are the ones I keep coming back to. There is an engineer who is slower than everyone else this quarter, because he reads every line the agent writes and asks it why before he accepts anything. His pull requests are the smallest we have, and they are the ones I never have to reopen. He is also the one I would trust with the most sensitive parts of our system.

Which brings me to the thing I did not expect to use. The speed at which we ship new features looks like an engineering problem, and from inside the sprint it feels like one. In practice it is a problem of working environment and process. People need clear direction and they need slack. Permission to move quickly has to come with protected time to review the parts that are mission critical. It also has to come with permission to say, out loud, that the agent wrote this and I do not understand it well yet. That is important information about the state of the system, and a team where people hide it is a team that will find out the hard way. And someone has to have spent enough time with the code to understand it, so that when there is a bug or an incident at an awkward hour, a person can fix it rather than ask the agent to guess.

None of that is technical. It is organisational research, the kind I read and taught and never expected to apply, and I now use it every week. Whether an engineer keeps learning while the agent works is not mostly a matter of character. Reading two thousand generated lines costs an hour, and who gets that hour is something I decide when I set the sprint. It matters more than it sounds, because people grow into whatever I hand them as their main work. Give an engineer a year of accepting what the agent produces and that is the engineer you will have at the end of it: fast, productive, and stuck the moment something breaks. Give another a year of reading, questioning and fixing, and you get someone who understands the system and can be trusted with it.

The next few years

I think this is what the next few years look like well beyond engineering. AI will change industries, and I suspect a lot of the change arrives as connections nobody planned, between disciplines, between industries, between technologies that had little to do with each other. A social scientist running an engineering team is one. Lawyers becoming the engineers of their own playbooks is another, and the legal teams handing routine work to agents will meet the same exam ours did, with their juniors. The interesting ground is between the fields. It is where I have ended up, and I find it a good place to stand.

Share

⚡ 01 Google Cloud launched Gemini Enterprise for Legal with four elite firms already signed

On 25 August, Google Cloud launched Gemini Enterprise for Legal, a purpose-built version of its enterprise AI platform, with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly named as launch customers. The product ships specialised skills for contract review, due diligence, citation verification, regulatory monitoring and data-subject access requests. It is available in preview, alongside a parallel financial-services edition released the same day.

⚖️ 02 State Farm’s outside counsel admitted its filings cited seven cases that do not exist

Musick, Peeler & Garrett was defending State Farm in a Los Angeles fire-loss suit when opposing counsel’s review found the firm’s motions in limine cited seven fabricated cases, nine fabricated quotes and thirteen fabricated holdings. Lead trial counsel Kenneth Katel told the court that a co-counsel had used AI to help prepare the filings, and accepted full responsibility. A sanctions request from the plaintiff is pending.

⚡ 03 Thomson Reuters shipped its own model to reduce its dependence on frontier labs

Thomson Reuters spent a reported $40 million to launch Thomson, a proprietary large language model trained on decades of Westlaw, Practical Law, Checkpoint and Reuters content. Its CTO confirmed the base is Snowdon, an Imperial College London model realigned from Alibaba’s open-source Qwen. For now it powers only the tabular-analysis feature inside CoCounsel Legal.

⚡ 04 LawToolBox shipped an MCP connector for Microsoft 365 and Claude

LawToolBox released a Model Context Protocol connector exposing 70 tools that let any MCP-compatible AI reach a legal team’s matter containers inside M365: deadlines, documents, applicable laws, meeting transcripts and the team’s own strategic analysis. It continues the interoperability push this briefing has tracked since early August.

⚖️ 05 The New York City Bar said AI can assist lawyers but cannot replace their judgement

The Emerging Companies & Venture Capital Committee of the New York City Bar Association published a policy paper concluding that AI tools may assist legal work but cannot substitute for professional legal judgement. It maps existing rules of professional conduct onto AI use and sets out a six-part framework: understand, verify, protect, supervise, communicate, be truthful.

Receive two articles and our weekly briefing by subscribing below.

Subscribe now

⚡ 01 Google entered legal AI with four elite law firms already signed

Google Cloud spent years selling infrastructure to law firms from a distance. On 25 August it launched Gemini Enterprise for Legal, a purpose-built version of its enterprise AI platform, with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly named as its first customers. A hyperscaler entering legal AI with four of the most prestigious firms in the world already attached to the launch is a different kind of announcement than a specialist startup’s next funding round.

The product packages specialised skills for contract review, due diligence, citation verification, regulatory monitoring and data-subject access requests, on top of the general Gemini Enterprise platform. Connectors are built to automatically inherit a firm’s existing ethical walls from its document and matter management systems, and research outputs are grounded in primary legal authority rather than model training data alone. It launched in preview, alongside a parallel financial-services edition of Gemini Enterprise the same day, with Google describing healthcare and life sciences versions as on the horizon.

The connector list is the actual story

The product’s Model Context Protocol integrations reach directly into the tools a legal team is likely already running: Harvey, Legora, iManage, NetDocuments, Docusign, Everlaw, RelativityOne, Thomson Reuters HighQ, CourtListener, Solve Intelligence and Courtroom5, plus Google Workspace and Microsoft 365. That is not a differentiator so much as an admission that no single platform will own the workflow.

The question worth asking. Google’s connectors reach into Harvey, Legora, iManage and a dozen other platforms a legal team may already run, which is the same list every other vendor is racing to plug into. Once every major platform claims compatibility with everything else, the question stops being which platform to buy and becomes something none of them answer: who decides what task goes where, and who signs off before it reaches a client?

Share

⚖️ 02 A law firm’s routine filings for State Farm turned out to be full of fake cases

Musick, Peeler & Garrett was defending State Farm in a Los Angeles fire-loss suit when opposing counsel found the firm’s motions in limine cited cases, quotes and legal holdings that do not exist. Motions in limine are exactly the kind of routine, high-volume litigation drafting billed at a law firm’s full rate, and here nobody caught the fabrication before it reached the court.

Plaintiff’s attorney Eric Khodadian’s review of the filings in Meni-Siliga v. State Farm, a suit stemming from a 2020 house fire in Carson, California and first filed in 2024, turned up seven fabricated cases, nine fabricated quotes and thirteen fabricated holdings or statements of law. Lead trial counsel Kenneth Katel told the court a co-counsel had used AI to help prepare some of the filings, apologised to the court and the plaintiff, and said he accepted full responsibility as lead counsel. It adds a name, a firm and a specific list of fabrications to a hallucination tracker this briefing noted in July had already logged more than 1,598 documented cases worldwide.

The question worth asking. A carrier the size of State Farm pays a firm like Musick, Peeler & Garrett full litigation rates to produce filings like these, and the drafting still went through an AI tool with no independent check before it reached the court. That is inexpensive work, done by expensive resources, with no supervision layer in between: the fabrication was not caught by the firm’s own process, it was caught by the other side. So where in your own workflow does the check actually sit, and would it have caught this?

⚡ 03 Thomson Reuters built its own model to stop depending on one AI vendor

Thomson Reuters spent a reported $40 million training Thomson, a proprietary large language model, rather than routing every CoCounsel Legal query through a frontier lab’s API. The company that owns Westlaw and Practical Law just demonstrated that even the largest legal data holder in the world does not think one AI vendor should run its whole platform.

Thomson sits on top of an intermediate model called Snowdon, developed by the Thomson Reuters and Imperial College London frontier AI research lab and, per the company’s CTO, realigned from Alibaba’s open-source Qwen. A dedicated team worked on bias, political neutrality and toxic behaviour, with human and automated red-teaming. It is trained on decades of proprietary content from Westlaw, Practical Law, Checkpoint and Reuters, and Thomson Reuters says it shows a meaningful uplift over its open-source base in following complex, multi-part instructions and in reasoning over dense, domain-specific text.

A routing strategy, not a replacement

For now, Thomson powers only the tabular-analysis feature inside CoCounsel Legal. Everywhere else, Thomson Reuters says it will keep deploying frontier models from OpenAI and Anthropic where they hold a clear advantage. CoCounsel Legal, in the company’s own words, remains multi-model by design.

The question worth asking. Thomson Reuters is not trying to replace OpenAI and Anthropic. It is deciding, task by task, which model is worth the cost for that class of work, and building the routing logic to keep making that call automatically. That is the same discipline enterprise legal teams need one layer up, at the level of the work itself: a procurement NDA and a novel cross-border restructuring memo are not the same class of work, and routing both to the same expensive resource, whether that resource is outside counsel or a frontier model API by default, is the waste this briefing keeps coming back to. So which of your own work is being done by a resource priced for something harder?

⚖️ 05 A bar committee said AI can help lawyers but cannot replace their judgement

The Emerging Companies & Venture Capital Committee of the New York City Bar Association published a policy paper this month concluding that AI tools may assist legal work but cannot substitute for a lawyer’s professional judgement. The paper names where AI is already embedded, in drafting, document review, research and workflow management, and says the profession still lacks a uniform framework for using it responsibly there.

Written for lawyers and business executives in the emerging-companies and venture-capital space, the paper sets out how existing rules of professional conduct apply to AI use, offers a six-part framework (understand, verify, protect, supervise, communicate, be truthful), and recommends firms adopt protocols, designate technology ethics officers, train staff and disclose AI use to clients where appropriate. It stops short, notably, of prescribing what that framework looks like in operational terms: who checks the output, on what schedule, before it reaches a client.

The question worth asking. The committee is right that AI cannot substitute for judgement, but it does not say where in a workflow that judgement gets exercised, or by whom, once AI is producing the first draft of everything from an NDA to a due-diligence memo. A principle without an operational checkpoint is something a team can point to after something goes wrong, not something that prevents it. Where does your own review gate sit, and is it the same gate for a routine NDA as for a novel structuring question?

🧠 What this week tells us

A hyperscaler entered legal AI as a distribution story, not a supervision story. Google Cloud arrived with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly attached, and with MCP connectors reaching into every tool a legal team already runs. What it did not arrive with is any statement about who signs off on the output. Distribution changes which brand sits on the login screen. It does not change who is accountable for what comes out.

Even Thomson Reuters will not bet its platform on a single model. A $40 million proprietary model, deployed to exactly one feature, with frontier labs still doing the rest, is a company saying out loud that model choice is a per-task decision. Legal teams should read that as a template rather than a vendor announcement. The same logic applied one layer up, to the work itself rather than the model underneath it, is what stops routine volume being priced like bespoke advice.

Interoperability is now table stakes, which makes routing the scarce thing. Google’s connector list and LawToolBox’s 70-tool MCP for M365 point the same way: within a year, everything will reach everything. When every platform can touch every system, the differentiator is no longer access. It is the decision layer that says which piece of work goes to which resource, and what has to happen before the output leaves the building.

Unsupervised capability is a liability, and the profession’s answer is still a principle. Musick, Peeler & Garrett’s filings for State Farm show what happens when routine, high-volume drafting has no review gate attached: the error was found by opposing counsel, not by the firm. The NYC Bar’s paper, published the same month, says judgement cannot be delegated. Both are true. Only one of them is a mechanism.

✳️ The Flank view

Take the week’s threads together: two platforms added capability, one law firm proved unsupervised capability is a liability, and one bar committee said the quiet part, that judgement cannot be delegated, without saying how a firm operationalises that day to day. None of that is routing infrastructure a legal team owns for itself.

Outsource legal work to supervised agents and the review gate stops being a principle a bar committee states and a firm hopes gets followed. It becomes an actual mechanism: agents that know a team’s own templates, terms and escalation rules, with a human checking the output before anything leaves the building. Inexpensive work stops being done by expensive, unsupervised resources. That is the gap every story this week left open.

✳️

Subscribe now

Hello! I’m Laura, a Legal Engineer at Flank. I build agents and supervise them, making sure that we’re putting as much trust in them as they’ve actually earned.

Most conversations about legal AI focus on whether the agent got it right. This piece is about something that took me a while to figure out - whether the way we’re supervising is actually working.

Subscribe now

Picture this…

You have an AI agent assigned to review an NDA. It flags 4 clauses out of 40 as non-compliant. Instead of jumping straight to those 4 clauses, you end up trawling through the entire document, top to bottom, “just to be sure”. By the time you’ve completed it, you’ve redone the entire review yourself - those 4 flags raised by the agent are never approached differently from anything else.

The problem

One of the easiest traps in agent supervision is spending so much time checking an agent’s output that you end up redoing the work yourself. This instinct isn’t irrational either. In a report published by Thomson Reuters, one US attorney stated: “Agentic AI, while exciting, to me removes oversight a step too far. I like the idea of prompting and reviewing a result”.

We’re clearly having to find the balance between two instincts, either: I don’t trust the agent being able to find every issue in this contract, so I’ll review the whole thing myself anyway - or, I have an agent configured with the perfect prompt and instructions, so I’ll hand off all of my work without a second thought.

Despite the growth of AI in legal, most legal teams haven’t fully reached a point where they can offload the bulk of their work to an AI agent. In fact, the same Thomson Reuters report found that only 16% of law firms and 18% of corporate legal teams use agentic AI tools.

Supervising agents is the blocker standing between where they are now and actually adopting agentic AI.

But what exactly is agent supervision?

In legal AI, agent supervision is the process of a human checking an agent’s work before it’s relied on and sent to a counterparty. The goal isn’t always to avoid reviewing the whole task, but to make the level of supervision proportionate to contract risk and the trust the agent has actually earned. Fundamentally, the agent does the work, the human checks for exceptions, without needing to complete the task from scratch.

In the context of contract review, that might look like:

That’s the model in theory. Nowadays, most people (and most tools) treat supervision as something closer to blanket quality control:

Reactive vs. proactive supervision

I used to think that supervising agent outputs meant finding mistakes, but I was wrong. The value of supervision isn’t in the catch itself, it’s that the catch improves the next hundred outputs. It’s a feedback loop.

A pivotal moment in my time as a Legal Engineer was understanding that there is a clear difference between reactive and proactive supervision.

• Reactive supervision is fixing an error after it’s already been spotted.

• Proactive supervision is noticing that a playbook built to review your own template doesn’t work for reviewing a counterparty’s template, and rewriting the playbook before that mismatch causes a string of errors, not after.

It’s easy to fall into the trap of favouring reactive supervision simply because it’s easily measurable. For example, let’s say an agent’s output comes back with 10 redlines. If you mark 2 as genuine misses by the agent, you’ve got a clean 80% accuracy figure. Proactive supervision, on the other hand, is rarer because the biggest wins are invisible. Nobody saw you rewrite that playbook before it caused a single problem. Likewise, you can’t score a mistake that never happened.

One, low-effort tactic to adopt tomorrow

I saw this play out with an indemnity clause. There was one contract that needed an indemnity clause but didn’t include one, and the agent kept missing it. At first, each miss looked like an isolated slip-up. It was easy to mark as a one-off correction and move on. But the more I started logging why the agent was missing it, the pattern was obvious. The playbook only told the agent to review the indemnity clause if it was already sitting in the counterparty’s contract. It never told the agent what to do if the clause was missing entirely. Three misses weren’t three isolated mistakes. They were one instruction the playbook never had in the first place.

Share

So, how can proactive supervision actually be implemented in a workflow?

Start with this: every Monday morning, before reviewing any new agent outputs, spend 15 minutes scanning last week’s corrections. Tag each one with a short reason - wrong fallback, missing context, incorrect analysis - whatever makes most sense to you. If the same tag appears three times or more, that’s not three unrelated mistakes. You need to flag it, fix the rule behind it, and continue your review.

The result

Within a few weeks, you stop seeing one-off misses and start seeing patterns. Updating a playbook because you caught a pattern is proactive supervision. Updating it only after being prompted by a client to do so is reactive supervision in disguise. The difference really isn’t the tool, agent or system you’re using. It’s whether your supervision learns from its mistakes.

I’m Khaled. I’m a technology lawyer working alongside enterprise legal teams, thinking about what work we should hand to AI, and what work we shouldn’t.

The Limits of Human Effort

Until now, there was only one way to keep attacking a problem.

Throw more of yourself at it.

You could feel like the most disciplined person in the world at that moment.

You could work for hours and hours. Keep pushing at the same thing.

But eventually, and true to human nature, you hit a point where more effort stops producing better work. You’re, dare I say… tired.

At some point, you’re just moving things around and convincing yourself you’re still working. So you decide, “You know what, I’m better off going at this fresh again tomorrow.”

There is only so much energy we can throw at a problem. Or there was. Now, a certain kind of intelligence has become a commodity, so we can throw (artificial) intelligence at the problem.

The Limitless Energy of AI

With agents, that intelligence doesn’t need to stop when we do. It doesn’t get tired. It just keeps going. You can even run a hundred subtly different versions of the same intelligence at the same problem, all at the same time. And agents keep going enthusiastically! (Maybe even a little too enthusiastically at times.)

An agent can work towards a goal, check its own work, see where the output fails to meet certain criteria that you set or it has set for itself, and just go again.

Subscribe now

And again. (And again).

The upside? We can produce deliverables much quicker.

Provide faster answers.

Write faster.

Research faster.

Review (and redline!) faster.

Now, what I’m paying attention to is the work we can throw intelligence at that we would never have thrown human energy at in the first place. Either because the work was too great for the return we might get from it, or because we simply didn’t have enough years left in our working lives…

Take every contract a company has ever signed.

Every single contract.

Ask, Where are we taking on the most liability?

What do we keep conceding? At what point in the negotiation?

What patterns have we missed?

The Risk of Too Much Intelligence

Previously, someone would have to read thousands of contracts. So it was never going to be done. And because we knew the size of the task was too large, we just never asked these questions. Not knowing was a risk we were willing to absorb.

The same goes for years of customer conversations, thousands of support tickets, or every internal request a team has received. We have always had more information than we could realistically process. Until now.

Now we can throw intelligence at it. And that changes what work is reasonable to consider doing. And this sounds great in principle. But there is something I’m realising I need to be careful about here… The limits of human energy was also a filter: If something took three weeks, we had to decide whether it was worth three weeks. If a point in a contract wasn’t worth spending time on, we probably left it alone.

That filter is starting to disappear.

Share

A great post from Anna Guo the other day made me think about this in the context of contract negotiations.

AI has made redlines incredibly easy to produce. But someone still has to review all those redlines. If the counterparty is overzealous in their use of AI, you get a contract back with redlines on everything. Changes that don’t really matter. Comments that just describe a change, without attempting to negotiate or explain why it’s needed.

Redlines that carry zero awareness of the nuanced context of the deal in front of them.

Someone just told an AI to review the contract.

There was no one stopping to think, actually, this isn’t worth fighting over.

No feeling of the inherent cost that every redline carries.

So agents can take on work that we never would have done.

They can keep going long after we would have stopped.

But when production becomes cheap, we lose one of the things that made us ask whether something was worth producing at all.

We are going to get much better at answering: Can we do this?

The harder question is: Should we throw intelligence at it?

✳️

⚡ 01 Legora reportedly chases $10bn-plus, nearly double its April valuation

Legora is said to be in talks to raise at more than $10bn, up from the $5.6bn mark this briefing tracked in April. Harvey, which closed $200M at $11bn in March, was separately reported on August 7 to be in talks for $500M at $15.5bn, as its revenue passed $350M. Neither round has closed.

⚡ 02 Relativity gives lawyers a conversational interface over the whole matter

claiR extends Relativity’s aiR Assist from indexed document sets to the full depth of a matter, metadata and all, inside RelativityOne’s existing permissions. It is live in early access with A&O Shearman, Foley & Lardner and K&L Gates. General availability is expected in early 2027.

⚡ 03 Mayer Brown puts its own name behind an AI verification layer

The firm and Scissero launched a service that configures structured-products documentation on an automation platform and verifies it against approved templates, before Mayer Brown’s own lawyers sign off. It starts with US SEC-registered products.

⚡ 04 Harbor launches an AI implementation arm built only for law

Harbor Deploy embeds forward-deployed engineers and delivery leads inside law firms and legal departments, arguing that the deployment businesses OpenAI, Anthropic and Microsoft built this year were not designed for a profession where confidentiality and duty to the client are not negotiable.

🏗 05 The UK holds the launch webinar for its first AI regulatory sandbox, built for law

The Advisory AI Growth Lab, which opened applications on August 3, will give around a dozen legal services providers up to nine months of coordinated access to the SRA, the Legal Services Board, the Council for Licensed Conveyancers and the ICO. Applications close September 27.

Receive one hand-written article and our weekly briefing by subscribing below.

Subscribe now

⚡ 01 Harvey and Legora both went back to the market this week

Harvey is reportedly in talks to raise $500M at a $15.5bn valuation, five months after its last round closed. Days later, Legora was reported to be chasing a valuation above $10bn, up from the $5.6bn it hit in April. Between them, the two platforms are discussing adding more than $9bn in paper value in under four months, a pace that says more about investor appetite than about anything a legal team can act on yet.

Harvey’s talks, first reported on August 7, follow a quarter in which the company says its revenue passed $350M, up from roughly $300M in May and $100M a year earlier. Neither round has closed, and neither company has confirmed the figures on the record. This is talks, reported by outlets covering the deal, not a signed term sheet. If it lands, it would be Harvey’s second valuation step-up this year, from $11bn in March to $15.5bn now.

Legora’s trajectory has moved on a similar clock. The company crossed $100M in ARR in April, roughly 18 months after launch, the fastest any enterprise software company has reportedly hit that mark, per Bessemer’s own tracking. A valuation above $10bn would nearly double the $5.6bn mark it reached the same month, when NVIDIA’s venture arm joined its Series D extension. Both companies are still, fundamentally, selling software that sits next to a lawyer and makes the lawyer faster. Neither round changes who decides what that lawyer still has to check.

The question worth asking. This capital is chasing model and workflow depth: faster drafting, more integrations, a bigger agent library. It is not chasing the question of who owns the routing decision once the software works as advertised. A $15.5bn valuation prices in every enterprise legal team eventually buying this kind of tool. It says nothing about who decides which piece of a matter an agent is trusted to touch, or who checks the output before it reaches the business. That is still, for now, an expensive in-house lawyer’s job. The same job all this capital is implicitly betting will eventually cost less.

Share

⚡ 02+03 Two products this week, one identical line around the lawyer

On back-to-back days, two very different companies shipped products that expand what AI can reach inside a legal matter, and drew the same boundary around what it is allowed to decide alone.

Relativity widens the aperture without removing the reader

Relativity’s claiR lets a lawyer ask natural-language questions across the full depth of a matter, not just the indexed document set aiR Assist already covered, but metadata and the connections between documents too, without the data leaving RelativityOne’s existing permissions. It is running in early access at A&O Shearman, Foley & Lardner and K&L Gates, with general availability expected in early 2027 at no added cost within RelativityOne’s integrated pricing. The lawyer still has to ask the question and read the answer. claiR widens the aperture, it does not remove the reader.

Mayer Brown draws the line somewhere else and lands in the same spot

Mayer Brown’s arrangement with Scissero draws the line in a different place but lands in the same spot. Client documents get configured on an automation platform, then verified by Scissero’s AI against approved templates and parameters, the kind of matching work a junior associate would otherwise do line by line on a structured-products issuance. What does not change is who signs: Mayer Brown’s own lawyers still provide the legal opinion the client is actually paying for. The service starts with US SEC-registered products and is expected to extend to exempt offerings, European products and defined-outcome products later.

The same week, Aloi raised $7M to build a “Judgment Graph” that captures how a firm’s own lawyers reasoned through past matters. That is a bet that the valuable thing to automate is not the sign-off itself, but making the firm’s accumulated judgement reusable by whoever handles the next similar case.

The question worth asking. Both products widen what AI can see inside a matter and leave the sign-off exactly where it was. If the boundary never moves, the software gets cheaper and faster while the expensive part of the process stays fixed. What would actually have to change for that boundary to move, and who inside your team has the authority to move it?

Share

⚡ 04 AI labs built deployment arms this year. Legal got its own.

OpenAI, Anthropic and Microsoft each spent 2026 building businesses that embed their own engineers inside client organisations, to get enterprise AI from pilot into daily use. On August 10, Harbor built the same thing for one industry only.

Harbor Deploy embeds forward-deployed engineers and delivery leads directly inside law firms and corporate legal departments, under what Harbor Labs managing director Justin Hectus calls an “Advise, Implement, Manage” model. The pitch is specific to law: general-purpose deployment teams, Harbor argues, are not built for a client relationship where confidentiality, professional judgement and duty to the client are non-negotiable constraints rather than compliance line items. It is the first offering of its kind built exclusively for legal, arriving into a deployment market the frontier labs effectively created this year by proving that shipping a capable model is not the same as getting it used.

The question worth asking. Getting AI into daily use still means paying skilled people, now embedded on-site instead of billing by the hour, to do the last-mile work of making an agent trustworthy for one task at a time. That is a legitimate business. It is also, structurally, another expensive resource solving a routing problem, just with a headcount answer instead of a software one. Once the agent is deployed, who is actually checking its output before it reaches the client: the forward-deployed engineer, or someone at the client with the authority to escalate?

🏗 05 The UK picked one industry for its first AI sandbox: law

A regulatory sandbox is usually a fintech instrument. This one launched with legal services and conveyancing as the sole pilot sector, and an explicit goal of making legal work “faster and more affordable.”

The Advisory AI Growth Lab opened applications on August 3 and held its launch webinar on August 10. Around a dozen participants, covering legal services providers, legal tech firms and conveyancing companies, will get up to nine months of coordinated, free access to the Solicitors Regulation Authority, the Legal Services Board, the Council for Licensed Conveyancers and the Information Commissioner’s Office, to work out how existing rules on client confidentiality and data protection actually apply to the AI they are already building. No grant funding is attached for participants. Applications close September 27.

The government’s own framing is unusually direct for a regulatory announcement: the goal is to “improve access to justice for the public by enabling high-quality, faster and more affordable legal services.” That is a government acknowledging, in its own policy language, the exact mismatch this briefing keeps coming back to. Inexpensive work is currently done by expensive resources, priced out of reach for anyone who is not already paying for it, and the bottleneck is not whether the technology works.

The question worth asking. If regulators already agree the problem is cost and access, not capability, what is actually stopping wider adoption today? Not the models. The routing infrastructure, meaning who decides what gets automated, on what terms, with what escalation path when it is wrong, is precisely what a nine-month advisory programme can clarify permission for, but cannot build.

🧠 What this week tells us

Capital is chasing depth, not ownership. Harvey and Legora’s funding talks price in more software: faster drafting, deeper integrations, a wider agent library. Between them the two platforms are discussing more than $9bn in fresh paper value in under four months. Neither round touches who supervises what the software produces, or who carries the risk when it is wrong.

Deployment is now its own market. Harbor followed the frontier labs into embedded engineering. The shape of the offer is telling: expensive people, now on-site rather than billing by the hour, still needed to get inexpensive tasks moving. When the answer to an adoption problem is headcount, the adoption problem is structural, not technical.

Every vendor keeps the human at the end. claiR and the Mayer Brown and Scissero service both expand what AI touches while leaving the sign-off exactly where it was. That is a deliberate, defensible choice. It is also the clearest signal in the week that nobody in the market has yet built the thing that would let the boundary move safely.

Regulators are quietly agreeing with the thesis. The UK built its first AI sandbox around legal services because the work is priced out of reach, not because the technology is not ready. When the policy language names cost and access as the problem rather than capability, that is a regulator conceding the bottleneck sits in how work is routed and supervised, not in the models.

✳️ The Flank view

Take the week’s threads together and they point the same direction. More capable models, a new deployment layer, and two products that each widen what AI can reach inside a matter. None of it changes who currently decides what gets automated and who checks it before it goes out. That is still an associate, a partner, or a newly embedded forward-deployed engineer: expensive people, doing the routing work by hand, because the infrastructure to do it otherwise does not exist yet at most legal teams.

That is the gap Flank is built to close. Outsourcing routine legal work to supervised agents means the routing decision, meaning what goes to an agent, on what terms, with what escalation path, is built once, into the system, rather than re-made by a person every time a matter crosses a desk. The agent knows the templates and the terms. A human still reviews the output before it leaves. Nothing shipped this week replaces that model. Most of it, read carefully, is still waiting for someone to build it.

✳️

Share

Hi there, I’m Taariq. I lead Forward Deployment at Flank.

Today I’m going to share some insights from these enterprise deployments.

Deploying an agent is a hiring problem

There’s a well-trodden path in enterprise deployments, which starts with gathering requirements and ends in user acceptance testing (UAT). The agile methodology built upon this and brought iteration cycles down considerably.

Agentic deployments can’t follow the same pattern.

Firstly, model capabilities change so quickly that designing a system based on today’s capabilities doesn’t quite work.

Secondly, agentic is a buzzword that’s been thrown around, but you need to take the word ‘agentic’ literally to fully appreciate it. You’re giving a system agency to act on your behalf. That’s a very different operating model compared with using a tool to achieve a goal.

In truth, there’s no playbook with agentic AI. The closest we have is comparing agents to employees and I’d like to explore that thread in this post.

Writing a job description

A customer comes to you with a specific problem to solve. Unless it’s a top-down mandate to ‘use AI’, the problem is generally articulated well.

In my opinion, this is the most important part to get right because there’s a tension between scoping this description ‘properly’ (i.e. exhaustively) and building whatever comes to the customer’s mind.

If you define this job description in a rigid way, the customer will only get a fraction of the value they wanted from the agent, and ultimately the deployment will fail.

This is because the nature of agentic deployment is very different to others. There will be things the customer doesn’t mention up front – not because they’re negligent, but because we’re looking to automate human processes that often contain some variability and judgement in the moment.

Trying to document everything up front is akin to writing an onboarding manual for a new employee – but only writing it once, without giving them any opportunity to change what or how they do the work once they figure things out.

The other end of the spectrum, where you very loosely define what needs to be built, is equally dangerous. This way of working is essentially a gun for hire – you build exactly what the customer wants at the time. For example, when configuring a drafting agent, we were asked to include a tracked-changes version of the agent’s drafted agreement vs the template. This was a check introduced by the business at some point to ensure no typos got through. Agents don’t make typos – so this really isn’t needed, but what the team was actually saying is that the tracked changes document is how they built confidence in everything they send out.

What the first version of the job description leaves out

Stress-testing the design is easier said than done, so I’m sharing what’s worked for me. The key to a solid design is to unearth the unwritten process. The written process is useful up to a point, but it doesn’t cover the complexity of the particular task.

The first port of call is to shadow someone whilst they do the job. They’ll usually show you a relatively simple example of what they do. Ask for one that doesn’t follow that particular blueprint – this is harder for someone to show you because it requires more effort for them to pull together all the sources, explain why it’s an exception, and then show you what they did (whilst being wary that they may have broken some rules by going off the happy path)

Which leads me to the next point: design for exceptions as well as the happy path. It’s often comforting to go down the happy path because that’s where the majority of the volume often is (everyone likes to tell themselves this, but I’ve found it isn’t generally true).

My favourite question when probing this is to ask ‘what’s the most complex one you’ve done this week – one that an agent wouldn’t know what to do with because of the nuances and intricacies.’ It usually yields things that an agent can do perfectly well – if the deployment strategist can re-articulate the task in general terms.

There’s a balance to this too. You don’t want to design something that’s exhaustive – that would go against the golden rule of not overfitting.

Share

Adoption is the most important test

You evaluate the performance of an employee soon after they join, and you do the same with an agent. You look at the outputs and you compare them to what’s expected. This is similar to the UAT phase in traditional deployments.

However, even when the agent’s evaluation is deemed ‘good’ that doesn’t mean it will be useful. That only happens when people trust the agent – and this gets built over time. At first, its outputs will be checked very closely, but this will gradually reduce as trust in the agent increases. Based on all of the deployments I’ve overseen, this trust-building phase is natural and no shortcut exists. Trust takes time to build (proportional to the outcome, in general) so this phase needs a lot of thought. The key question to ask before deploying an agent is “who’s doing the onboarding?”

This must be someone who does the job (and not someone from procurement or IT) and who has time to dedicate to the onboarding.

✳️

Share

⚡ 01 Wordsmith raises another $14M as Ironclad ships agents chasing procurement savings

Wordsmith’s Series B extension takes its total raised to roughly $114M, weeks after its last round closed. The same day, Ironclad launched agents that surface rebate and compliance obligations buried in supplier contracts and route them straight into SAP.

⚡ 02 Harvey becomes the first legal AI agent certified against AIUC-1, a standard barely a year old

Auditor Schellman ran more than 3,000 tests across data privacy, security, safety, reliability, accountability and societal impact, and Harvey reports zero critical failures. The announcement landed days after a growth investment from Goldman Sachs Alternatives and J.P. Morgan’s Growth Equity Partners.

⚡ 03 Nvidia and SK Group turn a summer handshake into a $500bn AI build-out

The two formalised the AI factory and next-generation memory partnership first floated as a letter of intent in July, spanning a planned 2GW AI factory in Korea and joint development of the memory that will feed the next generation of chips.

🏗 04 The EU AI Act’s transparency rules take effect on schedule. The high-risk rules due alongside them do not

Article 50’s disclosure and labelling duties became enforceable across the EU exactly as planned. The tougher Annex III obligations for high-risk systems, delayed by June’s Digital Omnibus, now land in December 2027. California’s own AI content-disclosure law took effect the same day, deliberately timed to match.

⚡ 05 OpenAI’s next model solves ten open problems in mathematics, with proofs a computer can check

An internal version of Astra produced machine-verified Lean 4 proofs for ten unsolved problems across eight fields, including a 27-year-old open question in group theory, for around $2,000 in compute. Every certificate is public on GitHub under an open licence.

Receive one expert articles and our weekly briefing by subscribing below.

Subscribe now

⚡ 01 Insure the agent, or supervise the work: two answers to the same question

Legal AI’s hardest unresolved question isn’t whether agents can do the work. It’s who answers for it when they don’t. This week produced two very different answers, from two very different layers of the market.

Crosby, the venture-backed AI law firm that reviews contracts in under an hour, announced it will buy professional liability insurance for its own agents, so they can complete legal work without a human lawyer signing off first. Founder Ryan Daniels was direct about where this is heading. Lawyer review of AI output, he said, won’t be necessary in the future. It remains, for now, the standard the rest of the industry still runs on.

The logic is coherent on its own terms. If the firm is financially on the hook when an agent gets it wrong, in theory it has every incentive to make sure it doesn’t. But it moves the mechanism of trust from verification to indemnity. Nobody checks the work before it goes out. Somebody just pays for it afterwards.

The software layer kept moving, unbothered by the accountability question

Further down the stack, the tools carried on at their usual pace. Wordsmith closed a $14M extension to its Series B, taking its total raised to roughly $114M. Ironclad shipped new agents that scan supplier contracts for buried rebates and compliance obligations and route them straight into SAP. Both are incremental tool improvements. Neither one changes who is answerable when the output is wrong, because in both cases a person is still expected to check it before it matters.

The second model is slower to describe and harder to fit in a press release. It’s also the only one of the two where the client, not an underwriter, still owns the standard the work is held to.

The question worth asking. If your agents got something wrong tomorrow, would you rather find out from your own supervision layer or from your vendor’s insurer? One of those answers means the work never reached the business. The other means it did.

⚡ 02 Harvey stacked up procurement signals this week. What they actually establish is a narrower question.

Harvey made three announcements in four days: an AIUC-1 certification, growth capital from two bank investment arms, and a new chief revenue officer. All three are aimed at the same audience, the enterprise buyers who decide whether software gets trusted with real work, and each is worth reading for what it establishes rather than for how it was framed.

On August 4, Harvey became the first legal AI company certified against AIUC-1, a standard for AI agent security, safety and reliability that has existed for barely a year. The auditor, Schellman, ran more than 3,000 tests across six domains: data and privacy, security, safety, reliability, accountability and societal impact. Harvey reports zero critical failures.

The framework maps onto standards enterprise security teams already use to vet vendors, including ISO 42001, the EU AI Act and NIST’s AI risk management framework, which is precisely what makes it useful as a sales document: it pre-answers procurement’s security questionnaire. What it can’t yet claim is a track record. No one knows whether an AIUC-1 pass predicts how an agent behaves in production, because the standard is too new for anyone to have checked.

The certification landed a week after Growth Equity at Goldman Sachs Alternatives and J.P. Morgan’s Growth Equity Partners took a stake of undisclosed size, on the back of a quarter in which Harvey says it added more than $100M in ARR. On August 5, Harvey named Steve Zad, previously of Rubrik, as its first chief revenue officer. These are the ordinary mechanics of a growth-stage software company preparing to sell more software. They say a lot about Harvey’s go-to-market ambitions and nothing about whether the work a legal team hands over comes back done.

What certification actually tells a buyer

AIUC-1 tests whether the agent is safe to run: data handling, security posture, failure modes, whether the system behaves predictably under adversarial input. That is infrastructure due diligence, not legal due diligence. What it does not test is whether the work is right. Nothing in the standard asks whether an NDA got redlined to your fallback position, or whether the contract left the queue at all. That’s a different audit, and nobody has built it yet.

The question worth asking. A safety certification tells a buyer the tool won’t misbehave. It doesn’t tell them the work got done to their standard, or that it left the desk of someone too expensive to be doing it. Those are the questions a services relationship has to answer, and no compliance framework answers them by proxy.

🏗 04 Two continents ran the same trick, on the same day

On August 2, the EU and California both let an AI transparency deadline land exactly on schedule, while quietly waving through the harder substantive rules that were supposed to arrive alongside it.

In the EU, Article 50 of the AI Act, the provision requiring disclosure when content is AI-generated and labelling for deepfakes and chatbots, became enforceable across the bloc precisely when the original text said it would. It wasn’t touched by June’s Digital Omnibus. What the Omnibus did move is Annex III, the high-risk system obligations covering things like biometrics, employment screening and credit scoring. Those now apply from December 2027 for standalone systems and August 2028 for high-risk systems embedded in products, more than a year later than the law originally promised. The Act’s most consequential paperwork requirement showed up on time. Its most consequential compliance burden did not.

California ran an almost identical play, on the identical day, on purpose. SB 942, the state’s AI Transparency Act, requires large generative AI providers to offer a free content-detection tool and embed provenance watermarks in AI-generated images, audio and video. Its operative date was expressly amended to land on August 2, the same day as the EU’s transparency rules. Two jurisdictions that otherwise regulate AI on almost nothing in common chose the same date for the same category of obligation: tell people what’s AI-generated. Neither one moved the harder question, what an AI system is allowed to decide on its own, at the same speed.

So what. Disclosure regulation is cheap to comply with and politically easy to enforce on schedule. Substantive regulation, telling a system what it can’t decide on its own, is neither. For a legal team building an AI governance programme, that gap is the actual risk map. The labelling rules are real now. The rules that would have forced a hard look at autonomous decision-making just bought everyone another year and a half.

⚡ 05 The same week, AI got smarter and less predictable

Three labs shipped three different kinds of news this week, and the gap between them says more about where the industry actually is than any single release does.

On August 1, OpenAI published something unusual for a model that hasn’t shipped: proof. An internal version of its next model, Astra, resolved or made substantial progress on ten open problems across eight fields of mathematics and theoretical computer science, including a 27-year-old question in group theory and two long-standing Erdős problems in combinatorics.

What made the announcement notable wasn’t the difficulty of the problems. It was the form of the answer. Every result came with a machine-checkable Lean 4 certificate, published openly on GitHub, so a mathematician doesn’t have to take OpenAI’s word that a proof holds. They can run the checker themselves. Fields medallist Timothy Gowers said he would recommend one of the results for a top journal without hesitation. The whole exercise cost roughly $2,000 in compute.

Meanwhile, Google’s most anticipated model of the year still hasn’t shipped. Gemini 3.5 Pro was announced in May, targeted for June, then July, then a specific mid-July date that also passed. As of this week it’s rumoured for August 12, unconfirmed by Google, which would make it the fourth slipped date for a single release. xAI is running the opposite playbook. Grok 4.5 launched three weeks ago, and Elon Musk says Grok 4.6 is coming within days, a major-version cadence measured in weeks rather than the months every other lab treats as normal.

The question worth asking. If your legal AI vendor is built on someone else’s model, which of these three release patterns are you actually exposed to? A verified proof, a missed date, or a version bump you didn’t ask for? Almost no legal AI buyer asks their vendor this question, and almost every vendor’s honest answer changed at least once this year.

🧠 What this week tells us

Every story this week is, underneath, the same story. The industry is racing to prove AI can be trusted with real stakes faster than anyone has agreed on what that trust should actually rest on.

Verification is the scarce resource. Astra shipped a proof a machine can check. Harvey shipped a certificate an auditor can check. Neither one checks the thing a legal team actually needs checked: the work itself. Every mechanism the market produced this week verifies the tool, and none of them verifies the output.

Regulators regulated the easy part. Disclosure and labelling rules landed on schedule in the EU and California, on the same day, by design. The harder question of what autonomous systems can decide unsupervised got pushed to December 2027 for standalone high-risk systems and August 2028 for embedded ones. Anyone building an AI governance programme should read that gap as the real risk map.

Accountability is splitting into two models. Insure the agent and let it run, or supervise the work and catch the miss first. Crosby picked the first and was refreshingly honest about it. Most of the rest of the market, so far, is still picking the second, mostly because nobody has offered them a credible alternative.

Capability keeps outrunning shipping discipline. One lab proved it can solve problems mathematicians have worked on for decades, and published the certificates so anyone can check. Another can’t hit a release date it set itself three times running. A third is shipping major versions on a weekly cadence with far less independent scrutiny. If your vendor sits on top of one of those, you have inherited its release pattern whether you priced it in or not.

✳️ The Flank view

Every certification, insurance policy and disclosure rule this week was, in its own way, an attempt to answer the question the market still hasn’t settled: who is responsible when AI does real legal work. None of them touch the problem underneath it, which is that inexpensive, repetitive work is still being done by the most expensive people in the building, and a badge on the tool doing it faster doesn’t change that. Harvey’s AIUC-1 certification says the agent is safe to run. It says nothing about whether the NDA left a general counsel’s desk.

Crosby’s answer, insuring the agent so it can work unsupervised, is at least honest about the stakes. But it transfers risk rather than removing it, and it keeps the work inside a vendor’s four walls rather than the client’s own. Insourcing enterprise legal work to supervised agents is a different bet: the client’s own playbooks, the client’s own escalation rules, and a supervision layer, run by the client’s lawyers, a partner firm, or Flank’s own team, that catches the exception before it ever reaches the business. Tools are a commodity. A certified tool is still a commodity. Outcomes, work that leaves the desk and doesn’t come back, are not, and nothing published this week claimed to deliver one.

✳️

This post is AI-generated and screened by a Flank team member.

Share

Hi, I'm Lorna and I've spent my legal career trying out different hats until I ended up designing my own.

At Flank, I lead the Legal AI Alignment function.

Figuring out what AI can’t do for me

I’ve had to ask myself some hard questions lately.

Where do I see the legal profession going? How has it changed over the last 15 years and how is it changing now? Where do I see myself fitting in?

I’m not going to fool anyone by pretending I have a tidy answer to the last question. That said, I do have a much clearer view of where a lawyer’s value actually sits now, and it starts with the blunt acceptance that it’s genuinely not where I spent most of my twenties assuming it was.

When I started out as a law student, I remember feeling daunted, uncertain and a bit of a misfit. Everyone else in my class seemed to have a clear trajectory planned - get a first, secure a training contract, qualify, specialise, and either get on track for partnership or eventually go in-house to become a general counsel. The handful of alternative paths you’d hear about were legal director, knowledge lawyer or innovation specialist.

Between my work as a barista during the day, library clerk on the weekends and conference assistant during the holidays, I had little bandwidth left to strategise about what it is I actually wanted to do after I left the cosmopolitan bubble that the University of Warwick was known for. Instead, I went for breadth and took any opportunity that came my way - doing research for the student TV, organising fundraisers for charities, a sponsored undergrad research programme into foster care, writing a dissertation on feudal systems, volunteering at the law centre’s dedicated unit for asylum seekers, to name a few examples.

Given my, let’s call it, eclectic background, I always struggled sitting down to write a convincing application to any law firm making my case for why I’d be a good fit for a training contract. I struggled convincing myself of that narrative, to be honest, so I’m not surprised that I kept getting rejection after rejection. Interestingly though, each rejection felt a little comforting - a confirmation that I didn’t belong or want to belong in the law firms I was applying to. However, this all changed when I came across Pinsent Masons. I felt I’d found a firm that seemed to tick all the boxes for me - friendly people, family friendly and invested in innovation. There was no stopping my resolve from that point on. Not even the initial rejections wore me down. Three whole attempts later, I was (at long last!) firmly locked into a training contract.

I found myself repeating the same patterns once I joined private practice. Looks like my modus operandi was pretty much established early on. Call it insatiable curiosity, impatience, short attention span or a combination of all of the above - I cannot help but try my hand at anything that interests me. Over time, this meant working with teams all over the country, across various practices (litigation, advisory, innovation) and taking on client secondments in completely unknown territory (like reinsurance, for example!). It was soon clear to me that I was not ready to commit fully to any one area of specialism. I was still searching for my identity as a professional. I envied people who could command a room on a particular topic and go deep - citing case law with precision and taking questions from the audience with the casual air of a seasoned professor.

When I went in-house after five short years in private practice, it felt like a breath of fresh air at first - shedding the label (and pressure) of my subject matter expertise (in my case, it was pensions law). I soon realised, though, that in focusing so hard on being the expert in the room during my early years in private practice, I had completely neglected the best part of being a lawyer - the human relationships. Throwing myself into an in-house role at a fast-growing fintech business gave me a taste of human partnership - the legal team is here to enable, to support, to protect (and yes, sometimes that means strongly recommending against doing something the business really really wants to do) no matter whether the initial advice was followed or not. Legal is a true ally of the business.

Unfortunately, we get a bad rep because, guess what, knowledge work requires a lot of thinking and precision, both of which take time and often surface some uncomfortable realities. We then end up with the label of ‘blocker’, or more commonly, ‘bottleneck’. You can tell that we really internalised this unfortunate label in my previous organisation where we used to call ourselves the ‘YES team’ to reclaim the narrative and share the messaging that, unlike what you might have heard about our kind, we were not going to be the ‘Department of No’.

However, Legal doesn’t have to be a yes-man nor a permanent stop sign. It can be the judgment layer that sits behind the decision, once we have built ourselves a filter for the noise - so the business can make decisions based on a full picture, knowing exactly who is accountable in that process too.

I learned very early on that a large part of why there is so-called friction in legal work is not down to lawyers being unnecessarily pedantic about contractual clauses. An incredible amount of time and effort goes into the absolute basic steps before any serious thinking can even start. Let me give you a couple of examples:

• Avoidable friction: Your CEO messages you about a customer emailing to terminate a contract. No other context provided, except for a screenshot of said email. You now need to go and find said contract, review the email thread, ensure there were no amendments made after this filed version of the contract was signed, speak to the person who owns the customer relationship to check the background to the email in the first place. A whole lot of investigation needs to take place off the back of one simple query. Almost all of this melts away if you have a contract record you actually trust.

• Necessary friction that’s probably here to stay: Your sales lead reaches out about an obscure question that a prospective customer has asked about your product architecture and how that’s mitigated by a particular provision in your contract. You need to go and speak to an engineer, check with your security counterpart, review the contract, assess the prospect’s actual concern and sense check that any response you share is aligned with your vendor due diligence responses. No tool will fill in for decisively knowing which engineer to ask, what the prospect is really worried about underneath the question they asked and the boundaries that your business is comfortable stretching to without reckless overexposure.

Legal work is complex and layered. It frustrates me when legal work is reduced to a binary state and when this appreciation for complexity is completely lost. Most of the friction actually comes from the preparatory steps we have to take before we can even start to think about the problem at hand meaningfully. Some of these steps can be avoided entirely (for example, if the contract database was up to date and reliable) but some steps cannot be avoided nor can they be skipped because even fact-finding requires specialist skills - you need to know what to look for and what questions to ask.

When I think about my relevance today, this is what I come back to. I think that a lawyer’s value lies in connecting the dots between information points and sharing their informed judgment. That means we need to be ruthless about eliminating any noise that stops us from doing this, but we cannot delegate the connecting of the dots itself. We cannot outsource this part of the thinking to AI because AI does not see AND feel everything that we think, see and feel.

It’s quite scary to work with a tool that knows a lot more than you do, especially when your entire USP to date has centered around being a subject matter expert. But expertise is changing shape. It’s about increased visibility rather than better execution. We have to hone our judgment, our taste-making, by leveraging tech to remove noise so our senses are sharpened. I don’t want to be an AI-assisted lawyer who is relying heavily on AI to do legal work faster - a) I find fun in the thinking part of the job, so this is a guaranteed fast track to boredom and b) what is the point of automating myself away so willingly without backfilling any improved value add?

There are a lot more pathways open in the legal profession than there used to be even five years ago. If no hat fits, there’s also never been a better time to design your own - legal engineer, alignment lead, AI specialist, AI governance expert, legal AI expert, fractional GC, fractional legal ops lead. The misfit problem sorts itself out once the paths multiply. I know that because the job I do now didn’t have a name when I was writing those training contract applications. The legal profession has never looked more unsteady, in a good way.

Where do I see myself fitting into this new version of the legal profession that’s actually still in formation?

Well, I’m still figuring it out. I know that I don’t want to be a human AI wrapper, nor do I want to be a technician purely deploying and maintaining AI systems that can do my job better than me. What I do know is that I want to do everything that I possibly can to preserve my uniquely human qualities - whether that’s sharing the camaraderie that forms when you’re getting a ridiculously protracted negotiation over the line, or the bonding that happens when you’re dealing with a security incident and need to reassure the business, or the instinct that kicks in when you’re mid-dispute and you sense exactly what the other side cares most about.

✳️

Share

Jake here, Co-founder and COO at Flank.

Most CLMs are not reliable. I’ve had that conversation more times than I expected to. The following conversation went further than others, and it changed what we decided to build.

A change to the schedule: Starting next week The Intake now publishes once a week, on Wednesdays, rather than twice as so far. Every article is written by hand by its author. This week in legal AI continues on Fridays.

Subscribe now

📄 The legal ops leader who would rather use a folder

Just a few months ago I realised no one trusted their CLM.

We were onsite at a new client, working on mapping the contracting process for one of their lower risk contracts. The end goal, to implement an agent to run that process. Fun, “bleeding edge” stuff, right? But the regional head of legal ops kept interrupting the workshop to complain about how other regions ran their version of the process.

And you should see our CLM, she said. Only a fraction of our contracts even land there… I’m thinking of packing it in and moving everything to SharePoint.

Whatever I was drawing on the whiteboard in that moment, I stopped.

To SharePoint?, I asked, trying not to sound judgemental or aggrieved. But seriously, transitioning from a purpose built tool to a folder?

So she explained, and her reasoning was sound. The CLM was a legacy implementation only a handful of people knew how to use. And a decent share of the company’s contract never landed in it anyway. Each region had their own flow. Signed contracts lived across email, Agiloft, SharePoint, DocuSign, and—inexplicably—a filing cabinet in the Rotterdam office.

Share

🔍 It turns out this is the norm

Later that night, back at my hotel, after a few drinks, I emailed another client who we’ve been working with for a couple of years.

T— A client today told me their [sic] CLM is functionally useless. No one using it. An entire team required to maintain/adminstrate. Legit?

I received a response a few moments later.

Yes.

I feverishly continued to dig, and I found industry numbers that suggest this is indeed the norm. World Commerce & Contracting puts the cost of poor contract management at around 9% of annual revenue, rising to 15% or more in complex industries, and its 2025 research found contract-related data scattered across an average of 24 different systems per organisation. Then, a widely cited Journal of Contract Management figure has 71% of companies unable to locate at least a tenth of their contracts.

A tenth of all contracts. Vanished. In a puff of smoke.

The next day, I was back at the whiteboard. But this time, I couldn’t get my mind off what M— had said the previous day, about her planned moved to SharePoint, and T—’s single word confirmation. I couldn’t shake it off.

Share

Her CLM was not the problem

M— I said. Sorry, but that CLM issue you described yesterday…

She glanced up at me eagerly.

I said, Is your CLM really so bad? Did you try others before you committed to this?

She looked puzzled. Our CLM is great, she said. We just we can’t trust anyone to use it right, and we don’t want to hire a team of contract managers to administer the thing seven days a week. And even if we did, it’s a recipe for human error.

The problem wasn’t the software. The problem was that the CLM had arrived with an invisible line item that no one had budgeted for: a team to run the thing round the clock, flawlessly. We continued the discussion over lunch with one of her legal counsels who had joined the company mere weeks earlier. He described a similar situation at his previous employer. A well built CLM with no team to run it. There, they had flirted with the idea of a dedicated team of contract managers to administer the system; they had even experimented with three of the legal ops team taking on this exact job, but the endstate soon became obivous: an expensive team who the business would become overreliant on, running an expensive tool no one else would understand.

A few days later on the flight home, I sent a short note to my co-founder, Problem with CLMs is not the software. It’s the implicit human work required to make the CLM trustworthy

Share

Keeping the record honest is a job nobody has

This is what it boiled down to. Throughout my research. In every discussion with every disillusioned customer of a legacy CLM. Trust was lacking because keeping the record honest is work. Expensive, human labour. Filing, tagging, updating, reconciling metadata against the actual reality of the negotiation, deduplicating, correcting typos. And the work never stops. It’s endlessly snowing new contracts and amendments. Renewals and renegotiations. And hardly any of this is legal work. It’s administrative. Often mind-numbing, inane, repetitive administrative work that is either done by well trained (and expensive) legal folk or else by no one at all.

Two bad options

Enterprises are left choosing between two bad options.

1. Hire a dedicated contract management team to run the CLM, and to ensure all signed agreements actually land in the CLM

2. Live with the CLM being untrustworthy, and watch that record creep further away from the truth each day until the system becomes unusable noise

So we built a third option.

3. Hire an agent to manage your contract record.

If you’d like to see it: flank.ai/flank-record

We’re ingesting everything. Even the filing cabinet in Rotterdam.

✳️

Images (but not text) are generated using ChatGPT.

Hi there, I’m Lorna. I lead Legal Alignment at Flank.

Every legal AI conversation arrives at the same question eventually: where’s the return?

I think it’s sitting in the part of the plan nobody wrote. The stretch after go-live, with no owner and no budget line.

The roadmap that ends at go-live

I’ve lost count of the AI delivery plans I’ve seen that end at rollout.

They’re often impressively detailed up to that point. Discovery, scoping, playbook building, testing, pilot, go-live, adoption workshops. Then the document simply stops, as if adoption were the finish line.

On more days than not, my work is the part that comes after that line. So I can tell you with some confidence that the go-live milestone is actually the starting line. Everything that determines whether an agent earns its keep happens after it.

Think about it. No business stops operating on launch day. Launch is when the real work begins - the customers arrive, the assumptions get tested, the operating model proves itself or doesn’t. We all know this instinctively about businesses. Somehow we forget it about AI agents.

Personally, I think this delivery-heavy planning habit explains a lot about why legal AI keeps disappointing people. Too much energy goes into testing and deploying. Not nearly enough goes into what happens next.

🏗️ You’re not buying legal tech, you’re hiring a new way of working

That’s because deploying an AI agent is a different kind of transaction to buying tech.

Traditional software is finished when it ships. You configure it, you train the users, and the vendor’s job is largely done. The tool does the same thing on day 400 as it did on day 4, whether you love it or not.

An agent doesn’t work like that. An agent is hired to do a job - review the NDA, turn around the supplier contract, tell sales whether they can accept 60-day payment terms on the deal that’s mid-negotiation. And how well it does that job depends on things that only exist after deployment: the playbook it follows, the feedback it receives, the escalations it raises and how those get resolved.

Share

I always like to think of it as onboarding rather than installation. A new joiner doesn’t arrive fully formed. They arrive with strong general skills and close to zero knowledge of how your organisation actually likes things done. The first months are corrections and context. Nobody considers that a flaw in the hire. That’s simply how capability becomes useful.

Agentic AI is a whole new way of working. It’s a new system inside your organisation, and like any system, it needs a solid build followed by regular maintenance. Skip the second part and the first part slowly stops mattering.

In fairness, the delivery focus makes sense

I want to be fair to the people writing those roadmaps, because the delivery focus didn’t come from nowhere.

A launch has everything an organisation knows how to fund. It has a date, an owner, a budget line and a ribbon to cut. If you’re a general counsel, your board wants to hear about AI progress, and “we’ve gone live” is a clean answer to give them. If you run legal ops, you’re likely measured on delivery milestones, and the roadmap reflects exactly what you’re being asked to prove.

Maintenance has none of those things. It has no end date, no obvious owner and no moment of glory. Nobody gets promoted for tending something that’s already live, and no budget template I’ve ever seen has a line for “keep teaching the agent”.

And procurement is built around the old model. You evaluated the vendor, negotiated the contract, ran the implementation. When you bought your CLM or your e-signature tool, nobody stood up a permanent team to keep tuning it. Why would agents be any different?

However, it’s crystal clear to me that agents are different, for the reason above. Their performance is shaped by what happens after go-live, not before it. Applying the buy-tech mental model to a hire-a-new-way-of-working problem is how organisations end up with agents that launch well and then drift.

🔍 What AI agent maintenance actually looks like

Day-to-day, maintenance is less glamorous and more interesting than it sounds.

It means scrutinising performance at scale, not just spot-checking the demo cases. It means collecting feedback from the people actually using the agent and iterating the playbook based on what they say. And it means measuring impact honestly, including the uncomfortable questions. Is human involvement genuinely tapering off? If not, why not?

Share

An example from our own operations at Flank. We run agents on our own standard contracts - events, sponsorship deals, NDAs, supplier agreements. No matter how comfortable I am with our internal playbooks on those documents, real end users interacting with the agent directly will always surface friction points that would otherwise have been absorbed by a human lawyer without anyone noticing.

Questions phrased in ways no lawyer would phrase them. Requests the playbook never anticipated, because a human lawyer would have handled them on autopilot and never thought to write them down. Before the agent, that knowledge lived in someone’s head and stayed there.

That’s the continuous learning loop the roadmaps leave out. And it’s the mechanism by which the agent becomes worth what you paid for it.

So what does success look like? I would describe it as a taper curve. In the early months, the humans in the loop make corrections often. Correction here rarely means the agent got something wrong, by the way. More often it means adjusting base positions, adapting the agent to your actual preferred negotiation posture on particular types of deals, proper tailoring to your organisation. Say your playbook holds firm on a 12-month non-solicit as standard, but on low-value sponsorship deals the business is relaxed about conceding it. The agent has to learn that posture, and the only way it learns is through the humans in the loop teaching it. In a word: training.

Over time, those corrections should genuinely reduce. Success is when human involvement sits at escalation points only - the moments where human judgment is needed. Even then, you keep a system for assuring quality throughout, like regular spot checks on samples of the agent’s output. The taper is earned, then verified on a schedule.

That's the continuous learning loop the roadmaps leave out.

As for who owns all of this: the legal function does, and that means legal ops just as much as the lawyers, wherever a legal ops team exists. I want to be clear that this is not another burden to load onto an already stretched legal team. Ultimately, agentic systems need a cockpit to be operated from, and someone at the controls. At Flank we have the benefit of a dedicated team for it - a legal AI alignment team, focused on aligning what the AI does to the organisation’s actual needs. Personally, I think every organisation deploying agents into legal work will eventually have one, whether it’s built by the lawyers, run by legal ops, or borrowed from the vendor. The name matters less than the cockpit existing at all.

And if I’m honest, I think this is the opportunity sitting in plain sight. For a GC, the alignment function is how legal AI becomes something you can stand behind in front of the board, with evidence rather than anecdotes. For a legal ops lead, owning the learning loop means owning the value story - you become the person who can actually show what the AI investment is returning, in a market where almost nobody can. That’s a far more interesting position to be in than being the person who ran the implementation.

This is where the legal AI ROI went

Now connect that back to the question everyone in legal AI keeps asking. Where’s the return?

MIT’s Project NANDA report, The GenAI Divide: State of AI in Business (August 2025), made headlines with the finding that 95% of enterprise GenAI pilots delivered no measurable P&L impact, despite $30 to 40 billion of investment. The methodology got picked apart, and fairly, but the pattern it described matched what many of us see on the ground: high adoption, low transformation. Tools that demo well and then sit brittle inside real workflows.

Share

Closer to home, Thomson Reuters reported in 2026 that 82% of legal departments either don’t measure AI return on investment or don’t know whether they do. In other words, most legal teams have no learning loop running on their own investment either.

Imo those two findings are the same finding.

A pilot that ends at deployment has no mechanism for compounding.

Usage spikes at launch, then drifts as the un-ironed friction points pile up. Workarounds creep back in. Eighteen months later the tool gets blamed for underdelivering, when what actually underdelivered was the plan around it. Meanwhile, the deployments that do generate returns are the ones where somebody kept scrutinising performance, kept collecting feedback, kept iterating. The MIT authors found the winners were workflow-integrated systems built with partners who stuck around after launch. That’s maintenance by another name.

The timing matters here too. Thomson Reuters also reports that only 15% of organisations currently use agentic AI, while another 53% are planning or evaluating it. That’s a very large number of delivery plans being written right now. On current form, most of them will end at go-live.

For a legal ops lead, owning the learning loop means owning the value story - you become the person who can actually show what the AI investment is returning, in a market where almost nobody can.

A test you can run this week

If any of this stings, there’s a check that takes five minutes.

Whether you’re the GC who signs the budget or the legal ops lead who writes the plan, pull up your AI roadmap, or the vendor’s delivery plan, or the business case that got the budget approved. Find the go-live line. Read what’s written after it.

If the answer is a renewal date and not much else, you now know where your ROI is going to go.

What should be written there instead? Ownership of the learning loop, for a start. Who reviews performance, who collects end-user feedback, who iterates the playbook, and how often. A definition of what tapering human involvement should look like for your use cases. A spot-check schedule that survives after the novelty wears off.

I won’t pretend the steady state is a solved problem. I can describe the taper curve, but I can’t yet tell you how long the taper should take for a given organisation, or at what point a taper that isn’t happening means the product is wrong rather than the playbook is simply young. We’re learning that in real deployments, one feedback loop at a time (which is rather the point).

The direction, though, I’m sure of. Most agent deployments that disappoint were never going to be rescued at launch. The gap sits in the long, unbudgeted, unowned stretch after go-live - and almost nobody is planning for that part yet.

Lorna Khemraz is Lead Counsel at Flank and leads its legal AI alignment team, which exists to align what agents do to what the organisation actually needs.

✳️

Subscribe now

⚡ 01 The White House accuses Moonshot AI of distilling Anthropic’s Fable to build Kimi K3, and of using export-banned Nvidia chips to do it

On July 23, White House science policy director Michael Kratsios accused the Beijing-based lab Moonshot of covertly distilling Anthropic’s Fable model, running the copying through a purpose-built internal system to stay hidden, and of accessing export-controlled Nvidia chips without a licence. No penalties have been announced. Kimi K3’s full weights are due to go public on July 27.

⚡ 02 Crowell & Moring publishes six months of firmwide Legora data: 2 million-plus interactions, 91% of its 700+ attorneys onboarded

Eighty-two percent of the entire firm has become a Legora user since the January 2026 firmwide launch, with nearly 70% of attorneys using it at least weekly. It is one of the more granular adoption data sets any large firm has published this year.

⚡ 03 Australia’s Northern Territory backs a $28B gas-powered AI data centre campus

Beetaloo Digital, a unit of a shale-gas producer sitting atop the Beetaloo Basin, plans two hyperscale campuses drawing up to 2 gigawatts, burning the region’s own gas reserves to power the compute.

⚡ 04 Google ships Gemini 3.6 Flash and 3.5 Flash-Lite, and teases Gemini 4, while Gemini 3.5 Pro still hasn’t shipped

3.6 Flash cuts output pricing by roughly 17% and posts double-digit benchmark gains over 3.5 Flash. A dedicated security-focused variant, Gemini 3.5 Flash Cyber, also ships. The mid-tier update lands quietly. The flagship Google has teased for months still doesn’t exist in production.

⚖️ 05 A federal judge in Michigan warns the DOJ over an apparent AI-hallucinated citation, but declines to sanction

Judge Hala Jarbou finds a government brief cites a Sixth Circuit case, Taylor v. Hott, that does not exist. She warns it “goes without saying” that federal filings can’t contain hallucinated law, but stops short of sanctions.

Receive two articles and our weekly briefing by subscribing below.

Subscribe now

⚡ 01 The distillation dispute

The White House says a Chinese lab built its newest frontier model by secretly copying Anthropic’s outputs. If that’s true, it’s a legal question this briefing has never had to ask before: what happens when the theft is the training data?

Model distillation, feeding a stronger model’s outputs to a weaker one so it learns to imitate the stronger one, is a normal, often licensed technique in AI development. On July 23, White House Office of Science and Technology Policy director Michael Kratsios accused Moonshot AI, the Beijing-based lab behind Kimi K3, of doing it covertly: running the copying through a purpose-built internal system and rotating access routes to stay hidden while pulling outputs from Anthropic’s Fable model. Kratsios also said Moonshot obtained access to export-controlled, high-end Nvidia AI servers to train the model, hardware barred for export to Chinese entities without a licence. Anthropic itself had flagged unusual activity back in February, tracing 3.4 million Claude exchanges to the Chinese startup. Moonshot has not responded to the allegations. No penalties have been announced.

And the punchline is almost absurd: Kimi K3’s own public launch a few days earlier was so popular that Moonshot had to pause new subscriptions within 48 hours, “the love,” as the company put it on social media, outstripping its GPU capacity.

The structural question for enterprise legal teams. Every enterprise legal AI contract has a clause, somewhere, about what the vendor can and can’t do with your firm’s inputs and outputs. This dispute is the first widely reported case of a frontier lab allegedly doing exactly that to a rival, at industrial scale, in violation of usage terms. If a well-resourced state-linked lab can allegedly do this to Anthropic and still ship a product before anyone can prove or stop it, what confidence should an enterprise buyer have that its own outputs, contract clauses, negotiating positions, and internal analysis aren’t being harvested by whatever model sits underneath a vendor’s product today? Provenance of training data is becoming a genuine legal due-diligence question, not a theoretical one.

The Flank read. Whether or not the specific allegations against Moonshot hold up, the incentive they describe, that outputs are valuable enough to steal at industrial scale, applies just as much to the outputs your legal team generates inside whatever AI tool it uses today. Flank’s model doesn’t ask a client to trust a single model vendor’s training practices as a matter of faith. Supervision sits above the model layer: a human reviews every agent output before it reaches the business, and the workflow, not any one underlying model, is what the client is actually buying. When the model layer is this volatile, the thing worth building trust around is the review step, not the vendor’s word about what its model was trained on.

Share

⚡ 02 The firm that published its numbers

Last week this briefing covered a survey showing 83% of in-house teams can’t prove their AI spend paid off. This week, one Am Law 100 firm published exactly the kind of granular usage data that survey found almost nobody has.

Crowell & Moring piloted Legora in late 2025 and launched it firmwide in January 2026. On July 22, the firm published six-month results: more than 2 million platform interactions, 82% of the entire firm, including 91% of its more than 700 attorneys, now counted as users, and nearly 70% of those attorneys using the platform at least weekly. That last figure is the one that matters most. Adoption numbers are easy to publish and easy to inflate with one-time logins. Weekly active usage at that depth, across a firm of that size, is a different claim entirely: evidence of the tool being embedded in how the work actually gets done rather than sitting unused after a launch announcement.

An associate in the firm’s Patents Group described the platform as having “accelerated my learning curve and, therefore, value to clients in ways I didn’t anticipate.” The firm also cited a recent trial where the team used the platform to prepare more thoroughly and respond more nimbly than an opposing team with more people on it. Both are anecdotes rather than measured outcomes, and Crowell & Moring didn’t publish a cost or time-saved figure to sit alongside the usage data, which is itself notable given how squarely last week’s ROI-measurement gap sits over this exact kind of announcement.

The structural question for enterprise legal teams. Weekly-active-usage data like Crowell & Moring’s is genuinely useful, and rare. But usage depth answers a different question than the one Axiom’s survey found 83% of in-house teams can’t answer: not “are people using the tool,” but “does the work cost less or get done better because of it.” If your own legal AI vendor can’t produce Crowell & Moring’s kind of usage number, and your own team can’t produce a cost or outcome number to go with it, which of the two gaps is actually the harder one to close?

⚡ 04 The model churn

Three models moved this week: one shipped quietly, one is still vapourware after months of teasing, and one got so popular it had to turn users away. None of it is visible from inside a legal AI product’s chat window.

Google shipped three new Gemini models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a narrow, security-focused variant called Gemini 3.5 Flash Cyber. 3.6 Flash cuts output token pricing by roughly 17% against its predecessor and posts double-digit gains on coding and reasoning benchmarks, but Gemini 3.5 Pro, the flagship this briefing has now heard teased across two editions, following a full architectural rebuild after the original base model was scrapped, still has not shipped. Meanwhile Moonshot’s Kimi K3, a 2.8-trillion-parameter open-weight model, launched the same day and proved so popular that Moonshot paused new subscriptions within 48 hours to protect existing users’ access, before the distillation allegations above had even surfaced.

The structural question for enterprise legal teams. Three models, three different states, shipped, promised, and overwhelmed, inside a single week, and almost none of it is visible to the person typing into a legal AI product’s chat window. Does your legal AI vendor tell you when the model underneath your workflow changes, and does your audit trail record which model produced a given output? A flat interface hides model churn by design. A governed workflow has to surface it, because the model that drafted last month’s NDA redline may not be the model that drafts next month’s.

⚖️ 05 The courts split on mercy

Two federal judges, in the same week, looked at lawyers who filed AI-hallucinated citations and reached opposite conclusions about whether a warning is enough.

This briefing has spent months tracking an escalation: sanctions climbing from four-figure fines toward $15,000-per-attorney penalties, bar suspensions, and disqualifications. This week ran in the other direction in two separate cases. On July 17, Michigan federal judge Hala Jarbou caught a DOJ brief citing Taylor v. Hott, a Sixth Circuit case that does not exist, and warned the government that hallucinated law in federal filings is unacceptable, but declined to sanction. Around the same time, Kentucky federal judge Thomas Cullen declined to sanction attorney Thomas Guyer over a brief filled with AI-generated misquotes and incorrect citations, finding Guyer had “owned the mistake,” had no history of misconduct, and was, in his own lawyer’s words, “incredibly remorseful.” Cullen was explicit that a warning could serve as “sufficient deterrent” on a clean record, while still insisting that generative AI’s growing role in practice, its becoming the “new normal,” doesn’t relax a lawyer’s duty to take reasonable measures to verify what gets filed.

The escalation this briefing has tracked ran the other way: a New York firm ordered to pay $10,500, a Sixth Circuit panel fining two attorneys $15,000 each, bar suspensions and disqualifications following repeat or egregious fabrications. The restraint this week introduced is narrower: a government attorney warned, not sanctioned, over a fabricated Sixth Circuit citation in an immigration case, and a solo Kentucky attorney with a clean record spared formal sanction in favour of a warning. Two judges, independently, treated a first offence with contrition as different in kind from repeat or careless disregard.

Neither case suggests courts are going soft on the underlying problem. Both judges were explicit that verification remains the lawyer’s job regardless of the tool. What they introduced is a distinction between the mistake and the response to being caught: remorse, a clean record, and taking ownership now appear to buy real leniency, in a way that a pattern of hallucinations, or a lawyer who fights the finding, does not.

The Flank read. The distinction these two rulings are drawing, first-offence contrition versus a pattern of disregard, is itself a supervision question, not a sanctions question. A firm that can show a court it has a real verification step built into how AI-assisted work gets produced and checked has a materially different story to tell than a firm that can only say a lawyer was sorry after the fact. That is the entire premise behind building supervision as core product rather than as a policy memo: every Flank agent workflow runs through human review before anything leaves the system, so the question a court is now implicitly asking, “was there a real check here, or just good intentions,” has an answer that doesn’t depend on how convincingly remorseful anyone sounds afterward.

🏗 The deadline that wasn’t deferred

This briefing has tracked the EU AI Act’s high-risk deadline sliding from April 2026 to December 2027 across four separate editions. The deadline that actually lands in nine days is a different one entirely, and it applies to almost everyone.

The political fight this briefing covered across April and May, over whether the EU AI Act’s high-risk compliance deadline would hold at August 2026 or slip, is over: the Digital Omnibus on AI was signed into law on July 8, formally deferring high-risk obligations to December 2026 and December 2027 in stages. But Article 50, the AI Act’s transparency regime, was not touched by that deferral and becomes enforceable on schedule on August 2, 2026: mandatory disclosure when a user is interacting with an AI system, labelling of AI-generated synthetic content, and disclosure of deepfakes. Unlike the high-risk provisions, Article 50 isn’t limited to systems classified as high-risk. It applies to any business using generative AI to produce content that reaches an end user, a considerably wider net than the compliance conversation of the last four months has focused on.

To put the two tracks side by side: the high-risk system requirements under Annex III (Articles 9 to 17 and 26) are now deferred by the Digital Omnibus to December 2026 and December 2027 in stages, while the transparency obligations under Article 50, which cover any generative AI producing content for end users and are not limited to high-risk, were left untouched by the Omnibus and bind as originally scheduled on August 2, 2026.

The practical risk is that legal and compliance teams who spent the spring tracking the high-risk deferral fight now read the Omnibus headline, “deadlines pushed back,” and stand down, when the obligation that actually lands in nine days was never part of that fight. Chatbot disclosure, AI-content marking, and deepfake labelling apply whether or not a given tool is high-risk, which in practice covers most of what a legal team’s own AI vendors are producing for it today: drafted correspondence, summarised research, generated first-pass documents.

The Flank read. This is a small, specific illustration of a pattern this briefing keeps documenting: regulatory complexity doesn’t resolve into a single deadline a legal team can put on a calendar and forget. It resolves into a set of overlapping obligations that a governance layer has to track continuously, not a compliance programme that gets built once. Flank’s supervision model exists for exactly this reason: every agent workflow runs through human review before output leaves the system, which means disclosure, labelling, and audit-trail obligations are met as a structural property of how the work gets done, not as a separate compliance exercise bolted on after the fact once the deadline is finally clear.

🧠 What this week tells us

A frontier-model dispute over stolen intelligence, three models moving in three different directions in a single week, a compliance deadline hiding behind a headline about deferral, a law firm’s usage numbers with no outcome numbers attached, and two judges quietly rewriting how much mercy a hallucination earns: none of these stories were coordinated, and all of them describe the same market working through the same unresolved question.

Intelligence is contested, not just cheap. The Moonshot dispute puts a legal question under a trend this briefing has covered as pure economics: if outputs can allegedly be stolen at industrial scale, provenance becomes a due-diligence question for every enterprise AI buyer, not just a pricing footnote.

The model underneath your tool keeps moving. Shipped, teased, and paused, all in one week. Almost none of it is visible from inside a legal AI product’s interface unless the vendor chooses to surface it, which means the model that produced last month’s output may not be the one producing this month’s.

Compliance arrives obligation by obligation. The EU AI Act’s high-risk deadline slipped to 2027. Its transparency rules didn’t, and they bind in nine days, on a wider set of systems than the deferral conversation suggested. Oversight keeps arriving in pieces, a transparency article here, a judicial distinction there, rather than as one governed system.

Usage data still isn’t outcome data. Crowell & Moring’s numbers are a real, rare data point on adoption depth. They still don’t answer the question last week’s survey found 83% of in-house teams can’t answer: whether the work costs less or gets done better. The gap between using AI and being able to prove what it did hasn’t moved.

✳️ The Flank view

Every story in this week’s briefing is a variation on the same structural gap: inexpensive work is being done by expensive resources, and neither cheaper intelligence, nor a firm’s usage numbers, nor a court’s evolving sense of how much mercy a mistake deserves, closes it on its own. A frontier lab allegedly stealing another’s outputs to build a faster model doesn’t change who reviews the work before it reaches a client. A compliance deadline hiding behind a deferral headline doesn’t change whether a legal team can actually show a regulator what its AI systems did. And a law firm publishing real, granular adoption numbers, genuinely rare, still isn’t the same claim as being able to show the work costs less or gets done to a verified standard.

Flank’s answer to all of it is structural, not a bet on any one model, vendor, or court’s mood this quarter: tools are a commodity, outcomes are not. We compete for the budget line where routine legal work already sits, the outside counsel, ALSP, and internal headcount spend that dwarfs any software line, not just a new tool to add to it. Agents execute the work under your playbooks, a human reviews every output before it leaves the system, and you don’t pay until the first workflow is live in production. Whichever model is running underneath any given vendor this week, whichever deadline just quietly bound, whichever judge just drew a new line on mercy, the one part of this week’s news that doesn’t change is that outsourcing the work to supervised agents is what turns “we use AI” into an answer a court, a regulator, or a CFO can actually verify.

✳️

Share

I’m Khaled. I work with in-house legal teams at Flank on getting agents into live workflows. Everyone agrees you should start with one workflow. Almost no one tells you which.

Last time I argued against trying to map your whole AI future at once. This picks that up and answers the question it left hanging: where to actually start.

🔍 First, you have to see the work

The honest answer is that before you can choose, you have to be able to see what you are choosing between, and most teams can’t. Not because they are disorganised, but because legal work doesn’t turn up as a set of neat, labelled workflows. In a smaller team it turns up as an inbox: requests from sales, procurement, HR, wherever, all blurring into one category in your head, legal stuff, one big pile. In a bigger org it is more structured than that, of course. You might have a commercial contracts team, an employment team, a regulatory or compliance one, each with its own remit. But that structure is organised around subject matter, not around the kind of work being done, and those are not the same thing. Even inside a tidy specialist team, the actual tasks are still all mixed in together.

But that structure is organised around subject matter, not around the kind of work being done, and those are not the same thing.

When you try to break that pile down, the instinct is to sort it by document type. NDAs here, MSAs there, DPAs in the corner. I don’t think that is the most useful cut, though, because what matters more is what you are actually doing to the thing. Take an MSA. One day you are reviewing a counterparty’s version against your positions. Another day you are drafting one from scratch off your own template. Another you are buried in redlines, answering the same objections you answered on the last three deals. Same document, three completely different jobs. Calling all of it “MSA work” hides that.

Share

Once you start sorting by the verb rather than the document, the real streams show up. There is reviewing paper other people send you. Drafting from your own templates. Handling the objections that come back on your paper. Answering the same questions the business keeps asking. Triaging whatever lands and sending it to the right person. And then there is the one people almost never count, which is all the work around the work: the chasing, the extracting, the summarising, the filing that sits behind every signed contract. Nobody writes that down as a workflow, so it never really gets dealt with, and in my experience it quietly eats more time than any single document type does.

The point of all this is that you can’t sensibly pick a workflow until you can see them laid out like that. So the first step isn’t choosing at all. It’s untangling.

Same document, three completely different jobs. Calling all of it “MSA work” hides that.

Then, narrow it down

Once they are laid out, you do have to pick one, and there are really two things I would look at.

The first is fairly obvious. Which of these is high volume and low complexity? You are not after the hardest or the most interesting work. You are after the stuff that comes in constantly and mostly follows a pattern, because that is both what an agent handles well and where your team’s hours are actually going.

Share

The second is less obvious, and I think it is the more useful of the two. When your team is completely underwater, where does that work actually go? Maybe it goes out to a law firm at eye-watering rates. Maybe offshore, or to an ALSP. Maybe you pull in contractors for a few months. Or maybe it just piles up while the business waits and grumbles. Wherever it ends up, follow it. That is where the cost is hiding. A workflow that quietly bleeds money to outside counsel every busy quarter is a far better place to start than the one that is merely irritating you day to day.

So now you have a shortlist: high volume, rules-based, and expensive when it overflows. There is one more trap to watch for, which is that the workflow causing the most pain often isn’t the real problem. It is usually the symptom of one, and if you automate the symptom it tends to come back a few weeks later in a slightly different form.

You are after the stuff that comes in constantly and mostly follows a pattern.

🧠 Then find the real problem, not the loud one

The thing I have found genuinely useful here is the old “five whys”. You take the problem you think you have and ask why, maybe five times, until you reach something that is actually a cause rather than a complaint.

Here is roughly how it tends to go.

“We want to automate supplier agreement review.” Why? “Supplier agreements take too long to turn around.” Why? “They sit in a queue waiting for a lawyer.” Why? “Every single one gets a full manual review.” Why? “We’ve never separated the standard ones from the ones that carry real risk.”

Share

The problem was never that supplier agreement review is slow. It is that a hundred near-identical, low-risk supplier agreements get treated exactly like the five that genuinely need a lawyer to stop and think. That changes what you would hand over. You don’t give an agent “supplier agreement review” as one big undifferentiated block. You give it the first pass on the standard ones, so your lawyers only ever see the handful that actually need them. Same area of work, much sharper thing to automate. And you would have missed it if you had stopped at the first answer, because the first answer is almost always the symptom.

So that is roughly how I would think about it:

1. Break the pile into streams based on what you are doing, not what the document is called.

2. Look for the ones that are high volume, rules-based, and expensive when they spill over.

3. Then keep asking why about the most promising one until you hit a real cause instead of a symptom.

More often than not, the workflow worth starting with is sitting right there at the bottom.

You still don’t need the whole map. You just want to be reasonably sure the first step is the right one, and this is how you check.

✳️

Every signed contract is a record of commitments. Most of them go into a folder that can’t answer a single question about what they hold.

This one is about metadata: the cheapest work in a legal function, and the work that almost never gets done. Here is why, and what finally changed it.

I’m Taariq. I lead Forward Deployment at Flank.

Your contract system holds a scan and nothing else

A few weeks ago I sat in a discovery session with the supply chain team of a large enterprise, walking through how a contract moves from request to signature. Somewhere in the middle of the workflow mapping, one of them said something almost as an aside. Metadata in their contract system is “treated as an afterthought”. It was the least controversial statement of the two-hour call with no internal divergences (which are often common with larger teams).

A colleague’s call the same fortnight made the extreme version concrete. A legal team handling five and a half to six thousand agreements a year, every signed contract scanned and filed into an internal system. No extracted terms, no renewal dates, no metadata at all. Six thousand times a year, a document full of commitments goes into a drawer that cannot answer a single question about what it contains.

A more important aspect of this ‘negligence’ also comes up often. Because past legal approvals are never memorialised anywhere, their sourcing managers cannot cite them. A clause legal approved two years ago comes back around, and rather than pointing to the precedent, the manager escalates it again. Legal reviews it again. The same position gets re-investigated as if it were new, by people who all vaguely remember settling it the first time.

Metadata is where the economics of a legal function show up in their purest form, and the teams above aren’t failing to follow their process. They’re following the economics exactly.

📄 Why the capture never happens

Recording what a contract says is the textbook case of inexpensive work. The parties, the term, the renewal mechanics, the positions conceded. No judgment required. A checklist could do it.

And yet it doesn’t get done, and it isn’t about laziness. At the moment of capture, that work is pure cost. The deal is closed, the business has moved on, and the only people who actually know what the document says are lawyers and deal managers whose next negotiation is already waiting. Asking someone who costs hundreds of pounds an hour to key renewal dates into a repository is a losing trade, and everyone in the workflow can feel it. So the field stays blank, or gets filled from memory at some point

The shape of the mismatch I usually write about is expensive people doing inexpensive work slowly. Metadata shows me the other shape, which I find more revealing: when the only qualified operator for a task costs more than the task is worth, the task stops being done at all. Nobody decides this: it’s simply what the economics produce, year after year, in team after team.

You buy the work back at a markup

Skipped capture only looks like saved cost. The work comes back, and it comes back more expensive every time.

A renewal date nobody recorded gets discovered by the counterparty, on the counterparty’s terms. A due diligence exercise asks a batch of questions the archive was theoretically built to answer, change of control, expiry, assignment, and answering them means paying people to re-read documents the team already paid people to write. The supply chain team I sat with described sourcing managers reconstructing deal context from old email threads, because the system that was supposed to hold it holds a scan and nothing else. The same team has no rulebook for template selection either, which is the same disease in a different organ: managers pick paper by experience, sometimes pick wrong, and the work gets done twice.

Share

The precedent problem may be the most expensive version. When approvals are never memorialised, every settled position is unsettled the next time it appears. The team escalates things legal has already blessed, which clogs the queue, which in turn slows the deals that actually need judgment. The work you declined to do at the cheap moment gets bought back later at the expensive one.

I think this is also why “we should tag our contracts” initiatives fail so reliably. Ask a builder why the snagging list never gets done and they won’t tell you the team forgot. They’ll tell you the crew that could do it is already on the next site. A tagging initiative asks the same expensive people to do the same underpriced task, just with more discipline this time. The economics that killed the habit are still standing there, untouched.

⚡ What changed

The reason I no longer treat any of this as an unfortunate fact of life is that the capture no longer needs the expensive operator.

Subscribe now

When an agent drafts or reviews a contract, it already knows the things a repository wants recorded. The parties, the term, the positions held and the positions conceded aren’t extra work to extract. They are the substance of the work itself. Capture stops being a separate task someone has to remember, and becomes a by-product of the work getting done. The losing trade disappears, because nobody’s hours are being spent on it. In the discovery session I mentioned, the team’s own first-choice use case wasn’t drafting or review. It was capture during the deal, not after it. They had diagnosed themselves before we said a word.

The back-book is harder: the teams I’ve seen think about this sensibly treat it as two different problems: capture-forward, which is now close to free, and backfill, which is a project you scope deliberately, starting with the contracts that are still alive.

None of this makes the design questions disappear. The first schema a team writes will be wrong, because a team that has never had metadata doesn’t yet know which fields it will want, and I’d rather see someone start wrong and narrow than stall for months designing the perfect taxonomy for a memory they don’t have. And renewal dates are facts, but “why we accepted this indemnity cap” sits somewhere between a fact and a piece of reasoning; a thin note can masquerade as the full rationale, so the line between recording and flattening deserves real thought. Neither question is a reason to wait. Both are questions you can only answer once something is being captured at all.

The eighteen-month test

Pick a contract your team signed eighteen months ago. Ask what it would cost, today, in someone’s hours, to answer two questions: when does it renew, and what did we concede to close it.

If the answer involves anyone re-reading the document, your department is paying expensive people to remember things it could have written down for nothing. And that price isn’t fixed. It gets steeper with every year of contracts filed into a drawer that can’t answer questions, and it comes due at the worst possible moments: the renewal the counterparty spotted first, the diligence exercise on a deadline, the position your own team settled and then paid to settle again.

The team I sat with a few weeks ago had already worked all of this out. Nobody had to persuade them that the memory mattered. What had changed was that keeping it no longer required their most expensive people’s time, and once that was true, the argument for the drawer full of silent paper collapsed on its own. The contract remembers nothing.

✳️

⚡ 01 Harvey acquires Benchmark, its third deal of 2026, to deepen its push into asset management

Harvey acquired Benchmark on July 16, a “decision infrastructure” platform used by firms representing more than $2 trillion in assets under management. Benchmark’s co-founders join Harvey’s product and engineering team. Harvey already serves 50-plus asset managers, including Blue Owl Capital, Bridgewater Associates and KKR.

⚖️ 02 The Arizona Court of Appeals rules that citing AI-fabricated cases is sanctionable, even without a lawyer involved

A self-represented litigant in a probate dispute admitted using generative AI without checking the results: two of eight cited cases did not exist. The court held that both attorneys and self-represented litigants can face contempt, dismissal, fines or fee-shifting, regardless of intent.

🏗 03 China’s new AI agent rules take effect, forcing ByteDance and Alibaba to shut down companion agents

The Interim Measures for the Administration of AI Anthropomorphic Interaction Services, issued jointly by five central government bodies, require anti-addiction systems, usage warnings and instant-exit mechanisms for any AI built to sustain an ongoing relationship with a user. Rather than retrofit the feature, Doubao and Qwen simply switched it off.

⚡ 04 Meta’s Louisiana data centre investment reaches $50B, up from $10B less than two years ago

The Richland Parish campus expands to nearly 10 million square feet and 5 gigawatts of capacity, becoming one of the largest data centres ever built. Meta’s own commitment has risen from $10B (2024) to $27B (October) to more than $50B now.

🏗 05 Illinois signs the first US law requiring independent third-party safety audits of frontier AI

Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, with broad bipartisan support. It applies to the largest AI developers, requires public disclosure of safety practices and “catastrophic risk” frameworks, and creates whistleblower protections. Takes effect January 1, 2027.

Share the latest news with your colleague:

Share

⚡ 01 Harvey just made its third acquisition of the year. It wasn’t a law firm tool: it was a way into $2 trillion of private-markets AUM.

Harvey has spent three years building the deepest brand in legal AI. On July 16 it announced the acquisition of Benchmark, a “decision infrastructure” platform that pulls a private investment firm’s scattered deal memos, financial models, emails and legal documents into one searchable system, used today by firms representing more than $2 trillion in assets under management. Benchmark’s co-founders, Alec Dunn and Connor Janson, join Harvey’s product and engineering organisation. Terms weren’t disclosed, but the strategic logic is not subtle: Harvey already works with more than 50 asset managers, including Blue Owl Capital, Bridgewater Associates and KKR, and this is its third acquisition this year alone.

What Benchmark does for a private-markets analyst is structurally the same job a legal AI copilot does for a lawyer: it makes an expert faster at reading, screening and summarising the material already sitting on their desk. It does not remove the deal-memo drafting or portfolio-monitoring task from anyone’s queue. Harvey’s own product suite tells the same story at scale. Agent Builder, its self-serve tool for constructing custom multi-step agents, still hands the configuration work to the buyer’s own team, and its per-seat pricing (roughly $1,000 to $1,200 a month per user) sits squarely in software-budget territory. Three acquisitions in a year builds a wider surface of tools. It doesn’t, on its own, change who does the underlying work.

The Flank read. Harvey is executing the textbook response to inexpensive work being done by expensive resources: it is buying more tools to hand to the expensive resource. That’s a reasonable growth strategy for a $200M-ARR software company, and Benchmark’s underlying product is a genuine improvement for the analysts who use it. But it is still the assistant model, applied to a new desk. The lawyer, or the analyst, opens the tool, gets a faster answer, and does the work themselves. The bottleneck doesn’t move. It gets a better interface. Flank’s bet is different in kind, not degree: agents that execute the workflow end-to-end, under a supervision model built as core product rather than assembled from acquisitions. The acquisition that matters isn’t a new tool for the desk. It’s the work leaving the desk entirely.

⚖️ 02 A new survey of 528 in-house legal leaders finds 83% still can’t prove their AI spend is working. The same week, a court ruled that citing AI-fabricated cases is sanctionable even if you’re not a lawyer at all.

Two stories this week describe the same adoption curve from opposite ends. Axiom’s 2026 In-House Legal AI Report, based on a survey of 528 legal leaders across six countries conducted by InsightDynamo, found that 83% of in-house teams cannot say whether last year’s AI spending actually paid off, and that just 7% have moved past piloting into organisation-wide, measured use. Only 17% of AI-using teams track ROI against a formal framework at all: 51% track it inconsistently, without one. Where AI is delivering measurable value, it concentrates in a narrow band of high-volume, document-heavy tasks: legal research, contract review, document summarisation. That is exactly the templated, repeatable work this briefing keeps returning to.

The same week, the Arizona Court of Appeals ruled on a probate dispute over the estate of a woman who died in 2023, in which a self-represented litigant admitted using generative AI without verifying the results: six of eight cited cases were deficient, two didn’t exist at all. Writing for a unanimous panel, Judge Brian Furuya rejected any suggestion that lack of intent should excuse it: “To accept a defense of lack of ill intent would legitimize reckless indifference to the truth and accuracy of court filings.” The ruling extends a body of sanctions case law this briefing has tracked for months (New York, Oregon, Mississippi, the Ninth Circuit) to a new category of user entirely: the self-represented litigant, the exact population MIT and USC found producing AI-generated writing in 18% of federal filings.

The Flank read. Put the two stories together and they describe the same missing layer from opposite directions. Axiom’s numbers are the demand-side version of a pattern this briefing has cited before in different clothing: teams adopt AI fast, and the spend keeps rising, but the organisation still can’t show the work costs less or gets done better, because nobody built the measurement and supervision infrastructure alongside the tool. The Arizona ruling is what happens at the other end of that same gap, when there is no verification step between an AI’s output and a court filing. Both are the same argument: access to AI is not the constraint any more, and hasn’t been for a while. The constraint is a governed system that turns access into a measured, supervised outcome. That is exactly why Flank prices itself on the outcome delivered, not the licence sold, and why every agent workflow runs through human review before anything leaves the system. The 7% who’ve scaled AI and the courts sanctioning the 93% who haven’t verified it are describing the same missing infrastructure.

🏗 03+05 Illinois and China moved in the same week, from opposite ends of the world, toward the same instinct: an AI agent operating without independent oversight is now a compliance problem, not just a product one.

Illinois and China regulate almost nothing else the same way. This week they converged on one idea: that an AI system acting with some autonomy, whether a frontier model reasoning about catastrophic risk or a companion agent sustaining a relationship with a user, needs a check that isn’t the vendor marking its own homework. On July 6, Governor JB Pritzker signed Illinois SB 315, the Artificial Intelligence Safety Measures Act, making Illinois the first US state to require regular independent third-party safety audits of the largest AI systems, with penalties up to $1M for a first violation. On July 15, China’s Interim Measures for the Administration of AI Anthropomorphic Interaction Services took effect, forcing ByteDance’s Doubao and Alibaba’s Qwen to shut down personalised companion agents used by hundreds of millions rather than retrofit the anti-addiction and instant-exit mechanisms the rules now require.

Two legal systems, one direction of travel

Illinois SB 315 targets the largest frontier AI developers with more than $500M in revenue, requiring independent third-party safety audits and whistleblower protections, effective January 1, 2027. China’s measures, issued by five agencies, target AI agents built to sustain an ongoing relationship with a user, requiring mandatory anti-addiction systems, usage warnings and instant-exit mechanisms, effective July 15, 2026.

Neither rule was written with enterprise legal AI in mind. One targets catastrophic-risk disclosure at the frontier-model layer, the other targets consumer companion products. But both share a premise that is directly relevant to how legal departments will be asked to procure AI going forward: that autonomy without independent verification is the thing regulators are now moving to close off, wherever they find it. Gartner has already said the general counsel should be asserting AI governance leadership inside their own organisations. ModelOp’s tracking shows AI governance-platform adoption jumping from 14% to 50% of enterprises in a year. The direction of travel, in two very different legal systems in the same week, points the same way.

The Flank read. This is the environment Flank’s model was built for, not a threat to it. Supervision that’s assembled after the fact, through a partnership with a governance vendor bolted onto someone else’s product, is a different claim from supervision that’s native to the product from day one. Harvey’s governance today runs through partnerships: Intapp for ethical walls, Aderant for billing. Flank’s supervision is the core product: tenured legal professionals reviewing every agent workflow before it reaches the business, under whichever of three models the client chooses (their own lawyers, a partner firm, or Flank’s own supervision team). As “was this independently audited” becomes a procurement question rather than a nice-to-have, the vendors who can answer it structurally, not contractually, are the ones left standing.

⚡ 04 TSMC just posted a record quarter and put another $100B into Arizona. Whatever happens at the model layer this week, the money betting on AI compute for the next decade didn’t blink.

Every legal AI product this briefing covers is a wrapper around someone else’s model, and every one of those models runs on someone else’s chips. TSMC reported record Q2 2026 results on July 16: net profit up 77.4% year-on-year, revenue of $40.2B, gross margin at 67.7%, and its high-performance-computing segment, the AI-accelerator business, now 66% of total revenue, up 20% in a single quarter. CEO C.C. Wei used the moment to announce a further $100B investment in Arizona, bringing TSMC’s total committed spend in the state to $265B. The same week, the US Commerce Department reclassified the UAE into its highest trust tier for chip exports, clearing licence-free sales of Nvidia and AMD’s most advanced accelerators and accelerating the $30B, 5-gigawatt Stargate UAE compute campus in Abu Dhabi.

None of this capital is being deployed cautiously. TSMC’s own numbers show why: 66% of its revenue now comes from the chips that power AI, and its gross margin is climbing because it holds roughly three-quarters of the global advanced foundry market. Every legal AI vendor’s cost base, and every enterprise buyer’s exposure to that cost base, sits downstream of a supply chain this concentrated.

The structural question for enterprise legal teams. The volatility at the model layer (Gemini’s rebuild, Grok’s aggressive pricing, GPT-5.6’s release) sits on top of an infrastructure layer that is consolidating around a handful of chip and compute suppliers, not diversifying. If the model your legal AI vendor licenses today is replaced, re-priced, or capacity-constrained by decisions made in Taipei, Washington or Abu Dhabi, does your contract, or your workflow, survive that switch? A vendor-neutral supervision layer is worth more the more concentrated the layer underneath it becomes.

⚡ The price war: xAI just priced a frontier-class model at a fraction of Claude’s cost. If that logic holds, it should be reshaping what your legal AI vendor charges you. Nothing says it will.

Last week’s briefing covered Harvey’s own disclosure that its token processing had grown 14x in six months, with a single large contract review capable of generating a $20,000 compute bill. This week, xAI shipped Grok 4.5, pricing it at $2 per million input tokens and $6 per million output tokens, by some estimates three to four times cheaper than Claude Opus 4.8, while claiming roughly double the token efficiency of comparable models on agentic coding tasks. It scores 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-Bench Pro, putting it in genuine contention with the frontier rather than trailing it. Google, meanwhile, is expected to ship its own rebuilt Gemini 3.5 Pro today, closing the loop on the architectural scrap-and-rebuild that wiped $225B off Alphabet’s market cap two weeks ago. As of publication, though, Google has not confirmed the specifications circulating in developer previews, including a reported 2-million-token context window and a “Deep Think” reasoning mode gated to a $250-a-month tier.

The structural question for enterprise legal teams. If a frontier lab can price a competitive model at a fraction of what another lab charges for comparable capability, the marginal cost of the intelligence underneath your legal AI tools is falling faster than most vendor contracts reflect. Does your legal AI vendor’s pricing move when the frontier model underneath it gets cheaper, or does the saving stay in the vendor’s margin while your renewal quote holds flat? A flat SaaS fee has no mechanism to pass that through. A model that prices on outcomes, not licences, has every incentive to.

🧠 What last week tells us

A software vendor buying its way into a new vertical, a survey confirming most legal AI spend still can’t prove its worth, two governments writing independent audits into law, and hundreds of billions of dollars committed to the chips and data centres underneath all of it: none of these stories mention each other, and all of them are describing the same market. Intelligence is getting cheaper and more available by the week. Oversight of that intelligence is becoming a legal obligation, not a competitive nicety. And almost nobody buying AI for legal work today can show, with real numbers, that the work now costs less or gets done better than it did before.

Tools consolidate, but the work stays on the same desk. Harvey’s third acquisition this year widens its product surface without changing who does the underlying work: the assistant model, applied to a new desk.

Intelligence gets cheaper, and there’s no mechanism forcing that saving through to you. Grok 4.5’s aggressive pricing and Gemini 3.5 Pro’s imminent rebuild both point toward falling marginal costs at the frontier, with no mechanism forcing that saving through to the enterprise buyer on a flat SaaS contract.

Oversight becomes law. Illinois and China, independently, in the same week, moved to require independent verification of AI systems operating with autonomy. It previews what legal AI procurement will soon demand.

The ROI still isn’t there. 83% of in-house teams can’t prove their AI spend worked. Courts keep sanctioning the output of AI nobody checked. Access was never the hard part.

✳️ The Flank view

Every thread in this week’s briefing is a variation on the same mismatch: inexpensive work is being done by expensive resources, and cheaper, more capable, more heavily audited AI does not, on its own, close that gap. Falling frontier-model prices and hundreds of billions in new compute make intelligence more abundant. New audit laws in Illinois and Beijing make unverified autonomy more legally exposed. Neither one touches the actual constraint Axiom’s survey just put a number on: 83% of legal teams still can’t show the work costs less, because access to a tool was never the same claim as the work done and proved it.

Harvey’s answer is to keep buying tools. Ours is structural: tools are a commodity, outcomes are not. Flank competes for the budget line where routine legal work already sits, not only a new software line, but the ten-to-fifty-times-larger services spend on outside counsel, ALSPs and internal headcount doing templated, repeatable work. Agents execute it under your playbooks, a human reviews every output before it leaves the system, and you don’t pay until the first workflow is live in production. As audits get written into law and frontier pricing keeps shifting under every vendor’s feet, that’s the one part of this week’s news that doesn’t change: insource the work to supervised agents, and the ROI question Axiom’s survey couldn’t answer stops being the open question it is for everyone else.

✳️

Software engineering has just spent three years running an experiment legal is about to repeat: collapse the cost of producing work and see what happens to the cost of checking it.

The results are in, and I find them worth reading closely, because my field got its first answer wrong before it found a better one.

Hi there. I’m Martin, CTO at Flank.

Subscribe now

The bill moved

Writing code felt like the constraint for most of my career. It was the part that filled the estimates, the part we hired for, the part that made projects late. Then, in the space of about two years, language models made producing code close to free. Engineers with AI assistants now complete nearly twice as many changes as those without, according to telemetry Faros AI collected across more than ten thousand developers. On paper, that is the productivity revolution everyone was promised.

The same dataset shows what the revolution actually did. Each change grew 154% larger. The time for a colleague to review one rose 91%, and it has kept climbing since. Google’s DORA research programme, the closest thing engineering has to a longitudinal study of itself, found in 2025 that teams adopting AI ship more and break more at the same time, with the hours saved on writing quietly reallocated to auditing what was written.

None of this surprised the older engineers. There is a strand of thinking, running back through fifty years of software literature, which holds that writing code was never the real constraint at all. Understanding, verifying and integrating it was. The generation step was simply expensive enough to hide that. Make generation free and the truth surfaces: the entire cost of software is the checking. The bill did not shrink. It moved.

🏗️ What we did about it

The first response was the honest, doomed one: review everything, harder. Senior engineers became full-time readers of machine output, and the bottleneck reformed one desk downstream, exactly where the throughput gains were supposed to land. What broke that pattern was a pair of moves that I think of as one idea.

The first move was letting AI into review itself. Every major model provider now ships a specialised tool or skill for reviewing code; GitHub’s review agent alone performed sixty million reviews within a year of launch. The point of these systems is only partly catching bugs. Their more interesting job is directing attention: annotating which change touches authentication, which migration brushes against the payments table, which innocuous-looking diff has a large blast radius. A reviewer whose eyes are steered to the dangerous ten lines is a different economic object from one reading four thousand lines at a flat rate.

The second move was making triage explicit. On a well-run team today, the most routine changes, the kind that pass every automated test and follow a well-worn pattern, ship with no human involved at all. Anything touching payments, passwords or customer data: a human review is mandatory, sometimes two. Between those poles every team calibrates its own line, and the line moves as trust accumulates. The same Faros telemetry shows nearly a third more changes now shipping with no human review at all, while review time on the remainder goes up. That reads to me as deliberate spending rather than carelessness. Attention is being taken away from changes that never needed it and concentrated on changes that always did.

That is the idea underneath both moves: review is a budget. Reviewer attention is finite, expensive and, unlike generation, does not scale with volume. So each change gets priced, in attention, proportional to what happens if it is wrong: what breaks, for how many people, how reversibly. Cheap changes cost near zero. Dangerous ones cost more than they did before, on purpose. No standards body wrote this down. Thousands of teams converged on it independently, because every alternative either shipped incidents or recreated the queue.

⚖️ Legal is early in the same experiment

Legal has reached the point in the curve where producing the work is fast and everything else is not. Models draft contracts and produce first-pass reviews in minutes. And then the verification begins. Thomson Reuters found that 41% of professionals will not use AI output without human review unless it is 100% accurate, a standard no human associate has ever been held to, and 91% admit to holding machines to a higher bar than people. The caution is earned. Stanford researchers benchmarked the leading AI legal research tools and found them wrong on between one in six and one in three queries, against marketing that said hallucination-free.

So lawyers check everything, and they are right to. But checking everything at the document level is the flat-rate response, the one my industry tried first. My colleague Taariq described the end state months ago: a lawyer who runs every NDA through a tool, then redoes the work by hand, every day, for six months. Adoption is now above half of in-house teams while the share reporting an actual reduction in cost sits in single digits, and I think the flat rate is most of the explanation.

Jake wrote this month that legal is an operational bottleneck, and that the right amount of human involvement is whatever lets a named person own the result, and no more. Supervision falls where the risk is. I agree with the whole argument, and I want to add the part I can speak to as an engineer: how a team actually gets there. Because “supervision falls where the risk is” is exactly where software landed, and the way it got there was by learning to price review.

🔍 Pricing legal review

Blast radius exists in legal work, and it is computable more often than the profession assumes: contract value, indemnity and liability exposure, the distance of a requested deviation from the playbook position, whether the paper is your template or the counterparty’s, and whether it resembles anything the system has processed before. Reversibility above all. A standard NDA that expires in two years is the legal equivalent of the change that ships itself. An uncapped indemnity in a seven-figure MSA is the one touching payments, and no sane budget prices them the same.

The hard part is the pricing engine, and I want to be precise about why, because it decides which products deserve trust. A language model asked to score its own confidence will cheerfully overstate it; models are poorly calibrated on their own uncertainty by default. Turning “the agent handled this” into “this is safe to pass” takes engineering around the model: confidence scoring tuned to the team’s actual risk thresholds, routing logic a lawyer configured, and recalibration as every correction feeds back. When a reviewer overrides the same call three weeks running, the price of that category was set wrong, and the system should be the thing that notices.

Get the pricing right and the lawyer’s morning changes shape. The queue stops being forty documents needing forty readings. Most of the volume has cleared against playbook positions overnight, logged and inspectable. What waits is a short list of decisions, each arriving with its reasoning attached: an indemnity outside the usual position on a contract large enough to matter, a governing-law clause the system could not map with confidence. The lawyer spends the hour on judgment, which is the thing the department actually pays them for. Gartner now expects that by 2029 around half of contract reviews will be delegated to self-service systems that escalate one in ten to a human. I read that forecast less as a prediction about AI and more as a prediction that legal will learn to price.

The question to ask

Here is what I am confident of, and what I am not. The trajectory seems clear to me; software is simply further along the same curve, and the physics of cheap generation and expensive attention are the same in both fields. What nobody can tell you is where your team’s line belongs, because early budgets will be mispriced in both directions. Some team will auto-clear a category it should not have, and the correction will be public and painful. Some team will keep hand-checking standard NDAs for two years longer than the evidence warranted, and that failure will never make the news, because it looks like diligence. It is worth saying that engineering has no survey proving how many teams run formal tiered review either. It is converged practice, visible in tooling defaults and telemetry rather than in standards. Legal should expect its own version to arrive the same way: quietly, team by team, ahead of the official guidance.

But the question to put to any system you are evaluating has, I think, already settled. Ask how good the drafts are, certainly. Then ask the question that decides whether the work ever leaves your desk: when this thing hands my team an output, does it tell them what that output should cost to review? A system that prices everything the same, whatever its accuracy, is asking your most expensive people to spend their attention at a flat rate on work that mostly never needed them. Legal departments have spent a decade learning what that mismatch costs when the unit is a contract. The next few years decide whether they repeat it when the unit is a review.

A legal team’s attention is the scarcest thing it owns. The discipline that matters now is refusing to spend it at flat rate.

✳️

Hi there, I’m Paul. VP Product at Flank.

Most legal AI budgets were sold on one promise: lower costs. Two years in, the spend has barely touched the cost line, and I don’t think the technology is the reason.

Share

The number that should bother you

I run almost everything I write for the outside world through a shared brain before it goes out: a draft LinkedIn post, a partnership deck, the language we use for a new feature. The brain is a set of documents holding how we think, what we believe, and how we say it, sitting behind the AI tools the whole team uses. Last week one of the checks I have running told me two parts of it had drifted apart. The product side and the marketing side had started describing the same thing in slightly different words, the kind of gap nobody notices until a customer reads both and asks which is true. A scheduled task caught it, pointed me to the spot, and I fixed it in a few minutes.

We built the first version of that brain in an afternoon, and it has grown for months since. It has made me, and everyone around me, measurably better at the work. I want to be precise about what it did not do, because I think that distinction is the one that matters most in enterprise legal right now. It made us faster. It did not take a single piece of work off our plates.

That sounds like a small thing. I think it is close to the whole point.

The number that should bother you

The ACC’s 2026 data puts corporate legal AI adoption above half, more than double the year before. In the same survey, only 7% of teams report an actual reduction in total matter cost. Two years of budget, board mandates and real enthusiasm, and the thing the money was meant to move has barely moved.

I don’t read that as a failure of the technology. I read it as a category error. Almost everything deployed so far makes individual lawyers faster: a better drafting assistant, a research tool that reads faster than any associate, a chat window that knows more law than the person typing into it. All useful. None of it changes who picks the work up off the pile. The lawyer is still doing the contract, now with a sharper instrument, and the sharpness of the instrument was never the constraint.

⚡ Two different jobs

There are two things you can do with this technology, and I find most teams have quietly done one while believing they were doing the other.

The first is to make your people more capable. This is what a general tool like Claude does well, and it is where the shared brain lives. The brain works not because it stores documents but because it grounds a group of individuals, each using AI on their own, in the same context. Everyone’s output now pulls from one source of truth. A signal that surfaces in a customer call can, if it proves out, climb into the strategy the whole team reasons from. The unit that gets smarter is the team, not the person. That shift, from the individual to the system, is the part I think people underrate.

The catch is that shared context rots. The brain that made everyone sharper in March is quietly misleading them by June if nothing tends it. So you point agents at it: small scheduled jobs that remove what is out of date, strip duplication, and flag where two parts have begun to contradict each other. That is the drift my check caught. Without that upkeep a shared brain becomes a shared liability, people stop trusting it, and a source of truth nobody trusts is just a folder. The maintenance is not overhead. It is the thing that keeps the brain alive.

But notice the ceiling. Done well, this makes everyone better at the work. It does not remove the work. I am still writing the post; my team is still drafting the contract. We do it faster and more consistently, which is worth a great deal, but the pile is the same size at the end of the day.

The second job is to take the work off their desks entirely, and it is a different category. Some work you do not want your team touching at all: not because they can’t, but because it is high volume, rule governed, and a poor use of expensive judgement. The routine NDA, the standard vendor agreement, the counterparty paper that follows a pattern you have seen a thousand times. Until recently the only way to move that off your team was to send it out, to outside counsel or an offshore centre or a managed service. Someone else’s people, still doing it by hand, still billing by volume, and the context and control left with the work.

An agent that owns the whole workflow is the other option, and its shape matters. The request arrives where it always has, usually an email, from a business user who learns no new tool and changes nothing about how they ask. It gets triaged and routed, to a person or to an agent. Where an agent picks it up, it drafts the document or redlines the counterparty paper against your playbook, and where it hits something high risk or low confidence, it stops and asks. The work leaves the lawyer’s queue. What arrives instead is a decision, not a task.

🔍 Supervision is the product, not the caveat

This is where people expect me to reassure them about accuracy, and I think accuracy is the wrong thing to fixate on.

The question a general counsel actually asks is not “is it right.” It is “can I stand behind it.” An output can be correct and still useless to a lawyer who has no way to satisfy themselves that it is, because the accountability does not transfer. Their name is on the work. Hand them a finished document and tell them it is done, and they will do what any careful person does with something they are answerable for: re-read all of it. At which point you have spent what it would have cost to draft it.

So the thing that makes this work is not the model getting the answer right. It is the surface the lawyer reviews it through, one that flags only the decisions a human needs to make and lets everything the agent is confident about pass. You stop checking the work and start checking the exceptions. I think that is the most important design choice in the whole system, and the one that gets the least attention, because it is less impressive in a demo than a clean draft appearing from nowhere.

Picture the lawyer’s side of it. A counterparty NDA comes back, already redlined against the team’s playbook. Most of it is resolved and marked as resolved. Two points are flagged: an indemnity outside the team’s usual position, and a governing-law clause the agent was not confident it had mapped. The lawyer looks at two decisions rather than forty clauses, makes the calls, and the response goes back out. The thing that used to take an afternoon is now a couple of judgements and a signature. None of it depends on the agent being perfect. It depends on the agent being honest about the two places it was unsure.

The amount you check also falls over time, and it does not start low. Early on a lawyer checks a lot, the way you would with a capable new hire whose judgement you do not yet know. As the playbook tightens and corrections feed back, the checking drops. I have seen teams barely review a given workflow after a few months, and others hold a high rate for far longer because the work warrants it. That curve, human input falling as the agent earns trust, is the real shape of adoption: slower than the demo implies, more durable than the press release. I don’t think anyone can tell you in advance where your curve settles or how fast it gets there. You find out by running real requests through it and watching.

Where this leaves you

If your AI spend has made your team faster and your cost base has not moved, I don’t think you have done anything wrong. You have done the first job. It is worth doing, and cheaper and quicker to start than most people expect. Build the shared brain; it compounds. But it was never going to shrink the pile, because making a lawyer faster and removing the lawyer from the routine work are not the same act, and only one of them shows up in the matter-cost line.

The second job is a different decision, and it feels less like buying software and more like changing where the work goes. It sits in the same budget you already spend getting routine work done outside the team, usually far larger than the software line, and it keeps the context and control inside the building rather than shipping them out. What holds it back is not the technology. On the harder, more variable work I am routinely surprised by how much it can do. It is that letting whole categories of work leave your team’s queue is a management decision before a technical one, and most people have not yet decided how much of the day they are willing to hand over.

That, I think, is the real question under the 7%. Not whether the agents are good enough, but how much of your team’s work you are prepared to stop doing, and how quickly you can learn to trust what comes back when you do.

✳️

Making everyone faster is the first job. Taking the work off their desks is the second, and it lives or dies on what the lawyer actually sees. In the upcoming weeks, we’’ll write more about that.

Subscribe now

⚡ 01 OpenAI ships GPT-5.6 and launches ChatGPT Work, its long-task agent for professional work

GA arrives across ChatGPT, ChatGPT Work, Codex, and the API after a 13-day government-coordinated preview. Another reminder that every legal AI tool sits on a frontier-model layer that changes underneath it, on someone else’s schedule.

⚡ 02 Google delays Gemini 3.5 Pro to July 17 for a full architectural rebuild, wiping $225B off Alphabet

DeepMind scraps the Gemini 2.5 Pro base model entirely rather than iterate on it, missing the June launch window CEO Sundar Pichai promised at I/O. The market reaction is compounded by four senior researchers departing in a single week, including Gemini co-lead Noam Shazeer.

⚡ 03 Harvey discloses its token processing grew 14x in six months, to roughly 13 trillion tokens a month

Co-founder Gabe Pereyra says a single 100,000-document contract review can generate a $20,000 token bill before anyone reviews the output, a preview of the “token pricing reckoning” he says is coming for every AI vendor charging flat SaaS fees.

⚡ 04 Norm Ai raises $120M at a $1.2B valuation to scale its outcome-billed, human-supervised AI law firm

The Series C, led by Khosla Ventures, brings Norm’s total raised to more than $260M in under three years. Norm Law charges clients on outcomes rather than the hour and staffs its agents with human attorneys, including Sidley Austin’s former executive-committee chair as its chairman.

⚖️ 05 The Florida Bar’s free legal-AI member benefit gets fresh coverage as a first-in-nation programme

Announced June 16 in partnership with Clio, the Preferred Bar Program offers eligible members four months of unlimited Clio Work access plus a responsible-AI training and certification track. It isn’t live yet: members can register interest now, with access expected later this year.

⚡ 06 Clio Work introduces “Skills,” letting its AI learn and repeat how a firm does specific work

No prompt engineering required: a lawyer tells Clio Work “remember this for next time” and the preference becomes a reusable, firm-wide asset applied to future matters.

⚖️ 07 A New York appellate panel sanctions an attorney and his firm $10,500 for AI-fabricated citations

The Second Department orders Michael Sanders, Esq. to pay $8,000 and the Law Offices of Michael S. Lamonsoff, PLLC to pay $2,500 to the Lawyers’ Fund for Client Protection, after a brief quoted a nonexistent Court of Appeals case, Xiang Fu Ji v City of New York.

Receive two articles and our weekly briefing by subscribing below.

Subscribe now

⚡ 01+02 Google lost $225B and four senior researchers in a week, while OpenAI shipped GPT-5.6: every legal AI product rents its intelligence from somewhere else

Not every story this week was about legal AI vendors directly. On July 8, reports confirmed Google DeepMind had delayed Gemini 3.5 Pro to July 17, scrapping the existing Gemini 2.5 Pro architecture for a ground-up rebuild rather than shipping the June update CEO Sundar Pichai had promised at I/O in May. The rebuild targets gaps in mathematical reasoning, image quality, and agentic workflows relative to OpenAI’s GPT-5.6 and Anthropic’s Fable 5. The market’s reaction was blunt: roughly $225B wiped off Alphabet’s market capitalisation, compounded by four senior DeepMind researchers departing in a single week, including Gemini co-lead Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic.

The same week, OpenAI moved the other direction. On July 9 its GPT-5.6 family (Sol, Terra, and Luna) reached general availability across ChatGPT, the API, and Codex, alongside ChatGPT Work, a new agent built for longer professional tasks, following a 13-day government-coordinated preview that began June 26.

Neither story mentions a law firm or a legal department. Both belong in this briefing anyway: Harvey, CoCounsel, and most of the legal AI vendors covered every week are wrappers around exactly this layer, Anthropic’s, OpenAI’s, or Google’s frontier models, whichever one a given product licenses. When one of those labs scraps a model days before its promised release and loses senior researchers to a rival in the same week, the volatility doesn’t stay contained to the model layer. It becomes a vendor-risk question for whoever built a legal workflow on top of it.

The question worth asking. Your legal AI vendor’s roadmap is only as stable as the foundation model it licenses, and this week showed that foundation can move without warning. Do you know which frontier model each of your legal AI tools actually runs on, what happens to your workflow if that model is delayed, deprecated, or replaced, and whether your vendor’s routing and escalation logic would survive a switch to a different underlying model? A supervised routing layer that isn’t captive to one lab’s release calendar is a resilience feature, not just an efficiency one.

⚡ 03 Harvey’s own numbers show what happens when AI work scales without a cost control layer: a single contract review can cost more in tokens than the lawyers reviewing it

Harvey has spent three years selling AI as a productivity multiplier. This week, co-founder and president Gabe Pereyra put a number on what that multiplier actually costs to run: token processing on the platform grew 14x in six months, and a single 100,000-document contract review can now generate a $20,000 compute bill before a human reviews any of it. Even a routine document draft costs roughly $20 in tokens, a fee that disappears into a flat SaaS subscription today, but won’t stay hidden as agentic workloads scale.

The volume behind those numbers is real: Harvey’s ARR is estimated at roughly $300M as of May 2026, up from $100M last August, serving more than 142,000 lawyers across 1,500-plus customers and 60-plus countries, including half of the Am Law 100. Pereyra’s own framing is that the fixed-fee SaaS model is heading for a reckoning: as AI shifts from single-shot chat answers to multi-step agents that plan, use tools, and reason across long horizons, the cost of a task stops being a rounding error and starts being a line item enterprises will demand visibility into: which model did the work, how it was routed, how many tokens it burned, and what outcome it produced.

The Flank read. A $20,000 bill to review one contract is not a pricing footnote: it’s the cost of running expensive, unrouted compute against work that shouldn’t need it. Most of what floods a 100,000-document review is standard-form: NDAs, boilerplate clauses, repeat counterparties. That volume is inexpensive work; it only becomes an expensive resource, a $20,000 token bill, or the billable hours of the lawyers who’d otherwise triage it by hand, when there’s no routing layer to separate the commodity share from the matters that actually need full model horsepower and full attorney attention. Flank’s model bakes that routing in at the front: supervised agents handle the high-volume, templated work directly, with a human review gate before anything ships, so the cost curve tracks the complexity of the matter instead of the size of the document pile.

The question worth asking. If your AI vendor’s costs are climbing 14x behind a flat subscription, that cost is being absorbed somewhere: in the vendor’s margin today, and in your contract renewal tomorrow. Do you know which categories of your own AI-assisted legal work are token-expensive because they’re genuinely complex, versus token-expensive because nothing is triaging volume before it hits the model? The second category is the one a routing layer can price down; the first is where the spend belongs.

⚡ 04 Norm Ai just raised $120M to prove that an AI-native law firm can charge for results instead of hours, with its own agents doing the work under attorney supervision

Every legal AI vendor sells software to law firms and legal departments who still bill and get billed by the hour. Norm Ai is building something structurally different: an AI-native law firm, Norm Law, that uses Norm’s own agents to do legal work, puts human attorneys on top of them to supervise it, and charges clients for the outcome rather than the time it took to get there. The $120M Series C, led by Khosla Ventures and announced July 7, values the company at $1.2B, unicorn status less than three years after founding.

The investor list reads like a bet on the model, not just the software. Alongside Khosla, the round includes Bain, Craft Ventures, Coatue, Vanguard, New York Life, and TIAA, plus Tony James, the former president and COO of Blackstone, and Jeff Hammes, the former chairman of Kirkland & Ellis. Norm Law itself is chaired by Mike Schmidtberger, the former chair of Sidley Austin’s executive committee, with partners drawn from Ropes & Gray, Paul Weiss, Davis Polk, Skadden, Cleary Gottlieb, Latham & Watkins, and Kirkland: Big Law pedigree, redeployed around agents instead of associates.

The Flank read. Norm Law is the clearest market validation yet of the diagnosis this briefing keeps making: inexpensive work is done by expensive resources because the billing model, not the work itself, has kept it there. An outcome-billed, agent-run, human-supervised law firm only makes economic sense if a meaningful share of legal work is repeatable enough to price by result, which is exactly the commodity NDA, redline, and procurement volume that sits in every in-house queue today. Norm built that model as an outside firm you hire. Flank’s answer is the same structural bet applied inside your own legal department: outsource that work to supervised agents that already know your templates, terms, and escalation rules, with a human reviewing every output before it leaves the system, so the outcome-based economics Norm is proving in the market work for your team directly, without routing the matter to an outside firm at all.

⚖️ 05+06 Florida’s bar just made legal AI free for its members. Clio just made its AI remember how each firm likes its work done. Neither is the same as a governed system

Two Clio-linked stories landed six days apart and describe the same shift from different angles. On June 16, the Florida Bar announced it would become the first state bar in the country to offer legal AI as a free member benefit; on July 2, Clio Work shipped “Skills,” a feature that lets its AI learn and repeat how a specific lawyer or firm handles a given type of work. Read together, they show the legal profession’s official infrastructure normalising AI access at the same time vendors are making that access more personalised, without either one constituting the governance layer a firm still has to build for itself.

What Skills actually does

Clio Work’s new Skills feature removes the technical barrier to customising an AI workspace: instead of prompt engineering, a lawyer tells the system “remember this for next time” or “save how you did that,” and Clio Work turns the preference into a reusable asset applied consistently across future matters, and, notably, across the firm, not just for the one lawyer who set it up. It’s available today to every Clio Work customer. The feature has four defining traits: no prompt engineering, since preferences are set in plain language during normal work; skills persist as a firm-wide asset rather than a single user’s private setting; they apply automatically to future matters of the same type; and they’re built on top of Clio’s library of 1 billion-plus primary legal documents, aimed at reducing hallucination and confidentiality risk.

The Flank read. Both stories are genuine progress: a bar association making AI accessible instead of leaving it to individual purchasing power, and a vendor making customisation cheap instead of requiring an engineering team. Neither one is a governance system. Free access to a tool and a tool that remembers your preferences both still assume a human is deciding, case by case, what gets escalated and what doesn’t, the exact infrastructure gap that keeps inexpensive, high-volume work sitting on expensive desks, because “we have AI access” and “we have a supervised system with review gates and escalation rules” are not the same claim. Flank starts from the second one: agents that already know your templates, terms, and escalation rules, with every output reviewed by a human before it leaves the system, access plus the governance layer, not access instead of it.

⚖️ 07 A New York appellate court hands down a $10,500 sanction for a Court of Appeals case that doesn’t exist, one entry in a global tally that has now passed 1,300

On June 23, the Appellate Division’s Second Department ruled on Landberg v. City of New York, a routine sidewalk trip-and-fall dispute. The case became notable not for the underlying facts but for the brief: it purported to quote a New York Court of Appeals decision, Xiang Fu Ji v City of New York, that does not exist, along with other misstated law and misrepresented holdings, the well-documented signature of generative AI used without verification. The panel, which had flagged the problem during May 20 oral argument, ordered sanctions under the state’s frivolous-conduct rule, 22 NYCRR 130-1.1.

[Include image: the sanction breakdown card from the Flank blog PDF, page 8: “Michael Sanders, Esq. (individually) $8,000”, “Law Offices of Michael S. Lamonsoff, PLLC $2,500”, “Total, payable to the Lawyers’ Fund for Client Protection of the State of New York $10,500”]

What distinguishes Landberg from earlier hallucination cases is that the court sanctioned both the individual attorney and the firm whose name appeared on the brief, a signal that courts increasingly see the failure as institutional, not just personal. It is also one entry in a fast-growing global record: Damien Charlotin’s publicly maintained AI Hallucination Cases Database now documents more than 1,300 court proceedings worldwide involving AI-fabricated legal material, of which roughly 496 involve licensed attorneys. Typical sanctions have climbed from around $5,000 in 2023 to $15,000 per attorney, with some 2026 rulings adding indefinite suspension on top of the fine.

The question worth asking. Landberg involved outside counsel on a routine matter, exactly the volume, lower-stakes work that gets the least verification scrutiny because it’s assumed to be low-risk. Does your organisation’s AI governance apply the same verification standard to routine filings as it does to your highest-stakes litigation, or does scrutiny scale with perceived importance rather than with whether a human actually checked the citations? The 496 attorneys in Charlotin’s database mostly thought they were doing something safe.

🧠 What this week tells us

Put together, this week’s stories price four things the market has mostly discussed in the abstract: what AI-run legal work is actually worth, what it actually costs to run at scale, what happens when nobody checks its output before it reaches a court, and how stable the model layer underneath all of it really is. A unicorn round for an outcome-billed AI law firm, a vendor’s own admission that its compute costs are compounding, an appellate sanction over a fake citation, a bar association’s free-access programme, and a frontier lab losing $225B in a week don’t share an obvious thread on their own.

Each story is a price tag on a different unknown. Norm Ai’s $1.2B valuation is a bet that a meaningful share of legal work is standard enough to price by outcome rather than by hour, which only works if that work is routed through agents built for exactly that repetition. Harvey’s 14x token growth is the cost side of the same coin: scale an AI system across everything a legal team touches, with no layer separating commodity volume from complex matters, and the bill grows faster than anyone budgeted for. The Landberg sanction is what happens when the review step that is supposed to sit between AI output and the outside world gets skipped, a $10,500 lesson that joins more than 1,300 others. And Google’s rebuild of Gemini 3.5 Pro is a reminder that every one of those vendors is itself dependent on a foundation-model layer that can move without warning.

✳️ The Flank view

Every story this week is really about the same missing layer. Inexpensive work keeps getting done by expensive resources: outside counsel billing by the hour, compute burning through tokens with no triage, courts absorbing the cost of unverified filings, wherever the infrastructure to route that work to supervised agents doesn’t exist. Norm Ai is proving the economics work when you build that infrastructure as an outside firm. Bar associations and vendors are proving that access and customisation are getting cheaper. None of it replaces the specific decision every legal department still has to make: to outsource legal work to supervised agents that already know your templates, terms, and escalation rules, with a human reviewing every output before it leaves the system, so the review step that Landberg skipped is never the one your team is missing.

✳️

Hi there, I’m Jake. COO and Co-founder at Flank.

Legal is the function every deal, every hire, every supplier waits on. I’ve spent eight years working alongside these teams. Almost everyone inside them knows it.

The strange part: because no one has ever removed the constraint, no one has ever priced what removing it would be worth. This is the argument for treating that queue as a design flaw, not a fact of nature.

Everyone waits on legal, and no one has priced what that costs

Legal is an operational bottleneck. Every deal, every hire, every supplier waits on it. That is a brutal thing to write about a function I have spent eight years working alongside, but it is true, and almost everyone inside these teams knows it is true. What is stranger is this: because no one has ever removed the constraint, no one has ever priced the upside of removing it. The gain does not appear in any business case, because there has never been a version of the company without the queue.

In April I published eight principles for how an enterprise legal function should be designed when execution is no longer constrained by headcount, and promised the full methodology would follow. It is now live at insights.flank.ai/method. This piece is the argument underneath it, and an honest account of the parts we argued about most on the way to publishing.

The tell

Start with how the queue shows up in the numbers. Gartner finds only one in five legal matters stays within budget. That gets read as a budgeting problem, and it is a structural one: the overflow model sends routine work to outside counsel at premium rates because there has been nowhere else to send it, so the overrun is priced in before the matter opens.

The tools have not touched this, because they solve a different problem. They are good, and lawyers genuinely like using them. But a copilot makes an expensive person slightly faster at inexpensive work.

A faster lawyer does not remove a bottleneck. The contract still passes through the same desk, joins the same queue, waits for the same person to come back from the same board meeting. The bottleneck simply gets a better interface. Moving the work is what removes it, and moving the work is a services question, which is why the budget picture matters. For every pound an enterprise spends on legal software, ten to fifty go on legal services: outside counsel, ALSPs, offshore delivery centres, contract reviewer pools, and the team’s own salary cost on routine work. The tools compete for the small line. The bottleneck lives in the large one.

A general counsel put it to us better than any positioning document could, explaining why they stopped shopping for a tool: “I don’t want my team looking at DPAs and NDAs anymore. I want it done. And I still want control.”

The analysts have started saying the same thing. Gartner now expects that by 2029 roughly half of contract reviews will be delegated to self-service systems that escalate only one in ten to a human, with legal tech budgets doubling to fund the shift. The forecast is not about faster lawyers. It describes the work moving.

⚡ The method in four lines

The Flank Method is long, deliberately, because it is a how-to guide as much as a point of view. But it reduces to four commitments, and if you read nothing else, read these.

First, legal is a bottleneck because the wrong people are doing the work. Expensive, judgment-trained professionals spend most of their capacity on routine, rules-based tasks. Remove that mismatch and the gain is structural rather than marginal, which is exactly why it has never been priced.

Second, the fix is less work reaching the lawyer, rather than a faster lawyer. Routine work routes to a supervised agentic system, so most of it never arrives on a legal desk at all. The staffing model changes. The backlog changes. The headcount does not have to.

Third, stay involved exactly enough to own the output. Human involvement exists to put responsibility somewhere a board, a regulator or a counterparty can find it. Sometimes that means checking every result. More often it means writing the playbook and auditing a sample.

Fourth, what remains is the work that always needed a lawyer. Contested terms, house positions, litigation, the deals that decide the company’s future. The lawyer moves above the process instead of sitting inside it.

The line that took longest to draw

The third commitment took us the longest to get right, and I think it is the most useful idea in the document.

Every conversation about agentic systems in legal eventually arrives at the same anxiety: who stands behind the output? The anxiety is earned. Icertis surveyed over a thousand US in-house practitioners this spring and found 47% would not detect an unauthorised or incorrect AI action until after it had occurred, sometimes weeks later. The lazy answers sit at the extremes. Review everything, which recreates the bottleneck one layer up. Or trust the system, which no serious legal team will do and no serious vendor should ask them to. The useful answer is that the right amount of human involvement is whatever lets a named person stand behind the result, and no more. For high-stakes work that might be every output. For a standard NDA run against a well-worn playbook, it is the playbook itself plus a sampled audit. Supervision falls where the risk is, and it has to scale faster than the volume does.

Share

Once you frame it that way, something clicks that I had not seen written down anywhere. Where that responsibility sits is what the words insource and outsource actually mean now. If your own team supervises the system and owns its output, the work is insourced. If a provider or a partner firm holds the standard on your behalf, it is outsourced. Same system, same agents, same playbooks. The only thing that moves is the locus of supervision. That reframe matters because it dissolves a false choice. Teams assume that getting routine work off their desks means sending it outside the building, with everything that has always cost them: the context lost in the handoff, the playbooks accruing to the provider, the institutional knowledge walking out the door. It no longer has to. The economics can match outsourcing while the judgment, encoded and owned, stays yours.

And around that supervision a new discipline is forming. Someone has to watch the overrides, notice the drift, keep the playbook honest. I find the most interesting open question is who does it, because I am not convinced it has to be lawyers.

🔒 What we have not solved

A method worth following should say what it has not worked out, so the published version ends with a section we nearly cut and I am glad we kept.

Junior development is the open problem I take most seriously. If associates do not grind through the routine work, where does judgment come from? Our current guess is that development shifts from doing the work to reviewing it, much earlier in a career than the profession is used to. I do not consider that settled, and I am suspicious of anyone who claims it is. The same goes for the optimal shape of a legal team once throughput is no longer tied to headcount, and for where the supervision standard lands as regulators catch up with all of this.

Share

These are live arguments, and the document says so. Where the Method does not match your experience, we would rather hear it than smooth over it.

Where this leaves you

The end state the Method describes is not a smaller legal department. A request arrives and is handled the moment it lands, drafted to the team’s own standard, with the few genuine exceptions flagged for a lawyer to decide in minutes. Lawyers spend their time on the work that was always worth it. Capacity stops scaling with headcount, and cost stops scaling with volume.

None of that requires a reorg, because the budget for routine work already exists, wherever it currently sits. It requires deciding that the queue is a design flaw rather than a fact of nature.

The full guide, including the practical chapters on intake, supervision, observability and adoption, is at insights.flank.ai/method. It is a working document and it will change as the practice does. Tell us where it is wrong.

✳️

Share

I’m Lorna. I lead Legal Alignment at Flank, and most of my week is spent supervising work an agent did first.

Most conversations about legal AI ask whether the output is right.

This piece is about a harder question: whether you can put your name behind it without redoing the work.

Share

The redline I couldn't sign off

A vendor agreement comes back from the agent looking clean. A few clauses redlined to our positions, a couple of fallbacks accepted, a liability cap left exactly where the counterparty put it. By every standard we set, the review is correct. And I still won’t send it.

I won’t send it because I can’t see why the cap was left alone. Maybe the playbook says that figure is fine for a contract this size. Maybe the agent missed it. From the document in front of me, those two possibilities look identical. So I do the thing that defeats the point of having had the agent do it at all, at least the way most of them are built today: I open the full contract and read it the way I would have read it if nothing had touched it. The few minutes the automation was supposed to save me become most of the morning.

I’ve seen this happen often enough, in my own work with agents and across the enterprises I’ve watched try to adopt them, that I’ve stopped treating it as a quirk. I think it’s the central problem in making legal AI genuinely useful, and that most of the industry is looking slightly to the left of it. We keep asking whether the output is right. The harder question is whether the person receiving it can stand behind it without redoing the work, and those turn out to be very different questions. I’ll say upfront that this is the question we spend most of our time on, so read the rest as someone who is invested in getting it right rather than neutral about it.

🧠 The thing I need is the reasoning, not the answer

A redline on its own is a verdict. It tells me what the agent decided and nothing about why it decided that, which means the only way I can verify it is to reconstruct the reasoning myself. That is the same labour as doing the review. No wonder it doesn’t feel like a saving.

What changes the maths is seeing the rule the agent applied. When the work arrives with its justification attached, “our indemnity fallback is mutual, capped at fees paid in the prior twelve months, so I accepted the counterparty’s clause because it already sits inside that limit,” the judgment I have to make shrinks. I’m no longer re-reviewing the contract. I’m checking whether I agree with the rule, and whether the rule was applied where it should have been. Both are quick. Reconstructing the whole thing from a silent markup is not.

The practical test is small. Before you approve anything, you should be able to answer two questions without reopening the source document: which rule produced this change, and was it applied in the right place. If the surface can’t tell you that, you don’t have a review. You have a guess.

There’s a version of this becoming a professional obligation, not just a preference. A proposal from the California bar, out for public comment earlier this year, would require a lawyer to independently review, verify and exercise professional judgment over any output produced with these tools. I find that framing clarifying. You cannot exercise judgment over a black box. You can rubber-stamp it or you can redo it, and a careful lawyer, faced with only those two options, will redo it every time. Visible reasoning is what creates a third option, the one where supervision is actually supervision.

Share

Correct by our standards is not the same as done

Here I’d argue against my own profession’s reflex, which is to read every change a lawyer makes to an agent’s output as proof the agent fell short. Sometimes it is. Often it isn’t.

The agent executes the playbook. The lawyer layers discretion on top of it: the read on this particular counterparty, the commercial temperature of the deal, the thing someone from the business said on a call that never made it into a written rule. A technically perfect review still gets changed because of all the judgment that lives outside the playbook, and that change is not a defect report. It’s the job. The human in the loop is where the lawyer’s discretion legitimately sits, rather than a fallback for a weak agent, and it would sit there even if the agent were flawless.

Once you accept that, you have to give up the idea that accuracy closes the trust gap. It can’t. There is no accuracy score high enough to remove the reviewer, because the reviewer was never only there to catch mistakes.

🔍 What the review costs is what decides everything

If accuracy doesn’t close the gap, what does? I’ve come to think it’s the cost of the review itself, measured not in minutes but in attention.

If marking up the agent’s work, or telling it where I disagreed, takes real effort, the kind that demands headspace I didn’t have much of to start with, I will quietly drift back into re-checking everything. And the moment I’m re-checking everything, the automation has bought the department nothing. The expensive resource is still doing the inexpensive work. We’ve just relabelled it supervision and told ourselves something changed. I suspect friction in the loop kills far more deployments than error rate ever does, and it does it silently, because nobody files a complaint about an agent they’ve simply stopped relying on.

This is why I think the review experience deserves the same design seriousness as the work being reviewed, maybe more, and it’s the conviction we’ve ended up building around. The test is whether a reviewer can respond at the granularity they actually think in, instead of being pushed back into drafting. In practice that means at least three ways to say no, each a different shape of “I still need to tweak this”:

• Reject a single change and give the reason in one line.

• Redline over the agent’s redline, when the position is close but not quite right.

• Comment without rejecting, when the point is a flag rather than a fix.

The reasoning belongs on that same surface, sitting next to the change it explains, so the decision and its justification arrive together rather than sending the reviewer off to hunt for one of them. And there’s a blunt way to tell whether the loop is working: a reviewer should get through the agent’s output far faster than they’d get through the same work by hand. The moment those two times start to converge, the agent has stopped earning its place.

Why an assistant can’t give you this

It’s worth being precise about why a general assistant, Claude Cowork, ChatGPT, a copilot in your Word ribbon, can’t hand you the same thing. I use those tools and rate them, so this isn’t a knock. The gap is structural.

With an assistant, you hold the pen. You paste the contract in, ask for the review, read the reply, then do the redlining and send it yourself. Ask it to explain its reasoning and it will, at length, because you’re sitting in the chat doing the asking. But nothing is ever waiting on your desk, finished, for a decision. The work was yours the whole time. There is no queue to clear, nothing to approve or dismiss, no exceptions to triage, because you never handed the task over. You went faster, which means you, the expensive resource, are still the one doing the routine work.

The supervision experience I’ve been describing only exists once an agent completes the work and hands it back for judgment. That is when reasoning on the surface, one-line rejections and a queue of exceptions start to matter, and none of it retrofits onto a chat window that assumes you’re still holding the pen. Getting that right is a different design problem, and it’s the one that decides whether the work leaves your desk or just moves through it faster.

🌱 Trust is a sequence, and the loop is how you walk it

None of this arrives at once, and I don’t think you can skip steps to get to the end faster. The pattern is fairly consistent, both in my own work with agents and in the enterprises I’ve watched move further down this road. At first you read everything the agent produces, in full, the way you’d check a new joiner’s work. Later you accept and dismiss without reopening the source each time. Eventually, on the narrow and well-behaved slices, the renewal reminders, the straightforward NDAs, you let it run and look only at the exceptions it flags.

You went faster, which means you, the expensive resource, are still the one doing the routine work.

The loop is what carries people from the first stage to the last. I’ve not seen anyone arrive at the far end because an agent crossed some accuracy threshold. They get there by reviewing, then reviewing a little less, until the reviewing itself tells them it’s safe to ease off. That’s worth saying plainly to anyone selling autonomy as a setting you switch on.

And there’s a dividend in this that’s easy to miss, which is the part I find most interesting to design for. When a reviewer dismisses the same kind of suggestion three times in a week, that’s a signal, and it usually means the playbook position is wrong or was never properly settled, rather than the agent misbehaving. Fed back in, those corrections tighten the rules, so the next month’s work is better because of the reviewing done this month. A well-designed loop compounds. A painful one just bleeds. The agents worth building are the ones that treat that signal as the point rather than exhaust.

Measure maturity on flow, not on the accuracy number

All of which points to a measure of maturity I’d trust more than a benchmark score. Picture two agents. One is right ninety-five percent of the time and miserable to supervise, every correction a small fight with the interface. The other is a little less accurate and a genuine pleasure to review, reasoning visible, decisions easy to express. I’d call the second one more mature, and I’d deploy it, because what determines whether the work leaves my desk is how well the person in the loop is served, not how high the accuracy figure climbs.

So when you’re judging an agent, the questions that predict adoption sit alongside accuracy rather than beneath it:

• Can you see the rule behind each decision, or only the decision?

• How many steps does it take to accept, reject, or amend a single point?

• When you disagree, can you say why in a way that feeds back into the playbook?

• Can you tell, at a glance, what it handled on its own and what it flagged?

None of those show up on a benchmark, and every one of them decides whether the work actually leaves the building.

The benchmark that actually decides adoption isn’t an abstract bar anyway. It’s whatever the team already does today. Where a contract is already reviewed in-house, the agent and its loop have to be lighter and faster than that existing review, not lighter than doing nothing. If supervising the agent costs more attention than the old way of just doing the work, the agent loses to the status quo, and it deserves to, however good its output looks in isolation.

I don’t think anyone has this fully worked out, and I’d be wary of anyone who claims they do. Most of the attention in legal AI, and almost all of the benchmarking, still sits on the quality of the output. I find myself more interested in the unglamorous part, the review, the rejection, the small act of putting your name behind a decision that something else drafted. That’s the part that decides whether the work genuinely leaves the desk and stays gone, or quietly finds its way back onto it a week later. The longer I spend on this, the more convinced I am that it’s where the whole thing turns, and it’s where we’ve chosen to put the work.

✳️