The Best Possible Client Experience in the Age of Legal AI
What a real client-first model looks like and the role of a legal quant
(This is the third piece in my legal quants series. The first (”Jane Street for Law”) argued the legal quant category exists. The second (”Origin Story”) argued they’re mispriced because no one has built the right environment around them.
This one argues that the legal quant’s real role is not to build tools, but to redesign the client experience around shared AI infrastructure, where clients, lawyers, and AI work in the same loop.)
How I got here
Many people know me as the lawyer who kicked off the vibe coding movement. People often ask how I got here, expecting some big origin moment. The real version is more boring than that.
I started with no-code. As a senior associate in BigLaw, the first internal tools I shipped were stitched together in Power Automate: document comparisons, PDF conversions, signature page collation. That wasn’t vibe coding. It was just no-code automation.
The first thing I vibe coded, letting AI write the code, wasn’t a legal tool. It was a website for my wife’s clinic. She commissioned me to build it the way she would commission any other vendor, and the real work wasn’t in the code. It was in the intention capture. What does the booking flow feel like? What colours, what tone? Weeks of back-and-forth on her taste and design preferences. Hours of gathering and curating photos and videos for the site. Refining details she couldn’t articulate until she saw a draft. She didn’t see the build until it was ready. The frontend came out genuinely good.
Legal tools came later. When Gemini 3 dropped and the agent loops got reliable enough, I started vibe coding them too, just for fun. Tabular review. Contract Simulation. I started sharing on LinkedIn because I wanted other lawyers to feel what I was feeling: the joy of using frontier models to build things that off-the-shelf tools cannot offer.
So quite naturally, my wife wanted to learn this new superpower I’d picked up.
A café in Seoul
We end up in a café in Seoul. She’s next to me with her laptop open. The thing she wants to build this time is bigger than the website. A patient management system. Database tables for patient records. Appointment scheduling that talks to the clinic site. A record of where each needle landed, session after session. A clean view for the front desk. The kind of system a small clinic would normally pay a software vendor mid-five figures for.
She wants to learn how to code, so I’m teaching her. Except I’m not letting her type anything.
What she doesn’t know is that I have Willow Voice installed and the FN button on my MacBook is wired up to it. She’s looking at the screen, and I’m holding the button down. She talks to me. “I want patients to be able to log in and see their next appointment. And I want the front desk to be able to override the schedule when someone walks in.” What looks to her like me listening is actually my Mac silently transcribing her natural speech into a clean, well-structured prompt and sending it to the Replit agent.
I let go of the key. Replit starts building. Her eyes light up.
Then she leans in. “Oh, can it do this? What if a patient wants to reschedule from their phone? Can it remind them the night before? What about the doctor’s view, does that work? Can we make it bilingual?” The follow-ups come faster than she would ever have typed them. I hold the FN button down again. Each question becomes another instruction. We test the limits of the tool live, on the fly, without her ever touching the keyboard.
Somewhere along the way her hand finds the FN button on its own. I’m not sure exactly when. From that moment, she’s the one pressing it, and I’m out of the loop.
By the end of the morning, she has a working patient management system. She built it. She started by talking to me. She finished by talking to the model herself.
That’s where the model I now apply to everything crystallised. My wife had no coding background. She had clear intent. The AI did the execution. Someone sat in the middle, watching her preferences become a product in real time.
A co-working space in Tokyo
A few weeks later, in April, I’m in Tokyo meeting my co-founder for the first time in person. We’re sitting in a co-working space trying to write our founder’s agreement.
This is not an easy conversation. Equity. Compensation. Runway. Commitment. What happens if one of us burns out. What happens if it works. What happens if it fails. We had spent the previous night just talking, an hour of unstructured conversation about our finances, our personal constraints, and what we each wanted out of the company. We recorded it all.
The next morning I dumped the transcript into Claude Code and asked it to do two things: analyse where we stood, and ask us the questions we hadn’t thought to ask each other. We turned on the “ask user question” tool so Claude would interview us directly, multiple choice, with options to click.
It was almost clinical. “If your co-founder did X, how would you feel?” Four options, A through D. Then the same question, flipped, to the other one of us. “If Jamie did Y, how would you respond?”
Twenty questions in, Claude had a sharper read on our preferences and our hard limits than either of us had managed to articulate alone. It produced a term sheet that genuinely reflected what we wanted, not the generic founders’ template that everyone signs and then quietly resents.
What happened in that room generalises the café scene. There were two humans with commercial intent and a real relationship to protect. The AI sat between us, surfaced assumptions neither of us had voiced, translated soft answers into structured terms, and produced a working document we both trusted.
That, to me, is the best possible client experience. The client, the counterparty, and the AI in the same loop, with somebody supervising the workflow.
Now look at law firms
The problem is that almost no law firm does this.
Two pieces from Artificial Lawyer this year sit uncomfortably side by side. A Thomson Reuters survey showed that in 2025 most major law firms made more money than ever before, with AmLaw 100 profit per lawyer rising more than 53% since 2019 and tech spending up another 10% on top of an already long-term upward trend. A month later, a senior in-house lawyer told a room of 500+ GCs in Stockholm what she thought of all that investment: “We get nothing. They haven’t changed and probably next year the same work will cost even more.”
Both can be true. The firms are spending on AI. The clients are not feeling it. The efficiency, wherever it’s going, is not coming out the other end of the invoice.
If AI is so transformative, why does the client experience look identical to 2019, except more expensive? Where is the efficiency landing? And what would it look like if a firm decided, from first principles, that the client should get to feel some of it?
There are three possible answers. One is what’s mostly happening. One is better. One is what I’d call the best possible client experience.
The worst case: AI as hidden leverage
The worst case is the most common one.
An associate uses AI to compress eight hours of work into one. The client doesn’t know. The billable target doesn’t move. The invoice looks the same. The work product may even look the same. The associate becomes, in effect, a human wrapper around the model — prompting, polishing, checking, and sending the output up the chain.
Often the partner doesn’t know either. The associate hands up a polished draft that looks like the product of a long night. The partner edits the margins and sends it on. The leverage is invisible inside the firm before it becomes invisible to the client.
I know this is happening because some of the people doing it are in the Legal Quants community I built. They are very good at this. They have figured out how to configure Claude Code with agent teams, shared skills, and memory files in ways their practice groups do not understand. They can do 3x, 4x, sometimes more, in the same time. They could be billing more. They aren’t. They’re going home early.
Why? Because they know the firm won’t pass on the upside. The hourly target stays at 1,800. The matter budget was set before AI. Doing more work would just mean more work. So the rational move, from inside the cage, is to do the assigned work faster and reclaim the evening.
This isn’t a moral failure on the associate’s side. It’s the system absorbing the gain at the lowest level because nothing else in the system was redesigned to receive it.
The same thing is starting to happen on the client side. In-house teams are using AI internally without telling the firm. The firm is using AI on the response without telling the client. Both sides pretend the work is being done in the old way. The client sends AI-generated instructions. The lawyer sends AI-assisted output back. Nobody on either side knows what was machine-generated, what was checked, and which parts of the analysis are just plausible-sounding synthesis.
That is AI slop moving back and forth between two professional parties, with legal fees layered on top.
It is not innovation. It is opacity with better software.
The middle case: AI as pricing advantage
The middle case is better. It’s also where the most thoughtful firms currently are.
A partner understands that AI has lowered the cost of delivery. Instead of billing the old way, the firm quotes a lower fixed fee. It undercuts competitors. It wins work it wouldn’t have won before. The client pays less. The firm still makes good margin because the work is genuinely cheaper to produce.
I lived this. At my old firm I had a software-focused fund client come in asking for an MFN election, the process where a new investor in a fund reviews the side letters of every earlier investor and elects to inherit any more favourable terms. Historically this kind of work was priced in the high five figures. It’s repetitive, document-heavy, and exactly the shape of task where an associate spends a week reading two hundred side letters and tagging clauses.
The client said: not this time. They wanted it capped at around twenty thousand.
My old boss turned to me and asked, “Jamie, can you AI this?” I said yes. I spent a week building an agent that ran autonomously for eight to ten hours over a weekend, processed everything, and produced the election with full reasoning. We came in well under cap, with 100% recovery. The client was extremely happy with the output.
In the moment, I thought we should productise this and own the MFN election market. In hindsight that was a tool-builder’s instinct, not a business one. Distribution matters more than the tool. The market doesn’t reward whoever builds the agent. It rewards whoever the GC already calls when an MFN election comes in.
But that aside, what we did was a real win on the second-best terms. I was happy because no one was writing off my time. The partner was happy with full recovery on a smaller fee. The client was happy with the output and the price.
It is also not the best possible client experience. It’s a cheaper version of the old workflow. The firm captured the work by being faster and cheaper than the firm down the road. The underlying relationship (what the client buys, how the work is scoped, where their team plugs into the loop) was untouched.
The firm is still using AI inside its own walls and deciding how much of the benefit to pass on. That decision is private. The next conversation about scope and pricing starts from the same opaque baseline.
The best case: AI as shared infrastructure
The best case is what happens when a firm decides to stop hiding the system.
I had a conversation recently with an in-house lawyer at a bulge-bracket bank. She told me, almost in passing, that when a legal question comes up internally, she now drafts the answer herself, with AI, end to end. She gets praise within the bank for being efficient. Then she sends the finished memo to her firm partner and asks for a yes or a no.
This puts the firm partner in a strange spot. What is he charging for? He isn’t doing the research. He isn’t doing the drafting. He isn’t doing the analysis. He’s reviewing a finished work product and giving it a thumbs up. Is that worth $2,000 an hour? Is it worth anything at all? It’s the most distilled possible form of judgment, with none of the execution that usually justifies the price.
If I were the partner, I wouldn’t try to defend the old number. I’d reframe the relationship.
I’d ask her to send me not just the final memo, but the prompt history that produced it. The whole thread. I want to see what she asked, what the model came back with, where she pushed, where she stopped pushing, and what she decided to keep.
Because the memo isn’t the only thing I’m reviewing anymore. I’m reviewing the reasoning trajectory. The memo might be perfectly right and her direction of inquiry might still be slightly off. She might be answering a question she didn’t realise was the wrong one. She might be missing a layer of risk that the model didn’t surface because she didn’t know to ask. Or she might be perfectly on the money, in which case I can sign off in five minutes and we both know the answer is robust.
That is a different product. It’s not a memo. It’s a meta-review of how she got there. And it’s something only an experienced lawyer with both legal judgment and AI fluency can credibly do.
One step further. Picture the in-house lawyer and the firm partner in the same room, or the same shared workspace. Client asks the question. AI takes a first pass, usually 80-90% of the way there. The lawyer reads it, says yes / no / tweak this clause / you’re missing this risk. The corrected output becomes the working answer. The client asks the next question. The loop continues.
This is exactly what happened in the Seoul café with my wife and in the Tokyo co-working space with my co-founder. Two humans with intent, the AI surfacing assumptions and producing first drafts in real time, somebody experienced supervising the loop. The only thing that changes when the lawyer is in the room is that the supervisor now has legal training, malpractice insurance, and a real view on what the firm has seen across other matters.
That partner is entitled to bill for the full hour. Not because they did everything themselves, but because their judgment is what makes the loop reliable.
This is the model that creates a stickier client relationship than anything else AI has produced. The firm isn’t selling hours. It isn’t even selling documents. It’s helping the client build a better legal posture.
Preference discovery: where this matters most
The place this changes everything is transactional work.
In a deal, the lawyer’s real job is not to produce the heaviest markup possible. It’s to understand the client’s commercial preferences and bottom lines, then express them in a way that protects the client without killing the deal.
The bottleneck has always been preference discovery. In the traditional process, commercial intent travels through too many layers before it reaches the document. A business person explains the position to in-house. In-house reformulates it for external counsel. The partner interprets it again and hands the drafting to an associate. The draft travels back up the same chain, each person reviewing the legal expression of a commercial point that may already have been softened, narrowed, or quietly misunderstood. By the time the issue appears in the contract, the drafting is competent but slightly detached from what the business meant.
That is the translation loss AI can collapse.
Imagine the business team sitting in front of a model before drafting starts. The AI runs a structured interview. What outcome do you actually need? Which points are non-negotiable? Which are preferences dressed up as positions? What’s commercially acceptable risk? What approvals would each fallback require? Where do you want to be aggressive? Where do you want to preserve the relationship? Where would you accept a worse term to close faster?
The business team answers. The AI turns the answers into a term sheet or an issues list. The lawyer then reviews it, tests the assumptions, identifies legal risk, and turns it into something usable.
The value isn’t that the AI drafts the document. The value is that the client and the lawyer share a clear, externally-readable view of the client’s actual position before drafting begins. Most of the back-and-forth that bills out as “scoping” or “clarification” disappears, because the working assumptions are now visible to everyone in the room.
Cut the middlemen, and the middle-AI
When I lay this out, people often jump to a different question. What about AI-to-AI? My agent versus your agent, arguing the contract in the background while we sit on the sidelines? Counsel on both sides instructing their models, the models exchanging markups, the principals reviewing the score at the end?
This is the fantasy that gets the most airtime in legal AI conversations right now. People imagine two adversarial agents negotiating like litigators, finding the zone of agreement faster than humans could, exchanging redlines while we drink coffee.
If opposing counsel sends an AI avatar to negotiate with me, I’m dropping off the call.
Not because I’m a luddite. Because it’s a worse design dressed up in newer technology. AI-to-AI imports the layered translation chain we already have, business to in-house to outside counsel, and adds another layer in the form of two agents that have to be configured by two more humans. Each layer is another place to lose what the principals wanted. You’re not removing middlemen. You’re adding middle-AIs.
It also smuggles in the wrong frame. Adversarial agents make sense in a courtroom, where the goal is to win. In a deal, the goal is a signed agreement that reflects what both sides commercially wanted. A heavy AI-generated markup with no commercial purpose isn’t strength. It’s noise. It slows the deal. It damages the thing the client hired you to protect.
The version that works isn’t AI-to-AI. It’s the version from the Seoul café and the Tokyo co-working space, scaled up.
The principals, the people whose deal it is, sit in the same room. The lawyers are present. There’s one AI in front of them. Preferences get surfaced live, in front of the people with authority to act on them. Drafts get produced in real time. The lawyer pushes back where legal risk lives. The business folks resolve where the business reality lives. The other side does the same on their end, and then both teams meet, in the same room and in front of the same AI, to close the gap.
That’s the first-principles solution. Cut the middlemen. Cut the middle-AI. Put the humans who own the decision in front of one shared tool, with experienced supervision in the loop, and let the machine do what it’s actually good at: surfacing assumptions, producing first drafts, and translating intent into structure.
We don’t need AI-to-AI. We need fewer translation layers.
The legal quant’s job
The Seoul café and the Tokyo co-working space worked because somebody was in the middle. My wife had clear intent and no code. My co-founder and I had clear intent and no formal structure for capturing it. In both rooms, AI was useful because somebody was actively translating intent into structured prompts and steering the model when it drifted. The role of the person in the middle is what mattered, not the fact that the person was me.
The reason none of this happens at most law firms is that nobody currently holds that seat. Partners are optimising for margin inside the existing pricing model. Associates are using AI in private to hit their hours faster. In-house counsel is procuring within an assumption that the work has the shape it had in 2019. Nobody is sitting across the relationship asking which parts of it should survive AI and which shouldn’t.
That’s the seat I’ve been calling the legal quant. The job is to look across the whole picture and decide which work belongs to the client, which to the lawyer, which to AI, and which to a redesigned process that doesn’t look like any of them. The starting question is what the work is for, not how to automate it. A memo, a markup, or a call may be the right output. It may also be a substitute for something better: a decision tree, a supervised AI exchange, a structured intake, a clearer escalation rule. Sometimes the deliverable itself is the wrong unit, and noticing that is part of the job.
The analogy I draw is to a quant trader sitting across a market, not to a legal engineer optimising one workflow at a time. This isn’t about coding as identity. It’s about leverage. And it’s broader than “lawyers who code” suggests, running from tool selection through professional responsibility, from build craft through firm adoption.
The people who can do this work exist. They’re rare, mostly invisible to the legal hiring market, and the firms that need them most usually can’t tell them apart from any other lawyer who uses AI. That’s the gap LegalQuants is meant to close. We’ve reviewed around 650+ applications and admitted roughly 120 lawyers who ship in their practice. They form the network. They critique each other’s builds. Increasingly, they’re also hiring each other directly, because it takes one legal quant to recognise another and put them in a seat where the redesign can happen. The standards are forming as the group chats revolve around common themes. The vetting is institutionalising into a real screen with explicit criteria, rather than my taste.
I’ll talk about what it takes to be one of them, and how a lawyer in a current firm gets there, in the next essay.
Jamie Tso’s essay brilliantly captures the shift from AI execution to shared infrastructure, but it exposes a massive governance gap: when the cost of producing legal work drops to near zero, the liability of certifying the output skyrockets. This is the literal reality of the Verification Tax—the economic and professional burden of ensuring autonomous agents comply with statutory standards.
To prevent this "vibe coding" movement from devolving into an uninsurable epidemic of algorithmic slop, the Legal Quant cannot rely on intuition alone; they must implement a strict Triple Audit Framework:
Legal Fidelity: Auditing the agent's core logic against statutory realities rather than statistical probability.
Procedural Fairness: Ensuring that every real-time adjustment within the shared workflow is completely transparent and contestable.
Linguistic Inclusion: Rooting out "Legal Logic Colonialism" to prevent Western-trained models from surreptitiously overwriting local jurisdictional nuances.
As Jamie notes, we don’t need adversarial AI middlemen; we need experienced human supervisors in the loop. The true mandate of the modern Legal Quant is to transition from a mere user of tools into a system architect who treats rigorous verification as the ultimate signature of legal legitimacy.