We Built the Firm a Memory

Series A documents came in for a client last week. The founder and I got on a call and walked through them — what to push on, what to concede, what had to change before any of it went back across the table. Ordinary work. The interesting part is what happened next.

At a large firm, the next step is an associate spending hours rebuilding context: re-reading the email thread, pulling the prior drafts, checking the notes so that nothing promised on the call quietly fails to appear in the document. That reconstruction is genuine work, and it gets billed as genuine work. It exists because the context lives in whoever happened to touch the matter, and it has to be reassembled by hand every time someone new picks up the file.

I asked our system to do it instead. Review the correspondence with the client. Review the notes from the call we just finished. Review every prior iteration of the documents. Then revise the drafts to reflect what the founder and I decided, drawing on every financing this firm has papered before this one. The turned documents came back in minutes, grounded in the full history of the matter.

That’s the firm’s side of it. The founder’s side is simpler: documents back the same day, a stronger position on the terms that mattered, and no line item for the hours of context reconstruction that normally pad a document turn. That is what the memory is for. It puts senior judgment in front of a founder quickly, at a price we can name in advance.

In the last thing I wrote, I argued that the work is collapsing in cost while the judgment isn’t, and that what the profession does with the difference is a choice rather than a forecast. I mentioned in one line that we had built the firm a memory, and then set the mechanics aside because they weren’t the point of that essay. They’re the point of this one.

The individual lawyer is the first chapter, not the whole book

If you’ve read Zack Shapiro on practicing law with Claude, you’ve seen the individual version of what we’ve built. One lawyer encodes his own judgment into reusable instructions, hands first-pass drafting and review to a model under his supervision, and produces the output of a much larger team. We do this too. We trained our system on how we review contracts and mark up a term sheet, so the first pass starts at our standard rather than a generic one.

Shapiro makes a point I think is exactly right: the leverage lives in the individual lawyer’s encoded judgment, not in some vendor’s idea of a firm template library. And he’s right about the harder limit too. You can feed a model the record of every deal you’ve ever done, but not the experience of having been in them. Michael Polanyi called that tacit knowledge — the things we know but cannot fully write down. It does not transfer cleanly to a machine, and it may never.

So we didn’t try to bottle our judgment and walk away. What we built is a memory, and it goes after a problem the individual approach leaves entirely untouched. When the record of what a firm has done lives in whoever happened to do it, the knowledge leaves when that person does. A vacation, a busy week, a departure, and it’s gone. That is the oldest problem in professional services, and giving every lawyer their own AI does nothing whatsoever to solve it.

A firm is two assets

Plenty of legal-AI companies sell a version of what I’m about to describe — the firm brain, the platform that makes you smarter. I’m not selling software. I built this inside a working firm and use it on live deals, which is a different exercise from building a demo.

Strip away the letterhead and a firm is two assets. The first is everything it has ever done: every agreement drafted, every deal negotiated, down to the call someone made under pressure at eleven at night. The second is how it decides — the positions it takes on a given term, where it pushes and where it gives way, what it treats as market in a sector this quarter. At almost every firm both of these live in human heads and scattered files, and neither one is searchable.

We made them searchable. The system has read what the firm produces: the document management system, our email, our meeting notes, our Slack. So when one of us has a question, we ask the firm instead of asking our own recollection. What’s our standard position on a valuation cap? Have we seen this protective provision before, and how did we handle it? The answer comes back grounded in our own work, with the source attached.

It only holds what we have actually done. It was thin when we started, and it gets deeper with every matter we close. That’s the part that makes it an institution rather than a product: nobody can buy the version of it that took us years of deals to accumulate.

The obvious question, and the line we drew

If a system has read everything a firm has done, the first thing any careful founder or general counsel will ask is what happens to confidentiality. It’s the right question, and it deserves a real answer rather than a reassuring one.

The system runs on enterprise infrastructure that does not train on our data and does not retain it outside our control. Client information stays inside the firm’s confidentiality perimeter.

Beyond the infrastructure, we treat anything that draws on prior client work as governed by the duties that have always applied — confidentiality and our obligations to former clients, Model Rules 1.6, 1.9, and 1.7, and the ABA’s 2024 guidance on generative AI in Formal Opinion 512. The ABA is explicit that a boilerplate line in an engagement letter is not informed consent for this. So our engagement letters explain in plain language how experience from prior matters may inform later work, and we get actual consent.

The case I think about most is the temptation to tell a founder what is market in their sector this quarter. Lawyers have always carried that knowledge from doing the deals, and a memory makes our own experience far easier to recall. But there is a line between drawing on our judgment and pulling a specific client’s confidential terms into someone else’s negotiation, and we don’t cross it. Rule 1.9 is clear that you cannot use a former client’s information to their disadvantage. Where a matter touches a client’s competitor, we wall it off. Where we can’t answer without exposing a confidence, we say so and stop.

I’d rather state that plainly than let the capability speak for itself, because the failure mode here isn’t exotic. It’s a firm that builds something powerful and never decides, on purpose and in advance, what it won’t do with it.

What’s live, and what we’re still building

The memory and the positions are live and in daily use. The rest is in progress.

Think about what a firm like ours does over and over: qualify a new matter, form the company, paper a financing, review an inbound commercial agreement, run diligence on an acquisition, turn a board meeting into minutes, keep filings current, get the bill right. Each of those has a recognizable shape — a precipitating event, a body of routine work that follows, and somewhere inside it, a small number of calls that actually matter.

We’re turning each one into a supervised loop. The system notices the trigger, does the first pass against the firm’s memory and positions, and hands a lawyer a draft that already starts at our standard. A lawyer reviews and signs off, or sends it back. Nothing reaches a client that a licensed attorney has not reviewed and approved. And we’re building each loop to learn, so that every correction a lawyer makes improves the next matter rather than evaporating.

What the memory makes possible past that is the part I find genuinely new. A negotiation brief showing how a specific opposing firm has actually moved across every deal we’ve done against them — where they open, and where they will not budge. A shortlist of the right investors for a founder about to raise, drawn from years of relationships and how those introductions actually played out. None of that is possible without the memory underneath it, and each one gets better every time we close another matter.

The pyramid becomes a rectangle

Start with the honest version of the numbers. My own estimate is that the system makes me roughly a third better at my job. That understates it, though, because the counterfactual isn’t me doing the same work more slowly. Without it I would run fewer matters at once, and I would drop things, because the context would live only in my head and my head cannot hold all of it. The point of a memory is that it doesn’t have to.

The math on hiring changes next. The work that used to justify the next associate — first-pass review, research, initial drafts, digging precedent out of old files — is increasingly done by the system under supervision. To be clear about what that means: every document that leaves here is reviewed, revised, and signed by an attorney. The system produces a first pass. We produce the work product. So our next hire is about judgment, client relationships, and the supervision this thing requires, not another two thousand billable hours of production.

Change the math and you change the shape. The traditional firm is a pyramid: partners on top, associate leverage underneath, profit flowing up from hours billed below. We’re building a rectangle — senior lawyers made more productive by the system, with junior hires doing what the technology can’t. The firm grows by getting smarter rather than by getting bigger.

That is also what makes different pricing real rather than aspirational. When every hour of attorney time produces more, and routine advisory work is substantially handled by the system, fixed-fee and subscription pricing stop being a marketing posture and start being arithmetic. For a growing company the payoff is counsel you can call without watching the meter, which is what clients have wanted from lawyers for about a century. I argued last month that the failure mode of this moment is capture — firms adopting AI internally, doing the same work in a fraction of the time, and keeping every dollar of the difference. This is the specific machinery by which a firm declines to do that. You can’t pass the collapse through to clients if your economics still depend on selling the hours it eliminated.

What stays human

AI is not practicing law. We are. It makes us faster and more thorough, and it leaves the decisions that matter where they belong. If you’ve spent fifteen years building that judgment, you’re holding the exact asset this technology makes more valuable.

So we drew one more line deliberately, and we wrote down which parts of this work will never become a loop. The hard conversation with a founder about a cofounder split, or a layoff. The negotiation that turns on reading the room. The deal structure none of us has seen before. Putting your name on a delicate call that could go either way. Those stay human, permanently and by design — not because the technology might get there eventually, but because the willingness to be accountable for a decision is the thing a client is actually buying.

Why I’m writing this down

I think the profession is about to divide into firms that understand this and firms that don’t, and I’d rather show the work than let that sort itself out quietly.

Most of the legal-AI conversation is still about a better chatbot for one lawyer. That’s a real gain and it isn’t the interesting question. The interesting question is whether a firm can become an institution that remembers what it has done — that doesn’t lose its accumulated judgment every time someone leaves, and doesn’t charge clients to rebuild context it should never have lost — while still leaving the calls that matter to the people whose names are on them.

We’re four lawyers who routinely sit across the table from firms with hundreds. We hold our own because the firm itself remembers. That’s the whole argument, and the rest is just building it.