Mind your business.
Not just your practice. A blueprint for Firm AI By John Hall and Thad Jampol

A corporation can mandate. A partnership has to convince. And a firm's work runs on projects, not departments. These facts shape how AI has to be built for a firm.

This holds at every scale, from an Am Law 100 firm with 1,200 partners to a private equity manager with 30. Professionals organize by expertise, not department, and each practice, sector, or strategy runs as its own business inside the firm, with its own clients or portfolio and its own economics. Partners own the firm, and they hold the relationships that bring in the work. They can walk, and when they do, they take entire practices, deal teams, or client books with them.

So Firm AI cannot be installed. It has to be adopted. There is no central switch a CEO or managing partner flips for the partnership. The software earns its place partner by partner, strategy by strategy, adapting to how each one already works rather than imposing a single firm-wide model. Adoption is not a rollout plan bolted on at the end. It is a property of the product, and it is the first thing Horizontal AI, built for traditional corporations, gets wrong.

The second difference is just as important. A firm's unit of work is the project, not the department. Deal teams assemble, diligence, negotiate, and disband. Matters get staffed, executed, billed, and closed. Funds raise, deploy, manage, and wind down. A firm is not a static org chart. It is a constantly re-forming network of project teams, each with its own members, obligations, and constraints.

So a Firm AI agent has to understand the project — deals, matters, engagements — not as a folder of documents, but as the unit that carries the rules of the work: its own access permissions, often complex and overlapping. Its own governance, what must be archived, retained, or turned into firm knowledge. Its own client, investor, and regulatory obligations. Its own economics, the margin or return it has to protect. All changing constantly as teams form and reshape.

None of this is uniform. Inside a single firm, a litigation matter and a regulatory investigation, a buyout and a private credit deal, an audit and a fund formation carry entirely different permissions, obligations, and economics. The firm runs dozens of operating models at once, one for every kind of work it does. An agent that applies a single logic to all of them gets all of them wrong, breaching a permission, missing a retention duty, or doing work the engagement was never scoped to pay for. Horizontal AI, seeing the firm as one undifferentiated business, has no idea there is anything to tell apart.

Firm AI starts from both realities. It respects the nudge, not the mandate. It orchestrates the project, not the department. It gives partners the autonomy they expect while giving the firm the oversight, guardrails, and accountability it requires. Many vendors trained on rollouts for other industries mistake this for a technology problem. It is not. It is how the firm is governed and how the work gets done. Firm AI is the only category built for the partnership and the project from the start.

A firm is two things at once. There is the work product itself: the contract, the audit, the brief, the deal model, the craft of the most capable and entrepreneurial people in the economy. And there is the firm itself: the structure that makes its people more than a roomful of sole practitioners. The firm is the platform. It gives a hundred individual experts one name, one standard, and a reach, scale, and breadth of offering no one of them could build alone.

Running that platform is the "business of the firm": how the firm develops business, originates deals, prices new work, staffs the right people, and collects the time it is owed. In a law or accounting firm, the work product is the obvious center, and the business of the firm surrounds it. In private capital, the deal professionals spend much of their day sourcing, screening, and diligence — the pre-work that gets a recommendation in front of the partners. Different shapes, same structure. Practice AI accelerates the work. Firm AI is built for the business of the firm.

And this is the half that increasingly decides who wins. Two firms can have equally brilliant professionals and equally good work product. What separates them is which one runs the business of the firm better: which one draws on its collective relationships, finds the better deals first, holds its margin under pressure, puts the right team on the right matter, and clears conflicts in hours instead of days. Growth, profitability, and reputation are won or lost here.

So why have AI companies ignored it? Because it is the hard half. Practice AI helps a professional review or produce work product: a contract, a spreadsheet, a memo. The business of the firm is not work product. It is the firm's whole operating fabric, and serving it means orchestrating processes across all of it. That is far harder than a better assistant, which is why the industry took the easy half and left this one alone.

And the business of the firm is expensive. Firms run on people, and so they scale linearly: more revenue, more operational headcount, flat or compressing margins. In a large firm, 15 to 20 percent of revenue is absorbed by business services headcount. The COOs in the room are no longer subtle about it. "We need a better operating model. We scale too linearly today." "My MP yells at me every week to reduce heads." "Practice AI is a race to the bottom. The business is where we can drive value." The managing partners have stopped asking nicely.

And even with the high costs, there are never enough hands to serve everyone well, so support flows to the rainmakers and the squeaky wheels and everyone else waits. High cost and uneven service: the worst of both worlds.

Firm AI is built to solve this. It does not replace Practice AI. It is complementary, built to augment and amplify the specialists running the business of the firm, not replace them. You cannot hire a seasoned conflicts analyst from a tech company, because the job does not exist there. Firm AI gives every professional the support once reserved for rainmakers, and it takes back the tax of belonging to a firm: the timesheets, the conflicts forms, the intake and compliance steps — none of it client or deal work, but all of it required. Firm AI takes it off their plate and hands back the hours.

The result is a firm that scales without growing its headcount, and serves its people better. The firm that builds for the business of the firm wins.

Today's AI is built for the professional at their desk: the associate reviewing the contract, the accountant generating a tax return, the analyst building a financial model. Practice AI makes that professional faster, and that is worth doing. But it leaves untouched the people whose job is not the work but the firm: the managing partner and CEO, the COO, the heads of practice, industry, and strategy.

Theirs is one of the hardest jobs in business. They are accountable for the whole firm, its growth, margin, returns, risk, and direction, but cannot run it the way a CEO runs a company. A partnership cannot be commanded. A firm leader governs by influence: herding partners who each own a piece of the firm, steering without orders, building consensus for moves a corporate executive would simply mandate. No AI has ever been built for that job.

Firm AI is built for exactly that.

It gives the leader a direct view of the firm. Today a managing partner learns how the firm is doing by asking the people below him, and gets back an answer that is slow, partial, and shaped by whoever assembled it. Firm AI sits on top of everything at once: the firm's memory, its matter and deal history, its pipeline, its conflicts and risk posture — all governed and tied to the same institutional facts. A managing partner who asks how the firm is performing in middle-market private equity this quarter gets a full-firm answer that is auditable. He no longer has to receive that filtered through multiple individuals. A GP who asks which LPs are most likely to anchor the next raise gets a ranked list that respects side-letter and MNPI boundaries, in seconds. A consulting practice leader who asks which sectors are under-resourced against pipeline gets a view across every engagement, methodology, and partner's calendar — from the firm's own systems, not approximated. It is a completely new paradigm and interface for firm leadership to not only understand their business better, but also manage it better.

It lets the leader steer without commanding. Because a partnership runs on influence, the most powerful thing a leader can do is make the firm's preferred way the easy way. Firm AI encodes the firm's chosen approach into the agents and playbooks partners already use, surfaces who is working with the grain of the firm and who is drifting, and turns a directive that would have been resented into a default that is simply there.

Every other category of AI helps somebody do their job. Firm AI is the first built to help the firm's leaders and partners run the firm itself.

Ask the managing partner of any firm what they care about most, and the answer is always growth: of the top line, the partnership, the next fund. Every other priority reports up to it. Firms that stop growing lose partners, and then the practices those partners built.

Every firm runs into throughput governors that constrain its growth, set by how many hours its people have. A private equity manager can screen only as many deals as its analysts can get to. A law firm can process only as many conflicts as its team can clear. These limits force a constant trade-off between throughput, quality, and cost. Want to do more, faster, without dropping the bar? You cannot, because doing more has always meant adding more people.

Firm AI loosens those governors with agents that scale with demand instead of headcount, and without lowering the bar, because the hard calls still route to a person. And an agent costs a fraction of a salary, based on compute and token consumption. So, a conflicts clearance playbook can suddenly run continuously instead of monthly. A KYC playbook can run always-on against sanctions and beneficial ownership changes. A deal screening playbook can run the manager's investment thesis against ten times the volume of inbound. A cross-sell playbook can surface the right opportunity to the right partner at the right moment — turning an audit relationship into an advisory mandate, or a first-fund LP into the anchor of the next vintage.

The economic unlock here is bigger than people think, because it is not just automation or productivity. It makes possible what was previously infeasible. Imagine a conflicts team that cleared 200 matters a month and could now clear 2,000. Or a fund that screened 300 deals a year and could now screen 3,000. Firms who would have filtered out a whole range of opportunities in adjacent segments for practical reasons will now be able to vet and assess them at scale. Those are new business models. Practice AI makes individuals faster. Firm AI makes the firm bigger.

A general-purpose agent does many things adequately and nothing specifically. An expert coworker agent — built around the vocabulary, obligations, and workflows of a particular profession — does the specific thing the firm actually needs. A deal screening agent that knows what a quality-of-earnings flag looks like. A conflicts agent that walks a corporate tree the way a senior risk lawyer does. A lateral-hire vetting agent that compares a candidate's book of business against the firm's client roster, conflicts map, and cross-sell opportunities. In software, expertise has always come from specificity. The general agent is a generalist. The coworker agent is a colleague who already knows the work.

And it starts expert and stays expert. It needs no ramp: it arrives fluent in the work on day one. And it does not turn over. The hardest thing any firm function does is train a person, season them, and watch them leave. Expertise walks out the door every year. The agent does not. The senior partner who has the final say at the investment committee can now bring that scrutiny to every CIM that comes in, screened by an agent that has read every IC memo the firm has written, the deals pursued and the deals passed on, and flags what an experienced eye would catch.

A firm with a large IT team or a DIY culture may ask why it cannot build this directly on top of a model itself. It can build something. But encoding a firm's vocabulary, obligations, and methods into a general agent is years of work from scratch, and that work is already done. Building instead of buying means spending those years getting to where the category already is.

This is the payoff of digital transformation that the first wave could only partially deliver. A decade ago, firms bought workflow tools that needed scaling numbers of data stewards and analysts to operate them. The cost compounded. The next wave is not that. An agent running the conflicts, intake, staffing, or LP reporting playbook does not need a salary, a benefits package, or a desk. It is a coworker that scales with the work.

The firm of 2030 will not have fewer professionals. It will have many more — multiplied by the expert coworker agents alongside them, each doing what a member of the business services team does, with the firm's full institutional method backing every action. The partnership grows. The overhead falls. The throughput governors lift. The shape of the firm changes.

Every firm leader has had the same conversation about the firm-wide agent — the one with its own identity, write access to critical systems, and the authority to act on behalf of the firm — and most have stopped it at the same moment. "How do I know it will respect the ethical walls, not hallucinate, not surface compensation data in the wrong place, and produce an audit trail I can defend to a court, an insurer, or a partnership vote?" Horizontal AI vendors are structurally unable to answer these questions, because the answers require an architecture the firm runs on — not a model with a wrapper around it.

Horizontal AI keeps the content of a document and strips the metadata around it. That works for marketing copy. It does not work in a firm, because in a firm the metadata is the compliance posture. Which partner is walled off because her brother works at the counterparty? Which deal team is on the buy-side versus the sell-side of a live auction, and who among them has MNPI from which public entity? Which documents are privileged, on litigation hold, or required to be destroyed after a set period? Strip the tags and you have not built an assistant. You have built a breach. Now give that breach an identity and write access to critical systems. The assistant that hallucinated an answer was a liability. The agent that acts on regulated information across a live matter is a different order of problem entirely.

There is a second breach hiding in the same approach. To make a firm's documents usable, horizontal AI copies them into a new store: a second repository of its most sensitive material, outside its retention rules, outside its access controls, outside the systems its policies were written for. A discovery request reaches it. A retention schedule does not.

Firm AI is built the other way. It reasons over the firm's own systems of record, where the data already lives under the firm's governance. When it does create something new — an analysis, a draft, a record of a decision — that artifact is stored inside the same controls, inheriting the access permissions, walls, and retention rules of what it came from. There is nothing the firm cannot see, and nothing outside its policies.

Walls for AI treats compliance as the bedrock the system runs on. Matter- and deal-based access control, independence, MNPI, retention, jurisdiction, privilege, and consent are not features that can be turned on — they are the grammar the system speaks. Every prompt is a query against the firm's policy. Every answer carries the source's controls. Every agent inherits the posture of the function it is acting for. Every action is admissibly logged. The walls are not a constraint on the agent — they are the environment the agent runs inside. Accuracy, scoped identity, real-time transparency, and an admissible audit trail are properties of that environment, not bolt-ons.

This is why a growing list of practice-AI vendors, such as Harvey, are moving toward Walls for AI as the compliance foundation under their tools. Not because anyone is forcing them, but because firms will not deploy them at scale otherwise. Vendors who treat compliance as an afterthought will ship faster this year. They will also be ripped out.