AI is the Operating System

Every firm has an AI story. A copilot in the document management system. A pilot with one of the research tools. A working group, a policy, a slide somewhere that says AI-first.

Underneath, though, most of them are doing the same thing: taking the way they already work and bolting AI onto the side of it.

A recent talk by Diana Hu, a partner at Y Combinator, named the thing that bothers me about that. Her argument is short. AI should not be a tool your company just uses. It should be the operating system the company runs on.

Using AI vs running on AI

If AI is a tool, you add a copilot to your existing workflow and carry on. D

ocuments get drafted a little faster. Research comes back a little quicker. The shape of the firm stays exactly as it was.

If AI is the operating system, almost everything is open to question. How the firm is structured. How it hires. How it measures good work. What it spends money on. You stop asking where AI can help with a given task, and start asking how you'd build the firm if AI were the default way work moved through it.

Those are very different questions. One makes you a bit more efficient. The other makes you hard to compete with.

The open-loop firm

Hu draws a distinction between open-loop and closed-loop systems.

Most firms are open-loop. Work gets done and decisions get made, but very little of it is captured in a way the organisation can learn from. Knowledge lives in inboxes, in matter files no one reopens, in the heads of the partners who've seen it all before. Status moves upward through people - slow, lossy, and dependent on someone finding the time to write the update.

A closed-loop firm is one where the work is legible. Every matter, every decision, every piece of advice leaves behind structured information the firm can read, query and improve from. Hu calls this being queryable. Can you ask your own organisation a question and get an answer - without booking three meetings to find it?

Most firms can't. And until they can, the AI they install has nothing good to feed on.

Open loop versus closed loop - the difference between work that vanishes and work the firm can learn from.

When the org chart stops earning its keep

This is the uncomfortable bit for legal.

A lot of management exists to move information around for the benefit of humans. Associates summarise for senior associates, who report to partners, who brief the committee. Layers of routing and translation, all because no one person can hold the whole picture at once.

When an intelligence layer can hold that picture, some of those layers no longer justify their place. Not the people - the routing. Humans move to the edge of the work, where the judgement actually sits: the client relationship, the risk call, the strategic choice, the decision that has to carry a named person's name. The middle, the part that was mostly about passing information along, gets thinner.

That cuts against the grain of a profession built on leverage pyramids. I don't think that makes it wrong.

The software factory, applied to legal

Hu describes a way of building she calls the software factory. Humans define what success looks like - the specification, the tests it has to pass - and agents do the building against that definition.

Put that into a legal frame and it's familiar. The lawyer's job becomes setting the standard: the house position on a clause, the risk tolerance, the precedent that counts as good, the test a piece of work has to clear before it goes near a client. The drafting, the first-pass review, the assembly - that's the factory floor.

Hu says she's watched teams halve their timelines this way while producing several times more. If you've ever looked at how much of legal delivery is high-volume, pattern-heavy work, you'll see where this goes.

Spend, not headcount

The instinct in a growing firm is to hire. More work, more people. Hu's provocation is the opposite: spend on compute first. Treat your token bill as the thing you scale, with a small, heavily leveraged team around it, rather than reaching for headcount by default.

Charlie Warren, another YC partner, puts a useful test on the same idea. If your revenue only climbs when your headcount climbs, you aren't really running on AI - you've just bought faster tools. The point is for the people to grow slowly while the work grows fast.

We should be realistic here. They're describing venture-backed startups, and a law firm isn't one. The maths only works where there's revenue or funding to cover the spend, and legal carries obligations a startup doesn't - confidentiality, privilege, data protection, duties that don't bend because the workflow got quicker.

A closed loop only works if the loop is governed. That isn't a reason to ignore the shift. It's a reason to build for it on purpose, rather than back into it.

Variance is what kills you

Warren's sharper warning is about consistency, and it's the one that should worry legal most. Clients will forgive you for being a bit slow. They'll forgive you for being a bit expensive. They will not forgive you for being unpredictable.

That's the catch with putting AI at the centre of the work. The upside is speed and volume. The risk is variance - the same kind of matter handled well one week and badly the next. In most services, that quietly erodes trust. In legal it can do real damage, because the price of an inconsistent answer isn't a refund. It's a missed limitation date, or a clause that doesn't hold when it's tested.

So the closed loop isn't only about leverage. It's how you keep the output consistent enough to be trusted. Set the standard, measure against it, and catch the misses before the client does. Speed is easy to sell. Consistency is what keeps the client.

Clients forgive slow and expensive. They don't forgive unpredictable.

Why this favours the firm being built today

So much of this comes down to timing.

A firm built around an intelligence layer from day one is a different animal from a traditional firm that has added AI at the edges. The first has its data, its workflows and its culture arranged so the AI can do its job. The second is forever retrofitting - trying to make decades of open-loop habit legible after the fact. Warren's version of the warning is blunter: don't try to buy your way in. You can't take an old services business, layer AI over the top, and expect the economics to follow. The habits are already set.

There's a second shift hiding in that. The most interesting new entrants aren't selling software to law firms. They're delivering the legal outcome directly - the firm itself rebuilt around the work, rather than a clever tool sold to someone else to use. Some of the largest legal businesses of the next decade may not be software companies at all.

We're already seeing the early version. The SRA approving its first AI-driven law firm. New entrants building AI-first from the ground up - we wrote about Crosby doing precisely that. These aren't traditional firms with a chatbot stapled on. They're much closer to the picture both partners describe: the firm as an operating system, with lawyers at the edge guiding it instead of routing everything through themselves.

What actually sets you apart

The easy conclusion is that the firm with the best AI wins. But when everyone has access to the same underlying technology, the technology stops being the thing that sets you apart. Reliability included, since the consistent output comes from the standard you set and the judgement around it far more than from the model itself.

What's left is judgement. Taste. Knowing which problems are worth solving and which aren't. The client experience around the work. Which loops you choose to close, and which you leave open on purpose because a person should be making that call.

AI as your operating system doesn't take the lawyer out of the firm. It moves them to where they were always worth the most. The harder question is what it does to everyone in between.

Narrow before you scale

There's a precondition to all of this that's easy to skip. Repeatable, consistent delivery - the thing the operating system exists to produce - only works if the work itself is narrow enough to repeat. Most firms are built to do the opposite.

A traditional firm is a generalist by design: many problems, many kinds of client, a different answer every time. That's a strength in the old model and a liability in this one. You can't build a system that does everything for everyone and also does it the same way twice. So becoming AI-native starts with the decision most firms avoid - choosing what not to do.

It's a narrow choice. A single service. A specific problem worth solving, in a market that's growing, where you can actually point to an advantage. A client profile defined sharply enough that you know exactly who you're building for. Get that right and everything after it gets easier: how you package the offering, how you price it, how you deliver it, and how you talk about it.

Scaling an AI-native company requires thinking like a technology company. The way you market, the way you sell, and the people you need to deliver a return on the investment you make into the product that is your new service.