The Knowledge Firm Comes Apart

In my last post I argued that most organizations shouldn't build AI capability in-house. Software is cheap but the ability to ship on AI's clock is still scarce. Most organizations should buy their AI solutions from vendors who do it full time.

I ended with a question that I’ve been contemplating that might overturn my thesis. What if knowledge-based companies (law firms, accounting practices, associations, consultancies) simply become AI implementation shops? If building software keeps getting cheaper and easier to build, maybe the capability I told them to outsource becomes the center of what they do.

I've been arguing with myself about it since. Here's where I've landed: almost every knowledge firm becomes an AI implementer but there will be different levels to this, and in the process, most of them come apart.

A knowledge firm sells judgment delivered through hours of labor. As AI takes over more of the labor, the job that's left is directing, supervising and standing behind AI work in a specific domain. That is implementation. Every firm that survives will do some of it.

What varies is sophistication. Some firms will use off-the-shelf tools. Some will configure them around their own workflows. Some will orchestrate a mix of vendors, agents and in-house pieces. A few will work at the frontier and sell that capability to everyone else. Those frontier firms pool scarce talent across many clients, and they see which new capabilities actually work months before any single client would. That's why the argument from my last post still holds.

The bigger change is what happens to the shape of the firm.

The economist Ronald Coase asked why firms exist at all. The answer that won him the Nobel Prize: coordinating work inside a company is often cheaper than buying it on the open market. Finding suppliers, negotiating, writing contracts and monitoring performance all cost money. When those costs are high, you bring the work in-house.

Agents attack exactly those costs. They can search, negotiate, coordinate and monitor at a speed and price no procurement department can match. As each piece of work becomes easier to coordinate with from the outside, it gets cheaper to buy than to own.

We've watched an industry come apart like this before.

From the 1920s through the 1940s, the big studios did everything. They owned production, distribution and the theaters, and kept stars, directors and crews on long-term contracts. Owning the theaters guaranteed demand, so a permanent payroll made sense.

In 1948 the Supreme Court forced the studios to sell their theaters. Television pulled audiences away at the same time. Without guaranteed screens, the studios couldn't justify the payroll. Stars and crews went freelance, agents started packaging talent for each film, and the studios kept what they were best at: financing, distribution and the film libraries.

Film suited the project model unusually well. Every film is a one-off with a start and an end. A cinematographer can work on almost any production. Nearly everyone was in one city. And guilds, standard contracts and screen credits made it safe to hire people you'd never worked with.

San Francisco and New York have dense talent markets too, but neither went fully Hollywood.

For tech, they never went fully Hollywood because software is an ongoing effort. A codebase gets maintained and extended for years, and the people who built it carry knowledge that's hard to hand off. So tech kept its core teams integrated and unbundled and outsourced everything around them: the cloud, the software stack, the contractors. The startup itself became the project, with venture capitalists playing the studio.

Finance is different. It runs on liability and reputation. Lending, trading and underwriting need capital, licenses and a counterparty that will still exist in ten years. That favors big institutions. Where risk could be ring-fenced, finance went project-based. Private equity funds are time-limited vehicles, underwriting syndicates form for one deal and disband, and boutiques spin out of the big banks.

So work comes apart when four things are true: it's episodic, the skills are portable, quality is visible and liability can be contained. Hollywood met all four. Tech fails the first for its core teams. Finance fails the fourth.

Most knowledge firms today bundle all three kinds of work under one roof. A typical accounting firm runs advisory projects, a recurring monthly close and the audit or tax signature. Those have always shared offices, partners and a brand. AI gives each one different economics, and I expect them to come apart.

Project work becomes a general contractor business. Strategy projects, deal advice, transactional legal work, campaigns and research are episodic, and the skills travel. A general contractor owns the client relationship, takes the liability, integrates the work and answers for quality, while specialists and agents do the trades. The firms that do well here will keep three things inside: defining the problem (the architect's job), keeping the end customer relationship, maybe one specialty they perform themselves, and the data from every engagement. A firm that only coordinates becomes a broker, and brokers compete on price.

Ongoing capability becomes a product business. Bookkeeping, managed IT, continuous compliance, member services and AI capability itself are living systems that need constant improvement. This is the territory of my last post. Clients should buy from firms that ship, usually by subscription. Those firms keep a tight core team like a software company and rent the stack beneath them.

Regulated work consolidates. Audit, tax sign-off, legal opinions, fiduciary advice and insurance need a licensed, capitalized counterparty that clients and regulators trust. Accepting liability at scale takes a balance sheet and insurance, which favors size. A firm that checks AI output across thousands of engagements also learns its error rates better than anyone, which lets it price risk well and win more work. The work underneath still unbundles. The licensed firm signs and stands behind the result, like the lead bank in a syndicate, while specialists and agents do much of the work.

Once we start seeing that world emerge, the real metric most companies will need to track is how much work their scarce budget of human judgement can oversee.

You’ll use agents and AI supplied by vendors and some built and maintained by you, to enhance their efforts. But the scarce resource will be human judgement and relationships that AI can’t replicate. Why would you waste that on owning and maintaining commodity AI where you have no advantage?

If you are good at helping HVAC companies deeply understand their finances so that they can maximize their margins and value, why would you steal hours from that to be tenth best at marketing those services to your customers?

Partly. Don't build AI capability yourself still holds for most organizations. The border of what is “AI capability” will change over time though. You do need in house capabilities but they don’t necessarily need to be at the frontier. It’s too expensive to maintain for all but a few specialist firms.

What really changes over time is how the companies are organized. Most organizations will split into a general contractor for projects, a product company for ongoing capability and a trusted counterparty for regulated/liability work, and each runs on different economics.

If you run a knowledge business, three questions matter now. Which of those three businesses are you actually in? Where you're in more than one, are you running each on its own economics, or letting the oldest one set the rules for all of them? And how many engagements can each hour of your best people's judgment cover a year from now?

The firms with clear answers will be worth more than they are today. The rest will be bidding against each other.