The Next $1T Company

On AI services, insurance claims, and who gets to define completed work.

Sequoia is right, and maybe so right that the argument almost becomes boring. Of course companies would rather buy completed work than better tools for doing work. Nobody wanted accounting software because they had a deep spiritual attachment to ledgers. They wanted the books closed.

The more interesting question begins after that point. When software stops selling tools and starts selling completed work, the company may not only produce the output. It may set the rubric, optimize against its own version of completion, and be treated by the buyer as if completion equals judgment.

Labor is being transferred. Intelligence, increasingly. But maybe also something stranger: the authority to define what done means.

In Services: The New Software, Julien Bek argues that the next major AI company may be “a software company masquerading as a services firm.” The market moves from copilots to autopilots: from software that helps a professional do the work to software-native services that sell the completed work directly.¹

This is a very good venture thesis. It is also, once stated plainly, almost obvious. Software has always wanted to disappear into the work. Most businesses do not wake up with a craving for more dashboards, more workflows, more interfaces, or more SaaS permissions to administer. They want the claim processed, the contract reviewed, the ticket resolved, the invoice collected, the candidate screened. They want the thing on the other side of the tool.

For a long time, software could not quite sell that. It could sell leverage. It could sell speed. It could sell organization. But someone still had to stand between the system and the world and say, “yes, this is right.” A tool could make the accountant faster, but the accountant still closed the books. A tool could help the lawyer search precedent, but the lawyer still carried the judgment of what the client should sign.

AI changes that bargain. If the system can produce the work, the tool starts to look like an unnecessary middle layer. Why buy the shovel if someone will sell you the hole?

But professional services were never only markets for labor. A client pays a lawyer not only because the lawyer can generate clauses, but because the lawyer works inside a professional world where clauses mean something. A company pays an auditor not only because the auditor can reconcile entries, but because the auditor can make the reconciliation credible to people who were not in the room.

Some of that can be automated. Some of it probably should be. A lot of professional work is more mechanical than professionals like to admit: forms, rules, comparisons, routine production.

Still, there is a difference between work that is correct and work that has been judged.

Correctness belongs to a system of rules. The books reconcile. The code passes tests. The contract includes the right clauses. The claim matches the policy language. The answer is legible to the institution that receives it.

Judgment begins when correctness is not enough. It asks what matters, what has been missed, what kind of situation this really is, and who is prepared to be answerable for the result. It is not merely the selection of an output. It is a stance toward reality.

That is where the language of truth starts to matter. Lowercase truth is already hard enough: did the system say something accurate, apply the rule properly, and produce a result that corresponds to the available facts? AI will keep getting better at this. In bounded environments, with clear verification rules and enough examples, it may become extraordinary.

Truth with a capital T is more annoying. It does not fit as cleanly inside a workflow. It asks whether the work remains connected to the thing it claims to serve, or whether the institution is only producing outputs that satisfy its own procedures. It asks whether the person or organization receiving the work can honestly say, “we understand what has happened here.”

That distinction may sound abstract until the autopilot starts making decisions in places where the output is not just an output.

Insurance is not a random example. The National Association of Insurance Commissioners already describes AI as being used in insurance claims handling and claims processing, including damage assessment from photos and historical data, while emphasizing that insurers remain responsible for compliance, consumer protection, and human oversight.²

So take a claim. A customer files after a basement floods. The AI service ingests photos, policy language, contractor notes, prior claims, and internal rules. It classifies the loss, checks coverage, identifies exclusions, flags suspicious details, produces the denial letter, and routes the file as closed. The service may be fast, consistent, and mostly correct. It may reduce obvious human delay. It may even be fairer than a tired adjuster trying to clear a queue before dinner.

Now imagine the denial is procedurally defensible. The right clause applies. The photos match the pattern. The system has seen enough similar claims to know that this one usually falls outside coverage. The customer receives a letter that is clear enough, polite enough, and legally reviewed enough.

Has the claim been judged?

Maybe. But notice how quickly the question changes. At first, we ask whether the system reached the right answer. Then we ask whether anyone understood what the answer meant. Was this really a routine exclusion, or was there a fact that only became legible when someone listened to the customer explain the sequence of events? Did the company honor the actual promise the customer thought they had purchased, or did it merely execute the promise that survived the fine print?

This is not a sentimental argument that every claim deserves a priestly human encounter. Many claims are boring. Many denials are right. Some customers are wrong, confused, or opportunistic. But the example reveals the structure of the problem. A denied claim is not only a classification event. It is a decision inside a relationship of trust, risk, vulnerability, and institutional power. Someone has to be able to say not only “the model applied the policy,” but “the company has judged what it owes.”

That difference matters because AI services will be especially tempting in domains where the buyer is not the person most exposed to the output. The insurer buys the claims autopilot. The claimant lives with the decision. The company receives efficiency. The customer receives reality as translated by a system they did not choose.

This is where Geist can enter without turning the essay into a Hegel seminar. For Hegel, Geist is often translated as spirit or mind, but the useful point for this argument is that mind is not sealed inside an individual. Cognition and judgment are formed through social, cultural, historical, and institutional life. The Stanford Encyclopedia of Philosophy connects Hegel’s objective spirit to the social and political forms through which human freedom and ethical life become actual.³ In plainer language: judgment does not live only inside an individual skull. It lives in practices, professions, norms, institutions, histories, and forms of recognition.

That creates a strange parallel to LLMs. They are trained on historical traces of human language, practice, and judgment, then released into current institutions where they may be asked to judge new cases. But a person keeps being changed by experience, feedback, contestation, and the movement of the world around them. A model’s training context is more static. Unless the institution around it creates real feedback, review, appeal, and revision, the system can apply an inherited picture of judgment to a situation whose standards still need to change.

That is the useful bridge back to AI services: a model can learn the shape of professional output before it can inherit, or be made answerable to, the role that made the output count.

But an institution is not only a dataset of prior moves. It is a way of standing inside a shared world and accepting a role within it. The insurer does not only classify losses. It sells policies that promise payment when a covered loss occurs. The adjuster does not only map facts to policy language. At their best, they help the institution understand whether that promise now applies.

This does not mean AI cannot participate in professional work. It means we should be precise about what kind of participation we are describing. There is a difference between automating the intelligence required to produce an answer and inheriting the authority to say that the answer has been judged.

Sequoia’s copilot-to-autopilot frame depends on an important distinction: intelligence versus judgment. Intelligence is the part of the work that can be learned, repeated, scaled, and verified. Judgment is the harder layer built from taste, experience, context, and responsibility. Sequoia’s wager is that today’s judgment will become tomorrow’s intelligence as AI systems gather data about what good decisions look like in a particular field.

A serious reply is that judgment does not have to stay mysterious. You can build more of it into the service. Require review for uncertain cases. Give customers a real appeal path. Audit the system. Track where the model gets things wrong. Make a person or team responsible for the final decision.

Good. That is exactly the work. But notice what has happened: the judgment has not simply moved into the model. It has moved into the design of the institution around the model. The hard question is still who can recognize the exception, change the rule, and answer for the company when the system produces the wrong kind of right answer.

And Sequoia will still be right in many cases. A lot of what we call judgment is probably pattern recognition wearing a nicer suit. Professionals routinely convert experience into intuition, and intuition into decisions that can later be written down. If enough examples exist, and if it is clear which decisions were good, AI will absorb more of that territory.

But the phrase “what good decisions look like” hides the central problem. Good according to whom? Good inside which institution? Good for the buyer, the client, the regulator, the shareholder, the person whose life is being sorted by the system?

Judgment is not only the ability to select the move that has historically been rewarded. It is the ability to ask whether the system is still rewarding the right thing, and whether the institution can live with what it has taught the system to prefer.

Once AI services become cheap and competent, organizations will have every incentive to treat judged work and acceptable output as the same thing. The output will arrive. The SLA will be met. The dashboard will be green. The claim will be closed. The customer will have bought the work instead of the tool. The system will look wonderfully efficient.

And maybe it will be. I do not want to pretend that every human checkpoint is a sacred encounter with Truth. Many are theater. Many are expensive rituals for producing institutional comfort. A human adjuster can also hide behind policy language. A human reviewer can also rubber-stamp the answer that keeps the queue moving. If an AI-native service can remove fake judgment from a process, good. There is no need to preserve human ceremony just because it has a professional title attached to it.

The harder question is how we tell the difference between fake judgment and real judgment before we automate both.

My hypothesis is that the decisive question for AI services will not be whether they can complete work. They will. It will not even be whether they can complete work accurately. In many domains, they will do that too. The deeper question is whether the service has a truthful relationship to the world its work affects.

That relationship may require human ownership, clearer liability, audit trails, professional norms, and real review. In claims work, that might mean the system can draft, classify, and recommend, but the institution must still create a real path for contestation, explanation, exception, and accountability. The point is not to bolt a person onto the end of the workflow as a ritual. The point is to design a system in which the institution can still be addressed.

This is less convenient than the autopilot story, because responsibility does not scale as elegantly as intelligence. It does not become cheap in the same way. You can automate a document, a classification, a recommendation, a workflow. But the authority to say “this has been judged” has to come from somewhere. If it does not come from a person, a profession, or an institution, then it may come from the market’s quiet shrug: the work was accepted, so the work was true enough.

That is not nothing. Much of civilization runs on “true enough.” But it is not Truth.

So yes, services may become the new software. The business logic is strong. The budgets are enormous. The customer always wanted the work done. The tool was just the awkward phase.

But if software starts selling completed work, we should be careful about what else gets bundled into the transaction. Labor may move. Intelligence may move. Revenue may move from SaaS budgets into services budgets.

Judgment is the thing I am less sure can move so cleanly.

Not because humans are magical. Not because professionals deserve permanent protection from automation. Not because every human decision is wise, ethical, or deep.

The reason is drier: judgment has to connect intelligence, responsibility, and the shared world where the work has to mean something.

The future of AI services will not only be about who does the work. It will be about who is still allowed, and still obligated, to mean it.