The curve is a line item now

A capability forecast is now a financial input.

Scott Alexander’s Astral Codex Ten essay on sigmoids is useful because it strips out one lazy form of comfort. Every exponential growth curve eventually flattens. That truth does not identify the point where it flattens, or whether the bend arrives before the economics of a small business have already changed.

For an owner, the sharper question is this: if AI capability keeps compounding for another few years, which cost lines, role designs, and margins stop making sense?

Alexander is arguing against a common forecasting dodge. Someone sees a rising AI capability curve, points toward a future level that would matter, and another person says the curve will become a sigmoid before then. That answer can be true in the abstract and still empty as a forecast. Curves flatten because something makes them flatten: compute supply, data, algorithms, economics, regulation, product limits, user trust, physical infrastructure, or another bottleneck. Absent a mechanism, the flattening claim is a hope wearing graph paper.

His examples matter because serious forecasters have made this mistake before. Fertility forecasts expected falling birth rates to level off too early. Solar deployment forecasts kept drawing flattening paths while actual deployment continued upward. One AI capability curve that looked ready for a bend was overshot by the next model release.

Curves do bend. The bend has to be earned.

In ignorance, Alexander argues for a Lindy-style prior. A trend that has already continued for years should be expected to continue for a comparable span unless a concrete model says otherwise. Skeptics who expect near-term flattening need to name the force that stops the trend, when it binds, and why it beats continued spending, better chips, algorithmic progress, synthetic data, tool use, and market demand.

Hadto’s audience should read that as business pressure, not futurist theater.

Small companies live inside thin financial models. A few labor hours per job can decide whether a line of work is viable. Two administrative touches per customer can decide whether a vertical can scale past the owner. One compliance review loop can decide whether a service stays local and manual or becomes a repeatable lane.

Continued capability growth makes some business cases that looked too thin worth another pass. Bad ideas do not become good by default. Customer acquisition still costs money. Physical work still takes time. Licensed work still has rules. Trust still has to be earned. Yet a model can draft, classify, compare, research, reconcile, route, and monitor without replacing the person who owns the promise.

The real change is that intellectual labor becomes variable in places owners used to treat as fixed.

Consider home services. Trucks and technicians remain necessary, but fewer owner hours may be needed to prepare estimates, inspect job packets, chase closeout evidence, write customer follow-up, review callbacks, and watch discount leakage. In dental support, licensed clinical judgment and payer compliance still matter, while evidence assembly, plan-rule comparison, attachment checks, and review packets can absorb fewer manual touches. Local B2B services may still depend on relationship selling, yet back-office work around account research, follow-up, proposal drafting, and promise tracking can become less labor-bound.

Margin is the point.

An owner who says “AI will plateau” may be right. The owner still needs to say where. A compute limit has to explain why competitors cannot buy more compute through ordinary software. Data-limit arguments have to survive workflows that create better task traces and review examples. Regulatory limits need to separate the regulated decision from document assembly, evidence checking, scheduling, routing, and customer communication. Trust limits have to name which decisions require human authority and which decisions only require visible evidence and a review gate.

Hand-waving about sigmoids does not answer any of that.

The practical move is to reprice the business around bottlenecks instead of job titles. Start where margin leaks now: owner approvals, rework, callbacks, claims rejected for missing proof, estimates waiting on private memory, intake that loses the customer promise, handoffs that rely on one senior person, and managers who rescue the same exception every week. Then ask what happens when the clerical, comparison, drafting, classification, and monitoring parts of those loops get two orders of magnitude cheaper or better.

Some bottlenecks will remain stubborn. Physical execution remains physical. Compliance authority remains authority. Customers still decide whether the promise was kept. Owners still choose what the business refuses to do.

Others move.

The scheduler’s scarce contribution may shift from typing notes to maintaining the dispatch rule. Estimators may move from writing every proposal to defining the evidence standard that makes proposals trustworthy. Office leads may spend less time chasing files and more time owning the exception queue. Technicians may turn private know-how into playbooks that train the next operator.

Those shifts change the price of a business. Lower coordination costs can make a previously uneconomic service line financeable. Fragmented verticals become more legible when workflows are encoded, inspected, and taught. Work that depended on the founder’s inbox can become transferable once decisions, evidence, and exceptions are visible enough for another operator to run.

Capability curves belong in SMB underwriting. They affect gross margin, management load, working capital timing, training cost, role count, review burden, and the price a buyer should pay for a business that still depends on one person’s memory.

Employees need the same model. Two more orders of magnitude of AI capability would test which parts of a job remain durable. “I do the task” is getting weaker as a defense. A stronger position is “I own the standard behind the task.”

Durable workers know the source material, the customer promise, the failure modes, the compliance boundary, the exceptions, and the tradeoffs. They can supervise agents, correct the playbook, inspect the evidence, and decide when the system should stop. Generic upskilling language misses the point because the issue is bargaining position.

People who only clear routine intellectual work are exposed when routine intellectual work gets cheaper. People who can convert practice into rules have a path upward. They become the person who owns the workflow, not only the person assigned to it.

For thin-margin businesses, that path matters. The next owner-operator may already be inside the company: the scheduler who knows which jobs become callbacks, the estimator who knows which photos decide scope, the claims specialist who knows which packet survives review, or the technician who knows when the written process is wrong.

The capability curve makes that conversion more urgent. It also makes it more plausible.

Hadto treats capability curves as business inputs because the way a company runs has to change before the P&L explains the change too late.

An ontology that only describes current work is backward-looking. It may say who performs a task today, which system stores a note, or which role approves an exception. Current-state description only gets Hadto halfway. The ontology also has to expose why the task is expensive, which evidence is required, where compliance binds, where owner attention enters, which steps are clerical, which steps require judgment, and which steps create training surface for another operator.

Ontology-as-ops makes the business inspectable. A changed labor line should reveal which bottleneck moved. Agent-prepared packets still need a named decision owner. Cheaper workflows should show whether margin improved or whether the gain was lost to rework, discounts, customer confusion, or a new review burden. Captured domain rules should show how that rule trains the next person and changes the owner’s dependence on private memory.

Alexander’s argument lands at Hadto as a rule for owners.

Do not assume the curve saves you by flattening before it reaches your business. Do not assume the curve saves you by replacing everyone either. Treat continued capability growth as a live underwriting assumption. Name the bottlenecks. Price the labor. Preserve the human authority that decides outcomes. Convert domain experts into operators who own the dispatch rule, the review standard, and the evidence packet.

The sigmoid may arrive soon. It may arrive later. Owners do not get paid for guessing the shape in the abstract.

They need to know which line item changes while the curve keeps going.

Source evidence used in this note: Scott Alexander, Astral Codex Ten, The Sigmoids Won’t Save You, published 2026-05-15 and reviewed 2026-05-17.