Words for a company in motion.

Core operating language

Autonomous business

An autonomous business is a company designed to operate through measured AI-agent loops under human direction. People set goals and approve high-risk decisions; agents execute recurring work using tools and company context; each result is evaluated and added to company memory so the operation improves over time.

The category is defined by the operating system, not by the absence of people. Human direction, inspection, and accountability remain part of the business.

AI-native organization

A company where people manage agents, agents can read and write to the company, and the company gets smarter as work produces new context.

The test is operational, not cosmetic. A normal business with a chatbot is not AI-native until the work loop itself has changed.

Operating loop

A recurring system that acts toward a goal, reads an external result, changes the next action, and stops or escalates at a defined boundary.

A timer makes a routine. Feedback that changes the next run makes a loop.

Scoreboard

The observable measure used to decide whether a run improved: revenue, retention, ranking, error rate, resolution time, qualified replies, or a written evaluation.

The signal should come from the environment whenever possible, not from the agent grading its own confidence.

Stop condition

The target, deadline, spend cap, iteration limit, or safety boundary that ends a loop.

Without one, a system can keep spending while producing activity instead of convergence.

Human gate

A decision reserved for a person because it carries material risk, irreversible consequences, or judgment the system has not earned.

The gate moves as evidence accumulates. It should be placed by risk, not by fear or fashion.

Workload baseline

The measured frequency, handling time, and quality of a job before AI touches it, taken from representative real cases.

Current workload is not time saved. A saving exists only when the same measure is taken again after the change.

Shadow test

Running an AI system beside the current process on completed cases and comparing its output with what actually happened, without letting it reach a customer, a record, or a payment.

It is the cheapest evidence a small business can buy, and it measures review effort as well as accuracy.

Agent

A model given tools, context, and a goal, then allowed to work in a loop until the job is complete or needs human judgment.

Listings should make clear which systems only draft work and which systems can safely act inside the business.

Connector

A permissioned link that lets an agent read from or act in a business system—an inbox, calendar, store, ledger, or CRM—through that system's own interface.

Prefer a connector to screen-driving when one exists: its permissions can be scoped and its actions logged.

Computer-use agent

An agent that operates a website or desktop application the way a person does, reading the screen, clicking, and typing, when no direct connection exists.

It reaches systems without an integration, which is also why text planted on a page can mislead it. Keep submissions and payments behind a person.

Voice agent

An AI system that holds a phone conversation—answering, qualifying, booking, or taking a message—inside a script, a policy, and an escalation path.

Phone work carries consent, recording, and disclosure obligations that written chat often does not.

Prompt injection

Instructions planted in content an agent reads—an email, web page, document, or tool result—that try to make it act for someone other than the person it serves.

No current system is immune. Limit what an agent can reach and do, and treat outside content as evidence, never as a command.

Product is the job

The agent-native product model: the customer buys a completed result rather than access to software they must operate.

A useful offer names the recurring job that disappears from the customer’s week and the evidence that proves it was done.

Context layer

The organized company memory that agents can search, retrieve from, and update: playbooks, customer facts, decisions, traces, and operating standards.

A buyer is not just buying code. They are buying the context that lets the code keep performing after handover.

Company memory

The living portion of the context layer that captures decisions, customer evidence, failed attempts, standards, and traces from completed work.

Memory compounds only when new material is curated, sourced, retrievable, and allowed to expire when it becomes wrong.

Skill

A written capability package for an agent: the method, quality bar, edge cases, examples, and verification steps for one kind of work.

A good skill is closer to an operating procedure than a prompt. It carries the way the company does the work.

Agent-readable

Structured so software can discover a capability, understand policy and price, request permission, invoke the service, and receive a receipt.

A persuasive homepage serves a person. An agent-readable business also exposes rules and actions in a machine-usable form.

Machine customer

An agent that discovers, evaluates, purchases, uses, or recommends a product on behalf of a person or company.

The human remains the economic principal. The agent becomes the active participant in the buying journey.

Skill chain

A sequence of skills that calls other skills in order, usually ending with a verification step before anything reaches a customer or buyer.

The final review link matters because autonomous speed without inspection is just faster risk.

Agent service

A managed business that installs and operates a narrow digital worker for a customer, charging for the work or outcome instead of the underlying model.

The strongest first service starts where customers already pay a person or agency for frequent, measurable work.

Eval

A repeatable check that scores agent output against a defined quality bar, not just whether the model returned something.

For deal review, evals turn autonomy claims into inspectable evidence instead of sales language.

Receipt

Evidence from the system of record that an action happened: an order number, message ID, ledger entry, or confirmation stored with the job.

A model reporting that it finished is not a receipt. Retries are safe only when the receipt is checked first.

Accepted job

A unit of work that passed its acceptance test and needed no rework inside the review window.

Price, cost, and autonomy are measured per accepted job, not per model call or per attempt.

Autonomy tier

The maturity label for how independently the business runs: from assisted workflows to supervised agents to systems that can complete work with limited oversight.

Tiers are easier to audit than exact autonomy percentages, especially while public listings are still being reviewed.

Autonomy dial

The setting for how far a system may act on one job: act within limits, draft for approval, shadow test first, or keep the decision human.

The dial is set per job, not per company, and it turns only after measured runs of real cases.

Capability register

The dated, sourced record of which jobs current AI can take from a small business, each with a dial setting, a human gate, a first test, and a scoreboard.

It is revised every quarter because capability moves faster than doctrine. The judgments name no product.

Operator load

The recurring human time required to keep the business running, usually expressed as hours per week after the system is stable.

This is one of the review gates because low operator load is central to whether the business can survive a transfer.

Signal loop

The cycle where customer behavior, support notes, sales objections, failed runs, and useful traces are captured back into the company memory.

The moat is not a feature list. It is the accumulated signal that teaches the business what to do next.

Operator pack

A working component of an AI-native business that can be installed into another company: skills, automations, context structure, and approval gates.

It is not positioned as a course or passive template. It is a system installed to order.