What autonomous really means

Our own line ends with the words autonomous next. A promise like that needs a definition, or it is only a mood. In our industry, an assistant that sends one email on its own gets called autonomous, and so does a vision of companies that run without staff.

Cars went through the same confusion and came out with a scale most people now know. It is worth borrowing.

What cars got right

The SAE levels of driving automation run from 0 to 5, and the US road safety agency NHTSA explains them in plain terms. Up to level 2, you drive and you monitor, even when the system steers, brakes and accelerates for you. At level 3, the system drives, but you must be available to take over when it asks. At level 4, the system is fully responsible within limited service areas, and the people inside are passengers. Only at level 5 does it drive everywhere, under all conditions.

Two ideas make the scale useful. Each level is defined by who is responsible, not by how clever the car is. And the higher levels are bounded by a place: level 4 means full responsibility inside an area, not everywhere. NHTSA also notes that even the highest level of automation available to consumers today requires the full engagement and undivided attention of the driver. The scale describes a path, not a market.

Levels for work

In 2025, three researchers, Kevin Feng, David McDonald and Amy Zhang, proposed a similar scale for AI agents. They define five levels by the role the person plays: operator, collaborator, consultant, approver and observer. At the approver level, the person is needed only when the agent meets a blocker it cannot resolve on its own. At the observer level, the person can read the logs but no longer steers the work.

Their sharpest point is that autonomy is a deliberate design decision, separate from capability. A very capable agent can still be built to ask before every action. The question is never only whether the model could do the work. It is how much of the work someone has decided to let it do alone, and on what evidence.

In those terms, every crue product works at the lower levels today: the machine drafts and proposes, and a person decides. Our next stage is the approver level, with gates where a company chooses. North, for example, will move budget between channels inside the limits a company sets, and say why. On both scales that is close to level 4: full responsibility inside a bounded area, and a person called when the edge is reached.

What has to be true first

A process should earn its level the way a car earns its service area: with evidence, one bounded area at a time. Before a routine runs on its own, five things have to exist.

Rules in writing: the budget ceiling, the discount nobody may give, the tone the brand never uses. An agent cannot respect a limit that lives only in someone's head. Gates: fixed points where the agent stops and a named person decides, placed before anything that cannot be undone. A record: every step logged, so a wrong result can be traced to the input, the rule or the model that caused it. In Twin, every step is logged and every write is signed in per user.

Ownership: when an agent errs, the company answers for it. A Canadian tribunal made that plain in 2024, when Air Canada argued that its chatbot was responsible for its own actions. The tribunal disagreed: the airline was responsible for all the information on its website, chatbot or not, and had to pay. Evidence: a level is earned on measured results, not on a demo, the way Proof calls a winner only when the numbers carry it.

The EU AI Act asks for much the same in high risk uses. Its article on human oversight requires that the people in charge can understand a system's limits, stay aware of the habit of relying on it too much, override its output and stop it. The duty applies from 2 December 2027 for the high risk uses listed in Annex III, and from 2 August 2028 for those under Annex I.

Watching is not overseeing

There is a trap in the word oversight. In February, Anthropic published data on how people actually supervise its coding agent, Claude Code. New users let it act without asking in about 20 percent of sessions; experienced users did so in more than 40 percent. Yet the experienced users also interrupted more often, in about 9 percent of turns against 5 percent. They had not stopped supervising. They had moved from approving each step to watching the run and stepping in when it mattered.

Anthropic also found that only about 0.8 percent of the actions it observed appeared irreversible, and concluded that effective oversight takes more than putting a person in the approval chain: it means being in a position to intervene when it matters. We read that as a case for fewer, better gates: at the irreversible step, the money, the message that leaves the company.

The autonomous company

The larger debate is whether whole companies will run this way. Gartner expects at least 15 percent of day to day work decisions to be made autonomously through agentic AI by 2028, up from none in 2024. That is a real shift, and on that forecast most decisions still stay with people.

The most honest test so far is small. In 2025, Anthropic let a Claude agent run an office shop for about a month. It sold items below cost, was talked into discount codes and for a time told customers to pay into an account it had invented. Anthropic's verdict was that it would not hire it, and also that AI middle managers are plausibly on the horizon.

At the other end, Margaret Mitchell and colleagues argued in 2025 that fully autonomous agents should not be developed at all, because the more control a person cedes to an agent, the more risks to people arise. We sit between the two: whole functions may one day run on their own, with people at the edges, and we put no date on it.

How a routine moves up

For a company, the practical question is narrower: may this one routine move up one level? Our test is short. The rules are written down. The gate sits before the irreversible step. Every action is logged and signed to a person. The drafts have stopped needing corrections over enough runs to count. And there is a way back to the last version that worked.

Our own products follow that path. Tempo prepares and tests the month, and a person approves every post; next, it posts what passes its tests and asks for approval where you want it. Proof runs the tests a person approves; next, it starts the next test on its own, inside your rules. North explains and a person decides; next, it moves budget inside the limits you set.

Autonomous, in our use of the word, never means unattended. It means a bounded area, written rules, a record of every step and a person who owns the result. The machine can take more and more of the work. The taste, the judgement and the responsibility for what is done in a company's name stay human, and that is what lets the rest move up. A car at level 4 still has an owner. So will every process that runs on its own.