The delegation ladder: five steps from locked down to AI-integrated

First written down in 1978, for the ocean floor

In July 1978, MIT's Man-Machine Systems Laboratory delivered a report to the US Office of Naval Research on the remote control of undersea vehicles. The problem was concrete: commercial divers in the North Sea were dying at rates the report described as alarming, on long jobs decompression consumed roughly an hour for every six feet of depth, and the Navy wanted unmanned vehicles with cameras and mechanical arms to take over inspection and salvage work. That raised a question nobody had ordered systematically before: how much should the computer on board be allowed to do? Thomas Sheridan and William Verplank answered with a typewritten table running across three pages, ten levels of automation, from a human who does the whole job at the bottom to, at the top, a computer that does the whole job if it decides it should be done, and tells the human only if it decides the human should be told.

Every ladder since descends from that table: the six levels of self-driving that SAE wrote down in 2014 are its best-known grandchild, and the columns keep the same logic through every generation, who gets the options, who selects, who approves, who starts, who gets told. The scale measures not what the machine can do, but what the human dares hand over.

For an engineering firm the ladder needs five steps, and one value holds constant across all of them: the signature. A named engineer reviews and signs at every step; what grows on the way up is how much has been checked before pen meets paper. The steps themselves are easy to recognise. The hard part is the climbs between them, which is why, after every step below, the climb is spelled out in plain words.

Locked down: AI is blocked, and used anyway

On 11 March 2023, Samsung's semiconductor division lifted its internal ban on ChatGPT. Within three weeks the company logged three leaks: one engineer pasted the source code of a faulty database program into the chat window and asked for a fix, a second uploaded code written to identify defective equipment, a third uploaded the recording of an internal meeting to get minutes out of it. The first counter-measure was a cap of 1,024 bytes per prompt. By early May, generative AI was banned on company devices altogether.

The ban solved less than it seemed to. When KPMG and the University of Melbourne surveyed 48,000 employees across 47 countries in 2025, 57 percent admitted using AI at work without being open about it, and the risky use clustered exactly where the rules were hardest: 67 percent of employees at organizations that ban generative AI reported it, against 33 percent where no policy existed. Microsoft's Work Trend Index had found much the same a year earlier: 78 percent of the people using AI at work bring their own tools.

More organizations stand at this step than annual reports suggest: AI blocked outright, or allowed in principle with no procured environment, no data policy and no answer to who may use what. What that produces in practice is quiet use, project data moving through private accounts, without logging, on terms nobody has read. The bottleneck sits in the policy rather than in the technology, and the shadow use makes step 0 the most dangerous place on the whole ladder.

Assisted: you and one AI, one activity at a time

A fixed pair: you and one AI on one activity at a time. It answers, summarizes and drafts; you review every line before anything moves on, because you do not yet trust what you have not seen come into being. That is the right instinct at this step. The bottleneck is your own reading time: the work goes faster, but you dare not look away, so your attention sets the ceiling.

What the step buys is real even so: the activity that used to eat an afternoon now fits between two meetings.

Delegated: whole activities leave your desk

Now you hand over whole activities: a requirements list out of a tender package, a quantity check against the model, the first draft of a memo. Several assignments are out at once, and the role has shifted from co-writer to commissioner: you formulate the assignment, review the finished result against the sources, and send back what does not hold. The bottleneck moves accordingly, from your reading time to your review capacity, and to how far you trust the verification underneath it. The deliverable that used to take weeks comes back the same day.

Supervised autonomy: the routines run themselves, you lead

The 1978 report had a name for this step: supervisory control, a scheme in which a system capable of autonomous decision-making over short periods, in restricted conditions, is remotely monitored and intermittently reprogrammed by a person. Sheridan's operators sat on a support ship and watched the TV picture coming up the cable while the computer down on the vehicle closed the control loop itself; the acoustic links of the day were too narrow for anything else. In a firm the same arrangement looks like this: every new delivery is reviewed as it arrives, every revision is diffed against the previous one, deviations are monitored continuously, and your week is built around flags and exceptions instead of read-throughs.

The bottleneck is trust in the routine: who may start work, what may pass onward without human eyes, how exceptions escalate. Those rules have to be in place before the volume scales, and once they are, the checking work that used to wait for a free moment runs continuously in the background.

AI-integrated: you steer by intent

Whole deliverables are produced and reviewed as flows: the material in, the finished deliverable back, traceable line by line. You lead by intent, follow up by exception, and spend the engineering hours where they decide something: the assumptions and the trade-offs. One thing stays where it always was, whatever the volume. The signature is a decision taken per deliverable, by a named engineer who can see what was checked and what was assumed, and it sits at the top of the ladder for the same reason it sat at the bottom: someone answers for the result.

The 1978 table carried its own warning about the upper rungs: as more automation is introduced, benefits accrue and risks come with them. The bottleneck at the top step is exactly that governance: choosing which work goes into the flows, and holding the machine's output to the same quality bar as a colleague's. What the step buys is capacity of a different order; the commitment the firm declined last year fits in this year's plan.

Built for the climbs

Yesper is the AI civil engineer for construction and infrastructure, and it is built for precisely these climbs: verification with a source on every claim for the step from assisted to delegated, routines that run on schedule for the step to supervised autonomy, organizational memory and codified standards for the step to full integration. Where a firm stands today says little about where it can stand next year. Pick the step you are on, name the precondition for the next one, and give a named engineer the mandate and the time to build it. The rest of the ladder is climbed the same way, one trust decision at a time, with the signature where it has always been.