non-HITL is not a prediction but a starting point: for asking better questions and looking further ahead.
Today a person has to be at both ends of a delivery: a courier you can identify, and you at the door. Not to do the work, but to carry the trust. When a network can carry that trust, the person is no longer needed in the loop. Delivery is the simplest way to see how.
Everything on this site follows from one idea.
Trust is moving from people to networks. Where people are irreplaceable, their value multiplies. Where people are less reliable, machines take over.
Not the carrying. They answer five questions that every trust network has to answer. These are the five pillars of trust. Here is each one, today and tomorrow.
TodayA courier in a uniform, from a company you know.
TomorrowA robot with a verified identity, known to the network the way a card is known to a payment network.
Already hereCard networks already give AI agents an identity.
TodayYou open the door, or you don’t.
TomorrowPermission for one door, one time window, one task. It expires when the job is done.
Already here“Least privilege” for agents is now standard practice.
TodayThe courier’s experience: the address, the stairs, the fragile box.
TomorrowA robot certified for the task: stairs, doors, the fridge. Homes are built for human bodies, so humanoids fit.
Already hereHumanoids already move 100,000+ totes in a warehouse.
TodayA scribble on a screen, a photo of the doorstep.
TomorrowEvery step signed and recorded, and the proof sent to your phone.
Already here“Liability follows the logs.”
TodayThe courier company answers. You complain; they refund.
TomorrowThe network settles who pays, and insurance is priced on the record.
Still missingLiability for machine work is unsettled.
Cost is one reason a person leaves the loop. There are four others → four reasons the loop opens
Too much to take in, too dangerous, too unpredictable, too hard to verify. Each is a limit of the person, and each is already pushing people out of the loop somewhere today. Click a reason to see where.
More signals than a person can read, in less time than a person needs to decide.
275 alarmsin 11 minutes, for two refinery operators, before the 1994 Milford Haven explosion
Air traffic, control rooms, grids, trading, code, MarsWhere a mistake costs a life, the first move is to take the person out of harm’s way.
Mostfront-line casualties in the war in Ukraine are now caused by drones
The battlefield, mine clearance, mining, nuclear and disaster sitesMachines that share one plan can work centimetres apart. One person nobody can predict widens every gap.
1 driverthe network can’t predict sets the safety gap for every machine around them
Roads, ports, warehousesA person’s word, memory and face no longer prove what happened, or who is speaking.
$25mpaid out in 2024 after a video call in which every other face was a deepfake
Contracts and claims, identity, inspection and auditA person reads a handful of signals at once and needs seconds, often minutes, to decide. Control rooms now produce thousands of signals, and some failures unfold in seconds. Past a certain volume and speed, the person in the loop is either overloaded or only pretending to check.
Attention fades over a shift. Under pressure, people fixate on one signal and miss the rest.
Setting the limits the machines work within, and taking the cases no rule covers.
A person needs tens of seconds to minutes to understand a situation and decide. Markets move in microseconds, grid protection in milliseconds, an alarm flood in seconds. Where events move faster than that, the person in the loop can only watch.
Weakest human pillar: Competence, because a person decides in minutes, not milliseconds.
TodayControllers keep aircraft apart by voice and radar, under rules that force regular breaks because attention fades.
Non-HITLEurope’s rules for low-level drone airspace already hand traffic to automated services: no controller could separate thousands of drones by voice. Manned airspace follows, with machines agreeing separation between themselves.
Weakest human pillar: Competence, because a person can’t track that much at once.
TodayOperators of refineries and plants watch alarm screens. Before the 1994 Milford Haven explosion, two operators faced 275 alarms in 11 minutes.
Non-HITLThe system absorbs the flood and acts within set limits; the operator is told what matters, and why.
Weakest human pillar: Competence, because overload hides the one alarm that matters.
TodayOn 28 April 2025 the Iberian grid went from normal to blackout in seconds. Only automatic protection works at that speed.
Non-HITLA grid of solar, wind and batteries is balanced by machines every second. People set the margins and plan the system.
Weakest human pillar: Competence, because a person decides in minutes, not milliseconds.
TodayMost share trading in the large markets is done by algorithms, in microseconds. The trader on the floor left the loop years ago.
Non-HITLAlready here. People moved to setting limits, risk and circuit breakers.
Already non-HITL
TodayEngineers click “approve” on more AI work than they can check. Past a certain volume, approving is no longer reviewing: it is hoping. And the more they rely on the AI, the weaker the very skills they need to supervise it.
Non-HITLVerification replaces review: the machine proves its own work with tests and records. People look only at what is hard to reverse.
Weakest human pillar: Authority, because a person can’t judge at machine speed.
TodayA signal to Mars takes between 3 and 22 minutes each way. A rover can’t wait for a driver, so it drives itself.
Non-HITLAlready here. People set the goal; the rover plans the route.
Already non-HITL
Here the person is not the weak link; they are the most valuable thing in the loop. So they move out of it: first to a remote control, then, when the radio link itself fails, out of the loop altogether. The war in Ukraine shows both steps, and why this is also where accountability matters most.
A body that can be hurt, and a remote link that can be jammed.
The decision to use force, and the responsibility for it. The more autonomous the machine, the clearer the human answer must be.
TodayDrones now cause most front-line casualties. Ukraine made unmanned systems a separate branch of its armed forces in 2024, and robot ground vehicles carry supplies and evacuate the wounded.
Non-HITLJamming cuts the link between pilot and drone, so drones get fibre-optic cables, then autonomous guidance for the final approach. The enemy cuts the loop, not the designer.
What must stay human: Accountability, for every decision to strike.
TodayUkraine is now one of the most heavily mined countries in the world. Clearing a field by hand is slow, and a single mistake kills.
Non-HITLDrones map the field, armoured machines clear it, and people check the result from a safe distance.
TodayDriverless haul trucks have run in Australian iron-ore mines for more than a decade, controlled from operations centres far from the mine.
Non-HITLWhole sites run with nobody in the pit. The operator works from a city.
Already non-HITL
TodayDecommissioning Fukushima depends on robots sent where people can’t go. Drones and robots go first into fires, tunnels and collapsed buildings.
Non-HITLMachines inspect, cut, carry and search. People plan and decide from outside.
A network that coordinates thousands of machines can’t plan around someone it can’t predict. Not because people perform badly, but because they perform differently every time. So machines get their own space first: lanes, port zones, warehouse floors.
Varies with mood, fatigue and habit, and can’t share a plan in real time.
Mixed and open spaces, and the exceptions: the blocked road, the strange cargo, the child in the street.
Not because people drive badly. A network that coordinates thousands of machines can’t plan around someone it can’t predict. Ports, logistics corridors and night lanes go first.
Weakest human pillar: Competence, because a person is unpredictable.
TodayAutomated container terminals in Rotterdam and Shanghai already run cranes and vehicles from software.
Non-HITLWhole port zones close to people while machines work, and open only for inspection.
Weakest human pillar: Competence, because a person can’t share the plan.
TodayHumanoids already move 100,000+ totes in a warehouse, and robot floors are fenced off from people for safety.
Non-HITLNo lights, no people. Maintenance by appointment.
Weakest human pillar: Competence, because a person is unpredictable.
Trust used to rest on a person: a signature, a testimony, a face on a call. Generative AI makes all three cheap to fake, and memory was never reliable. A signed machine record becomes the stronger proof.
Forgets, rebuilds memories honestly but wrongly, and can be impersonated.
Judging disputes and intent: what the record means, not only what it says.
TodayIn 2024 a finance employee in Hong Kong paid out about $25 million after a video call in which every other participant was a deepfake.
Non-HITLMoney moves on a cryptographic identity, not a familiar face. Agents verify each other before they act.
Weakest human pillar: Identity, because faces and voices can be faked.
TodayA signature and a person’s word settle what was agreed and what happened. Courts already face deepfaked testimony.
Non-HITLA signed machine log becomes stronger evidence than a person’s word.
Weakest human pillar: Evidence, because people forget and can be faked.
TodayAuditors test samples and inspectors sign forms, both relying on what people say happened.
Non-HITLEvery step is signed as it happens. Audit becomes continuous and complete, not a sample.
Weakest human pillar: Evidence, because a person checks a sample, after the fact.
Across all four reasons, the same pattern appears → the person against the network, pillar by pillar
Start from what people are like, and compare them with a machine inside a trust network, pillar by pillar.
Hard to verify at a distance: voices and faces can be faked. A machine carries a cryptographic identity.
Can be fooled or pressured into going beyond the mandate. A sandboxed machine cannot.
Unpredictable, inconsistent, tired, slow on routine work. But in the unknown, where no rule exists yet, the person still leads.
Forgetful: memory fades and is rebuilt, honestly but wrongly. A signed log doesn’t. “Liability follows the logs.”
Here the person is the strong link. A person can answer in court and in conscience; for machines, liability is still unsettled.
Positions are our judgement, for discussion. Each pillar is built from technology, process, regulation, economics and people; see what has to move.
Which is exactly where human potential grows → where humans win
When machines carry the routine trust, the scarce thing is what only people do. The sectors built on it gain the most: each person is amplified by the machines around them.
Machines run the diagnostics and logistics; clinicians spend their time on judgement and the patient.
One researcher directs thousands of automated experiments.
AI tutors do the drilling; teachers and coaches do motivation and judgement.
Generation is free; choosing what is good becomes the value.
Agents prepare and execute; people set direction and carry the deal.
When most things are made and served by machines, “made by a person” is worth more.
What does this mean for the work people do today? → jobs today, with non-HITL at the centre
The question is not whether a job disappears, but which part moves to the network and what the person does instead. Six families of work, with non-HITL at the centre. Click one.
Air traffic controllers, grid and plant operators, rail signallers, security analysts, ship’s officers of the watch.
From watching screens → to setting the rules
Speed and informationTruck and van drivers, couriers, crane and forklift operators, train drivers, drone pilots.
From driving one vehicle → to running a fleet
PredictabilityHazardMiners, deminers, firefighters, nuclear technicians, offshore and power-line workers, soldiers.
From going in → to sending in
HazardAuditors, inspectors, claims adjusters, notaries, compliance and identity analysts.
From checking samples → to judging disputes
ProofCode reviewers, accounts payable, buyers, loan officers, schedulers, back-office teams.
From approving every item → to writing the mandate
Speed and informationProofClinicians, nurses, teachers, coaches, negotiators, leaders, craftspeople.
From the paperwork → to the person in front of you
Value multipliesDecision support now. Machine-run drone airspace is already being built. Machine-run grids and plants from around 2035.
Watch screens, answer alarms and decide in seconds, in shifts limited by fatigue.
The network resolves routine conflicts and alarms itself, and explains what it did.
Exception manager: sets the margins, takes the cases no rule covers, and investigates what went wrong.
Mines and ports now. Machine-only corridors around 2035. Inside the home around 2040.
One person, one vehicle, one shift.
Vehicles drive themselves in machine-only corridors and depots, and one remote operator supervises many.
Fleet controller, then exception handler: the person called when a vehicle meets something it has never seen.
Already in mining and on the battlefield. Demining, energy and emergency services follow over the next decade.
The person goes in, and carries the risk.
The machine goes in first: it maps, cuts, clears and carries. The person directs from a safe distance.
Mission commander: plans, decides and answers for what the machines do.
Around 2035, when a signed machine record counts as evidence.
Check a sample, after the fact, based on what people say happened.
Every machine action is signed as it happens, and checks run on everything, all the time.
Judge of the exception: what a record means, who was at fault, what is fair.
Around 2030, as standing mandates let agents act for people and firms.
Approve item by item, often without the time to check.
Agents act within a standing mandate: what, how much, for how long. Only what is hard to reverse comes back to a person.
Mandate designer: decides what an agent may do, and audits how it used that power.
Gradually, from now.
A large part of the day goes to records, scheduling and logistics.
Machines carry the records, the scheduling and the logistics, with proof. The person carries the relationship.
The same job, multiplied: more time for judgement, and for the person in front of you. See where humans win.
Between the weak link and the strong link sit new markets → nine markets that open
None of these is waiting for a better app. Each is waiting for trust to be carried without a human. The year is our estimate of when that happens.
When does each market open, and what has to happen first? → the path to 2050
Our scenario, in five steps. Click a step to see what changes and which markets open.
Why does it take this long? Technology is only one of five layers under each pillar → what has to move
Machines that are identifiable, limited, reliable and recorded.
Ways to check machines without slowing them down.
Rules that recognise a machine’s mandate, record and liability.
Someone paid to run the trust, and someone to insure it.
People willing to delegate, and to let machines in.
Identity, authority, competence, evidence and accountability each need all five layers before a market can open. See the pillar-by-layer analysis →
Agentic economies, the new labour equation and polycompute. “Liability follows the logs.”
ftsg.comThey named human-in-the-loop theater: an approval step that works as an accountability sink.
thoughtworks.com/radarFutures drawn as paths with milestones over decades: the model for our path to 2050.
shell.com
Transformation and programme director. More than 28 years leading transformation in telecoms, energy, logistics and technology across Europe, South America and Africa.
hugomgomes.com
Responsible AI and AI Governance Lead. Driving AI strategy in regulated industries, bridging technology, regulation and business, and making it accessible to everyone.