Finance, transport and other sectors

18.1Finance, banking and insurance

In context

Agentic finance and trading by autonomous agents

Algorithmic trading has been around for a long time, but 2026 saw a further step emerge: autonomous agents (Chapter 6) connected directly to platform interfaces so that they analyze the markets, decide and execute orders on their own, with no human click. The OpenClaw agent, already encountered, is the most widely publicized example: using community skills, you link it to an exchange or a broker (stocks, cryptocurrencies, prediction markets), give it an instruction in natural language, and it monitors prices continuously to act at the right moment. The promise is appealing, but the risks are just as great, and the course offers no investment advice. First, no guarantee of gain: entrusting money to an agent creates no market edge, and on some platforms a large majority of participants lose. Next, operational risk: a model that hallucinates or misreads a piece of data amid volatility can place an unwanted order and rack up losses in a matter of seconds. Then security risk (Chapter 20): an agent connected to API keys and exposed to prompt injection becomes a prime target, and several skill marketplaces have seen malicious modules circulate, designed to steal those keys and drain accounts. Finally, regulatory risk: autonomous high-frequency execution can verge on market manipulation. This is why the sensible practice that is emerging is that of an AI that assists rather than acts autonomously: the agent researches, monitors and proposes, but the human validates and keeps a hand on consequential actions.

18.2Transport and autonomous mobility

Diagram18.1. The levels of automation (SAE scale). In 2026, driverless taxis (level 4) operate commercially in several cities, within defined areas. Level 5 (driving everywhere, in all weather) remains a distant horizon: it is the "long tail" of rare and unforeseen situations that still resists.

In context

The state of autonomous driving in 2026

Behind the SAE scale, the real landscape has come into focus, organized around two opposing bets. Waymo (an Alphabet subsidiary) is clearly in the lead: by mid-2026, its driverless taxis (level 4) handle around 500,000 paid rides per week across about a dozen American cities, and the service is expanding internationally (Tokyo, London). Its recipe: a costly combination of lidars, radars and cameras, high-definition maps of every street, and a cautious, city-by-city rollout. Tesla is making the opposite bet: cameras alone and massive learning from its fleet's data, in the hope of an autonomy that generalizes everywhere (and, eventually, of turning private cars into taxis); its robotaxi service, launched in Austin in 2025, remains to date far more modest in size and coverage, with most trips still keeping a safety operator. Others are entering the race (Zoox, an Amazon subsidiary; Uber, which is integrating autonomous vehicles; Chinese players such as Baidu, Pony.ai or WeRide). The provisional verdict: the feasibility of level 4 in a bounded urban area is proven, but scaling up runs into the long tail (the rare situations), local regulation, and public trust, which the slightest accident calls back into question.

In context

The car and the "trolley problem"

Autonomous driving has revived an old problem of moral philosophy, the trolley problem: faced with an unavoidable accident, should a car be programmed to, say, sacrifice its passenger rather than several pedestrians? These scenarios have fueled countless debates (and a vast global survey, MIT's Moral Machine, revealing that intuitions vary across cultures). But many engineers consider the question partly ill-posed: in reality, a vehicle almost never "chooses" a victim; it seeks above all to brake and avoid; and it generally does not have the information (who is where, how many people) that these theoretical dilemmas assume. The real ethical debate is more prosaic, but more pressing: what overall level of risk is acceptable for authorizing these vehicles, how to compare their safety with that of human drivers, and who is liable in the event of an accident (Chapter 21)? The trolley problem is an excellent thought experiment, but it must not obscure these concrete questions, which will truly decide deployment.

18.3Agriculture, industry and commerce

In context

Agentic commerce

An emerging trend extends the logic of agents (Chapter 6) all the way to purchasing: the agent no longer merely recommends a product, it builds the cart, pays and completes the transaction for you. Since late 2025, interfaces such as ChatGPT's "Instant Checkout" (via the ACP protocol co-designed by OpenAI and Stripe) allow purchases within a conversation; Google (UCP protocol), Visa ("Intelligent Commerce") and Mastercard ("Agent Pay") are building competing rails. The mechanism rests on tokenization: rather than handing your card number to the agent, you delegate to it a token with restricted use, capped, limited to one merchant, with thresholds requiring human validation (consistent with the consequential-actions rule, Chapter 6). The promises are large (some forecasts cite hundreds of billions of dollars in "agentic" purchases by 2030), but three obstacles remain. Fraud first: a transaction carried out by an agent lacks the behavioral signals on which anti-fraud systems rely (Chapter 20). Disintermediation next: if you buy from your assistant, merchants and banks risk losing the direct relationship with the customer. Reliability finally: the first consumer tests have sometimes disappointed (a much-publicized trial converted far worse than a classic purchase journey). Many of these payments will also be able to rely on stablecoins, notably for micro-transactions between machines (Chapter 11).

In context

Digital twins

A notion increasingly structures industry: the digital twin, a virtual replica of an object, a machine, a factory, even an organ, continuously fed by the sensor data of its real-world counterpart. AI plays two roles here: running the twin (simulating it faithfully) and exploiting its data to predict and optimize. One can thus test a modification on the twin before touching the real system, anticipate a breakdown (the predictive maintenance already mentioned), or optimize a production line without stopping it. The applications go beyond the factory: twins of power grids, of buildings, of entire cities, and even, in medicine, organ twins (a virtual heart to test a treatment). The principle echoes that of world models (Chapter 5) and of simulation in robotics (Chapter 13): having a faithful virtual double in which to try, at lower risk and lower cost, what one cannot afford to experiment with directly on the real thing. The limit is the same everywhere: a twin is only as good as the fidelity of its model and the quality of the data that feed it.

18.4A common denominator: data and trust

Key takeaways (Chapter 18)

Finance is an advanced use of AI (trading, fraud detection, credit, insurance), but with risks of discrimination, volatility and hallucination.
The autonomous vehicle is making progress: robotaxis (level 4) operate commercially in 2026, but level 5 (everywhere, all the time) remains distant because of the "long tail" of rare situations.
AI is also spreading into precision agriculture, industry (predictive maintenance, quality) and commerce (recommendation, dynamic pricing).
Common denominator: everywhere, AI runs on data and promises to optimize, but runs up against opacity, bias, reliability and the gap between demonstration and deployment. The stake is trust.

After the economic sectors, Chapter 19 descends closest to the intimate: daily life, relationships and psychology.

18.1Finance, banking and insurance#

18.2Transport and autonomous mobility#

18.3Agriculture, industry and commerce#

18.4A common denominator: data and trust#

Key takeaways (Chapter 18)

18.1Finance, banking and insurance

18.2Transport and autonomous mobility

18.3Agriculture, industry and commerce

18.4A common denominator: data and trust