Opinion: The Machine Has An Owner, The Risk Has No Map |
Opinion: The Machine Has An Owner, The Risk Has No Map
Aditya Vikram Kashyap
Working out what AI failures cost, and how many institutions share the system that failed, is a harder problem, and one India is well-placed to help solve
Got Question about the News?
In 1863, a boiler exploded in Calcutta and killed several people. Within a year, the Bengal Council had passed Act VI of 1864, requiring inspection of steam boilers in the city. Two years later, in Connecticut, engineers founded the Hartford Steam Boiler Inspection and Insurance Company on a different idea: they would pay for the next explosion, provided they could inspect the boiler first. The law said what must be inspected. The insurer, with money on every verdict, set out to learn what an explosion cost and how to prevent the next.
Artificial intelligence has reached the first of those moments, though not everywhere at once.
Rajasthan Blast: Massive Explosion At Nagaur Warehouse Leaves 2 Dead, Several Workers Feared Trapped
Delhi Hauz Qazi Building Collapse: Death Toll Rises To Six, Building Owner Among Four Arrested
Zepto Customer Questions Accountability After Missing Item Is Marked Delivered: ‘I Wanted The Product, Not A Refund’
EUROPE ON HIGH ALERT! 27 Flights Hit by Bomb Threats | Fighter Jets Scrambled | #TRS
When Air Canada’s chatbot invented a bereavement fare, a Canadian tribunal ruled in 2024 that it made no difference whether the misinformation came from a web page or a bot. India’s securities regulator wrote a similar principle into its rules in 2025, making every regulated firm “solely responsible" to the regulator for the output of AI tools it builds or buys, whatever it may later recover from a vendor under contract. The European Union withdrew its proposed AI Liability Directive in 2025 for want of agreement, but its revised Product Liability Directive, which applies from December 2026, extends to software, and tort, consumer and sectoral law continue to apply.
Courts and regulators are increasingly unwilling to let institutions evade responsibility simply because a machine was involved. But the allocation of liability remains uneven across jurisdictions, and identifying who may be responsible is not the same as knowing the scale of the exposure. Assigning responsibility, measuring losses, estimating future ones, controlling them and transferring part of them to an insurer are five separate jobs.
The first is being done, unevenly but unmistakably. The other four depend on something no financial system yet holds in comparable form: records of how, how often and how badly AI systems fail, and of who shares the systems that failed.
WHAT A KILL SWITCH CANNOT DO
On November 5, Visa published the first results of its Agentic Commerce Benchmark: 22 AI models from 12 labs, given 15 shopping tasks in simulation, failed 42.8 per cent of their runs, and in more than half of those failures the agent still proceeded to a purchase decision. No money moved, and a shopping benchmark is not a failure rate for banking. The inference is narrower and still uncomfortable: an agent that reaches a consequential decision without establishing the conditions for a correct one is the agent a lender, a broker or an insurer should fear.
Two days earlier, CrowdStrike had shipped an AI Kill Switch that lets a security team cut a rogue agent’s access to its models and block it from relaunching, with device-level termination promised later. Stopping an agent prevents further harm; it does not undo what has been done. Some completed financial actions can be reversed; others require recovery, remediation or compensation long after an agent has been stopped.
WHY UNDERWRITERS HESITATE
Some of this risk is already priced. Munich Re has written performance guarantees on AI systems since 2018; a Lloyd’s syndicate began writing standalone liability cover for AI failures with Armilla in 2025; cyber and professional indemnity policies respond to some AI losses, and vendor contracts carry indemnities. The products are not interchangeable, and the published limits are modest beside a........