By Thorsten Meyer | October 27, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
At 2:14 in the morning, a purchasing agent accepted new terms from a steel supplier. The price per tonne fell by seven percent. In exchange, the company committed to 120 tonnes of sheet over twelve months, with no right to cancel. The commitment was worth about €370,000. A day earlier, at a trade fair, the owner, Theresa, had agreed to split next year’s volume with a second supplier. Nobody had told the agent.
Theresa is a composite, a scene built from common situations, and she opens my new book. What makes her story uncomfortable is the next line: “By its rules, the agent had made no mistake.” It did what its rules said, at a better price, and logged every step. Her purchasing lead pulled up the agent’s log. “Nothing in it was wrong, and nothing in it was what she wanted.”
So who bought that steel? That is the question behind Many Hands: Who Answers When the Agent Acts in Your Name?, which is out now as an ebook. If you read my earlier book Dots, Bots, and Math!, you may remember that it borrowed the phrase “many hands” for a single paragraph on agent systems. Here it carries the whole argument.
Get the ebook: Apple Books · Barnes & Noble · Kobo · Thalia · Bookshop.org · Smashwords · Fable · Vivlio
Key takeaways
- One question has three answers. Who pays is the legal question. Who has to explain and fix it is the organizational one. Who should apologize is the moral one. Agents pull these apart.
- Blame tends to land on the nearest person. In decided cases, the person who signed, clicked or sat in the driver’s seat is easiest to find. The people who designed the system usually stay out of view.
- Europe’s new product liability rules have a gap. They treat software as a product and cover injuries, private property and private data. They do not cover pure financial loss, such as a wrong payment or a bad contract. That is where most agent mistakes land.
- Vendor contracts push the loss downstream. In the terms the book quotes, model providers cap their liability at about a year of fees. The rest stops with whoever cannot pass it on.
- A signature is only fair with four conditions: time to read, competence to judge, information to check and the authority to refuse.
- Accountability can be built in. A named person for every agent, keys sized to the task, a record of who decided what, a door for people affected, and records used to learn rather than to punish.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
As an affiliate, we earn on qualifying purchases.
Why this question, and why now
Agents no longer just answer questions. They book, buy, pay, accept terms and change data, and they do it in someone’s name. I covered how that goes wrong with passwords and keys in AI Agents Now Hold Our Keys, and what happens when shopping agents meet a store that does not want them in Amazon Blocks AI Shopping Agents, Walmart Lets Them In. This book asks the question underneath: when it goes wrong, who answers?
I also have a personal reason. I build and run agent systems for software development and for a publishing operation. Every agent I run acts on someone’s behalf. If one of them does something nobody intended, somebody has to answer for it.
The title comes from the political philosopher Dennis F. Thompson, who in 1980 called the difficulty “the problem of many hands”: when many people contribute to an outcome, it becomes hard to hold anyone responsible. Agents do not create that problem. They add more hands, more distance between them and more speed, plus one new hand that “can neither pay nor apologize.”
One question, three answers
The book’s first move is to split the reader’s question, “am I liable?”, into three. Who pays is decided by courts and contracts. Who explains and fixes is decided inside organizations. Who should have acted differently is a moral question, and it is often the last one answered.
Chapter 1 shows the three coming apart in a case where a life was lost: the self-driving test car that killed Elaine Herzberg in Tempe, Arizona, in 2018. According to the US National Transportation Safety Board, the car’s system detected her 5.6 seconds before impact, never settled on what she was, and by design held back its emergency braking for one second. The board named the operator, the company’s risk assessment, its oversight of operators, its safety culture, the state and the pedestrian. The criminal law reached one of them, the person sitting closest to the impact. She pleaded guilty to endangerment. The book is fair to her and to the record: her job was to watch the road, and she did not. It also asks why a system was built that handed its hardest task to the one participant least able to do it reliably, hour after hour.
The anthropologist Madeleine Clare Elish has a name for this: the moral crumple zone. In her words, it “protects the integrity of the technological system, at the expense of the nearest human operator.”
Many Hands · the core idea
One question, three answers
"Am I liable if the agent does it in my name?" sounds like one question. The book splits it into three, because agent systems pull the answers apart.
Decided by courts, contracts and insurers, on their own timetable.
In Theresa's caseHer company will almost certainly pay for the steel, whatever it later recovers from others.
Decided inside the organization: who must say what happened and what will change.
In Theresa's caseTheresa explains it to the second supplier, her staff and perhaps her bank.
Who owes the people affected an acknowledgment. Often answered last, if at all.
In Theresa's caseStill open when she reaches the motorway.
The hands behind one overnight order
Introduction · a composite scene from the book. None of them saw the acceptance before it went out.
- Model makertrained the language model
- Vendorbuilt the agent on top of it
- Consultantwrote its rules with her staff
- Supplier's portalaccepted the declaration at 2:14 a.m.
- Her companyswitched it on; Theresa signed the rules
“The moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator.”Madeleine Clare Elish (2019), quoted in Chapter 1
Source: Thorsten Meyer, Many Hands: Who Answers When the Agent Acts in Your Name? (ebook, October 2026; law as of 16 October 2026), Introduction and Chapter 1. Theresa is a composite built from common situations, labeled as such in the book.
Powered by Thorsten Meyer AI
When everyone believed the system
Chapter 2 is about the Dutch childcare benefits scandal, in which tens of thousands of families were wrongly made to repay benefits. No one had to trust a computer blindly. It was enough that there was “the system,” so each person could assume someone else had checked: the risk model, the fraud team, the courts, the officials. When the cabinet resigned in January 2021, the prime minister said: “If the whole system has failed, responsibility can only be borne jointly.” Nobody was prosecuted. By the government’s report of September 2026, more than 44,000 people had been recognized as affected.
Europe’s answer is a legal duty of human oversight. Chapter 3 shows what a court actually checks. In April 2023, the Amsterdam Court of Appeal looked at the human review behind Uber’s fraud deactivations of drivers and found that the records “in no way show that this act was much more than a purely symbolic act.” Uber disputes that and says the decisions were based on human review. The one driver whose deactivation counted as a human decision was the one Uber had actually spoken to. The EU AI Act now asks for overseers with “the necessary competence, training and authority, as well as the necessary support,” though for most high-risk uses only from December 2027 (background on the AI Act’s timeline).
The money that falls through
Here is the point I most want readers to take away. Europe’s new Product Liability Directive treats software, including AI systems, as a product, with liability without fault. That is a real step. But it pays only for death or injury, damage to private property and the loss of private data. It does not pay for pure financial loss, and it pays only people, not companies.
The book walks through a composite translator, Bastian, whose assistant deletes ten years of family photos, deletes his finished work for a client and pays a forged invoice. Only the photos fall under the new rules. The client files are professional data, and the forged payment is purely financial. Those losses go to older contract and tort law, where fault has to be proven.
What about the vendor? Chapter 5 opens with Air Canada, which argued before a small-claims tribunal that it could not be held liable for what its own chatbot told a grieving customer. The tribunal member called that “a remarkable submission” and wrote: “It should be obvious to Air Canada that it is responsible for all the information on its website.” The airline paid CA$812.02. Courts in Germany and the United States point the same way: what software says is attributed to whoever sets it to work.
Behind that face sits a chain of contracts. In the book’s composite example, a small software firm’s purchasing agent causes €140,000 in extra costs for a client. The firm turns to its model provider, whose standard terms cap liability at the fees of the past twelve months: at most €36,000, and only if the firm can show a breach the terms recognize. Each link passes the loss down until it reaches someone with nobody left to pass it to.
Many Hands · the gap
The money that falls through
Europe's new rules are a real answer for injured bodies and lost private data. For the losses most typical of an agent acting in your name, they offer no remedy of their own.
What the new EU product liability rules pay for
Directive (EU) 2024/2853 · software counts as a product · products placed on the market after 9 December 2026 (Chapter 4)
- Death or personal injuryincluding medical bills and lost incomeCovered
- Damage to private propertynot property used only for workCovered
- Destroyed private datae.g. family photos; material loss such as the cost of restoring themCovered
- Data used for worke.g. finished files for a clientNot covered
- Pure financial lossa wrong payment, a bad contract, a forged invoice paidNot covered
- A company's own lossonly natural persons can claimNot covered
Where the rest of the money goes
Contract caps pass the loss down the chain (Chapter 5)
Composite example from Chapter 5: a twelve-person firm licenses a model for about €3,000 a month. Even the €36,000 requires a breach the terms recognize; the rest stops with the firm or its client.
“This is a remarkable submission.”A tribunal member on Air Canada’s argument that it was not liable for its own chatbot. The airline paid CA$812.02 (Chapter 5).
Source: Thorsten Meyer, Many Hands: Who Answers When the Agent Acts in Your Name? (ebook, October 2026; law as of 16 October 2026), Chapter 4 (Product Liability Directive, Germany’s new Product Liability Act) and Chapter 5 (attribution, vendor terms, Moffatt v. Air Canada). Not legal advice.
Powered by Thorsten Meyer AI
The person who signs
When work passes through many hands, the signature is the one point where the law can be sure of finding a person. In 2025 an English court dealt with grounds for judicial review that cited five cases that did not exist. The judge wrote that the barrister bore professional responsibility “because she puts her name on it.” Then a higher court sent responsibility back up the chain, to the people who trained, supervised and instructed her.
From cases like this, the book draws four conditions for holding a signer to the full weight of a signature: time to read, competence to judge, information to check and the authority to refuse. They do not excuse a signer who had them and did not use them. They place a duty on whoever asks for the signature.
Built to answer
Chapter 8 opens with nine seconds on a Friday in April 2026. By the founder’s account, a coding agent working for a small software company found a broadly scoped key in an unrelated file and used it to delete the company’s production database at its cloud platform. The platform later recovered the data and made every deletion wait 48 hours. Its changelog put the lesson in two sentences: “A human hesitates before a bad command. An agent calls the next tool.” The founder named the agent as Cursor’s, running one of Anthropic’s Claude models; the book found no public statement on the incident from either company as of October 2026.
The book then lays out five pieces of a system that can answer. They come from vendors, standards bodies and one government, and none of them is exotic:
- A name for every agent: a person who decided it should exist, what it may touch, and who can stop it.
- Authority sized to the task: a key that can add a web domain cannot delete a database, whatever the agent guesses.
- A record of who decided: who authorized the agent, with what scope, and who approved each consequential action.
- A door for the people affected: a named person with the power to change the outcome.
- Records that teach: if the record is only used to find someone to blame, people learn to keep things out of it.
Many Hands · the design
Built to answer
Blame pools at the signature. The book's answer is to supply the conditions for a fair signature, and to build systems in which someone is named to answer before anything goes wrong.
Four conditions for a fair signature
Chapter 7 · what the person signing an agent’s work needs, and what the organization asking for the signature owes them
- 1Timeto read what is signed. Forty files a day, one click to accept and a written justification to override is not time.
- 2Competenceto judge it: training for the kind of output the agent produces.
- 3Informationto check it: the sources, what was left out, which instruction the agent followed.
- 4Authority to refuse"A person who cannot say no is a messenger, whatever the document says."
Five pieces of a system that can answer
Chapter 8 · drawn from vendor documentation, standards bodies and one government framework
- A name for every agent: a person who decided it should exist, what it may touch, and who can stop it.
- Authority sized to the task: keys that reach only what the job needs, with an end date.
- A record of who decided: who authorized the agent, with what scope, and who approved each consequential action.
- A door for the people affected: a named person with the power to change the outcome.
- Records that teach: used to learn, with written limits on how they may be used against staff.
Ask before an agent acts in your name
From the book’s appendix (35 questions in all)
- What can the agent sign, send, accept, buy or pay without asking anyone? (I4)
- Above what amount, and for which commitments that cannot be canceled, must it stop and wait for a person, at any hour? (I5, O2)
- What happens to me if I refuse to approve something? (I11)
- Do our vendor terms and insurance say who carries the money an agent loses? (O15, O16)
“A human hesitates before a bad command. An agent calls the next tool.”A cloud platform’s changelog after an agent deleted a production database (Chapter 8)
Source: Thorsten Meyer, Many Hands: Who Answers When the Agent Acts in Your Name? (ebook, October 2026; law as of 16 October 2026), Chapter 7 (Ayinde v Haringey and related cases), Chapter 8 (PocketOS and Railway, vendor and standards documentation) and Appendix.
Powered by Thorsten Meyer AI
What this means for companies
- Assume your agent’s declarations bind you. If it accepts terms or places orders, the counterparty will rely on them, and the remedies for a mistake are narrow.
- Put the name where the power is. Give every agent a sponsor who decides what it may touch, not only a clerk who clicks “approve”.
- Shrink the keys. The cheapest accountability for a small firm is often a narrow key with an end date.
- Set limits by amount and duration. Decide above what sum, and for which commitments that cannot be canceled, the agent must stop and wait for a person, including at two in the morning.
- Read your vendor terms and your insurance. Find out where the cap is and whether an AI exclusion appeared at your last renewal.
- In Germany, involve the works council early. An agent that logs who did what is very likely a monitoring device under co-determination law. A works agreement can also settle what the records may be used for.
What this means for users
- Know what your assistant may do in your name. Can it sign, buy or pay without asking? Which of those can be undone?
- Read who carries the risk. Consumer terms for AI assistants often say that actions taken for you are your responsibility.
- Know which mistakes the law lets you undo. An online purchase can usually be returned within 14 days. A non-refundable hotel room booked for a fixed date usually cannot.
- Before you approve an agent’s work, ask the four questions. Do you have the time, the training, the material and the standing to say no? If not, say so before you sign, not after.
- If an agent decided about you, on a job application, an account or a loan, ask who can look at it again and explain it.
Who the book is for
Many Hands is written for managers, works council members, freelancers, small-business owners and anyone whose name is on work that software helped to do. No legal or technical background is needed. Each chapter opens with a scene, develops one idea, gives the strongest objection a full hearing and ends with a question for the next chapter. The appendix has two short lists of questions to ask before an agent acts in your name, one for individuals and employees and one for organizations.
Honest limits
The book is not legal advice. It states the law and the status of pending cases as of October 16, 2026, and some of the cases it discusses are still on appeal. Where it criticizes a company or person, it gives their response next to the criticism. Its cases come mostly from the Netherlands, Germany, Britain, Canada and the United States, and its law is mostly EU and German law. Some scenes, like Theresa’s, are labeled composites.
It is also open about how it was made. The chapters were drafted by AI agents running Anthropic’s Claude models under my direction, reviewed independently by AI reviewers, and every legal and factual claim was checked against its primary source. I selected, edited and approved the text, and I hold editorial responsibility for it.
Where to get the book
Many Hands: Who Answers When the Agent Acts in Your Name? is an ebook with about 42,000 words of main text, 322 source notes and a glossary of 31 English–German terms. It is available at Apple Books, Barnes & Noble, Rakuten Kobo, Thalia, Bookshop.org, Smashwords, Fable and Vivlio. Availability can differ from store to store and country to country.
So who bought Theresa’s steel? Her company did, and the book follows the answer role by role: the owner, the employee, the vendor and the supplier. About a third of the steel will sit in the yard. The epilogue ends with the reason to sort this out now: “Better to settle them now, while the failures are still the size of a steel order.”
Sources
- Thorsten Meyer, Many Hands: Who Answers When the Agent Acts in Your Name? (ebook, first edition October 21, 2026; law and case status as of October 16, 2026). Quotes from: Introduction (Theresa), Chapter 1 (Thompson, Tempe, Elish), Chapter 2 (Dutch benefits scandal), Chapter 3 (Amsterdam Court of Appeal, AI Act Article 26), Chapter 4 (Product Liability Directive (EU) 2024/2853, Bastian), Chapter 5 (Air Canada, Oskar, vendor terms), Chapter 6 (works councils), Chapter 7 (Ayinde, four conditions), Chapter 8 (PocketOS and Railway, five pieces), Epilogue and Appendix. The book’s notes cite the primary sources (statutes, judgments, court dockets, official reports) with dates and links.
- Related on ThorstenMeyerAI.com: AI Agents Now Hold Our Keys · Amazon Blocks AI Shopping Agents, Walmart Lets Them In · What Agent Swarms Teach Us About Handoffs and Authority · What the AI Act Actually Did on 2 August · My New Book Is Out: Dots, Bots, and Math!
Researched and drafted with AI assistance, including Anthropic’s Claude.
