Who is responsible for AI mistakes?
Who is responsible for AI mistakes? As a rule of thumb, the company that puts an AI to work answers for what it does, much as it answers for its other tools and for its staff. An AI is not a person and cannot carry responsibility itself. This article gives a general orientation. It is not legal advice, and the details depend on your country, your line of business and your contracts.
Why "the AI did it" is not an answer
A customer deals with your company, not with your tools. If a bakery's till prints the wrong price, the customer talks to the baker, not to the maker of the till. It is the same with AI. If an AI agent, a program that uses AI to carry out tasks for you, sends a wrong price, your company sent a wrong price.
Whether you are bound by that price, and whether you can turn to a supplier afterwards, are legal questions for an adviser. For everyday practice the safe assumption is simple: what leaves your company under your name is yours.
What you can settle today
You cannot settle the legal questions alone. You can settle the internal one: who in your company looks after what. An agent that nobody looks after is the real risk. A few clear roles are enough. In a small company, one person may hold several of them.
| Role | Answers for | Example |
|---|---|---|
| Management | Which tasks AI may do, and the rules for them | "Agents may draft customer emails. They may not send them." |
| Agent owner | What the agent is for and how well it works | The head of sales owns the follow-up agent |
| Administrator | What the agent can reach | The follow-up agent sees deals, not the payroll |
| Approver | Each proposal they say yes to | The sales rep reads the draft before it goes out |
Three situations, sorted
- The agent acted alone, within its rules, and the result was wrong. The rule allowed too much. Management and the agent owner tighten it.
- A person approved a wrong proposal. The approver answers for the yes. But if proposals are too long or too unclear to check, the agent owner has work to do too.
- Someone used an agent for something it was not meant for. That person answers first. If the agent should never have had that access, so does the administrator.
One page to put in writing
A small company does not need a thick AI policy. One page covers it.
- Which tasks AI may do, and which it may not.
- Which actions need a person's yes.
- Who owns each agent, by name.
- How a mistake is reported, who corrects it and who tells the customer.
- A date on which the page is reviewed.
Show this page to your legal and insurance advisers once. They can tell you what your line of business adds to it. Also read what your contracts with software suppliers say about mistakes before you rely on an agent for anything that costs money.
Responsibility is not blame
The point of naming owners is that someone acts when something goes wrong. If saying yes to an agent's proposal means carrying all the blame, people stop saying yes and the agent stops being useful. Make it normal to report a wrong result quickly. Treat each report as a reason to improve the setup. And keep a record of what the agent did, so that the discussion is about facts.
How Worknano handles this
In Worknano an agent works within the permissions of the person who asked. So every agent action belongs to a named user. Agents propose changes, and that person sees the values and confirms them. Decisions that belong to a manager go through the Approvals module, where every approval and rejection is logged with the time and the user. Approvals follow the structure of your organisation, so the request reaches the right manager.
- Approvals: who signed off, and when
- Roles and permissions
- Organization: teams, roles and reporting lines
Create a free workspace and give every agent action a named owner.