Who is responsible when AI gets it wrong, or acts in my name?

The short answer

Who is responsible when AI gets it wrong? In practice, the customer looks to you. They dealt with your business, not with your software. If a promise is wrong, they come back to the person whose name is on the work.

Who answers in law, in a particular case, depends on the facts, the contract and the rules where the business works. That is a question for a lawyer. This post explains how the arrangement works and what one piece of EU law says, so a better question can be asked.

What you can control is where a tool is allowed to decide. Felesh does not run the business. It drafts, and you decide. The notes, dated promises and next steps Felesh drafts from your sentence wait for your yes, with the reason beside each. Tasks are added directly.

Can AI make mistakes?

Yes, and it is better to plan for it than to be surprised by it.

A tool can read a sentence the wrong way. It can take a date you meant as this Friday and place it a week later. It can answer a customer from a document that was true last year and is not true now. None of this is exotic. It is what happens when something works from words and has no way of knowing what you had in mind.

The useful question is whether a mistake reaches you before it becomes a promise, or after a customer acts on it.

What waits for your yes

After a conversation, you type one sentence about what happened. Felesh drafts a note, the promise with its date if you said one, and the next step. Each draft shows its reason and waits for your yes. In the panel, each draft is marked “Review draft”.

The reason matters as much as the draft. You see why Felesh wrote what it did, so you can tell a sound draft from a wrong one without reconstructing the call from memory.

Tasks work differently. A task is added directly, with no approval step. It is a line on your own list, so check your tasks the way you check any list you keep yourself.

That is the line. Notes, dated promises and next steps wait for you. Tasks go straight onto the list.

What if a customer gets a wrong answer?

This is the part people worry about most, because it is where the software speaks to someone outside the business.

Customer messages from your website chat, Instagram, WhatsApp and Telegram arrive in one inbox. Felesh answers from what you taught it: the documents you uploaded and the instructions you wrote.

When Felesh cannot answer, it does not fill the gap. The message goes to “Needs you” with the reason it could not answer. Nobody is notified, so the message waits there until you look. When a person joins the conversation, Felesh stays quiet.

There is an honest limit here. If a document you uploaded is out of date, the answer will follow it. Felesh answers from what it was given, so keep what you teach it current. That part of the work stays with you.

A walk-through: one call, one wrong date

Take an ordinary afternoon. You finish a call with a customer and promise to send a revised quote by Friday. You type one sentence: what happened, and what you promised.

Felesh drafts three things. There is a note of the call, the promise with a date, and the next step. Each shows its reason.

Suppose the date reads wrong. Perhaps the draft has put the promise a week later than you meant. You see that in the draft, with the reason beside it, before it becomes a dated promise in your records. You do not give your yes to a date you would not stand behind.

Now suppose it reads right. You say yes, and the promise is there with its date, where you can find it later. You can ask Felesh in your own words what you promised that customer. It answers from your own records, which is why those records are worth checking before you approve them.

In both cases the same person decides. Felesh did the writing.

What does “deployer” mean in the EU AI Act?

If you read about the EU AI Act, you will meet the word “deployer”. It comes from Regulation (EU) 2024/1689, which defines it in Article 3.

Article 3(4) says a deployer is a natural or legal person, public authority, agency or other body using an AI system under its authority, except where the AI system is used in the course of a personal non-professional activity.

Article 3(3) says a provider is a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model, or has one developed, and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge.

The Act gives the two roles different obligations. For high-risk AI systems, for example, Article 16 sets out obligations of providers and Article 26 obligations of deployers. Article 25(1) also provides that, for a high-risk AI system, a deployer is considered a provider in certain circumstances, such as putting its own name or trademark on a system already on the market or making a substantial modification to it.

On liability, recital 9 of the Act says that its rules are without prejudice to existing Union law, and that the rights and remedies that law gives to consumers and other persons on whom AI systems may have a negative impact, including compensation of possible damages under Council Directive 85/374/EEC, remain unaffected and fully applicable. The Act does not itself answer who pays in a particular case.

Whether a given business is a deployer, a provider or neither for a given tool depends on the facts and on the text of the Act. That is a question for a lawyer who knows the sector.

Where to start

Start with one conversation. Type what happened, read the draft and the reason beside it, and decide whether you would stand behind it. If you would not, you have found the mistake while it is still cheap to fix.

To see how drafts, reasons and your yes fit together, read about drafts and approvals.

This is general information, not legal advice.