Who pays when your AI agent breaks a client's data?

We used to build chatbots that only talked. Now we build agents that take action: they read emails, update CRMs and move money. That change moves the risk from a bad answer on a screen to a bad action in a client's systems.

The core rule is short. You cannot guarantee that an AI agent will never hallucinate or never be prompt injected. Before you give an agent write access to a client's database, CRM or payments, get the client to sign an AI behavior indemnity clause that puts that risk in writing.

Why AI agents carry more risk than chatbots

An AI agent is software that uses a large language model (LLM) to decide and then act inside other systems. A chatbot only returns text. If a chatbot is wrong, a person reads the error and ignores it. If an agent is wrong, the error lands in the CRM, the database or the bank account.

Attackers know this. In November 2025, Anthropic published a report on how hackers use agentic tools for autonomous cyberattacks. Every business owner who deploys agents should read that as a warning: the same autonomy that saves your team hours also gives an attacker a new way in.

Who is liable when an AI agent deletes data?

Nobody can answer this cleanly today, and that is the problem. Imagine that you build an autonomous agent for a client. The agent has write access to their database. Someone tricks it, and it deletes a table. Is that your fault, the client's fault, or nobody's?

Two failure modes make this question unavoidable:

• Hallucination. The model produces a confident output that is wrong, and the agent acts on it.
• Prompt injection. An attacker hides instructions inside an input, such as an email, and the agent follows them.

You cannot remove either risk with better prompts or better code. You can only reduce it, and you can decide in advance who carries what is left.

What an AI behavior indemnity clause does

An AI behavior indemnity clause is a contract term in which the client accepts the risk that an AI agent acts wrongly. It names the risk (hallucination and prompt injection) and records that the client knew about it before go-live.

Here is what 15 years of building software taught me: the technology is usually ready before the paperwork is. With agents, the tech is ready, but your legal protection probably is not. The clause closes that gap before an incident, not after it.

How to deploy a write-access agent safely

Use these steps for every agent that can change data in a client's systems.

1. List what the agent can write to. Name each system: database, CRM, email, payments. This list tells you, and the client, how large the risk is.
2. Tell the client the limits in writing. State that no one can guarantee an LLM will not hallucinate or be prompt injected. Do not soften it.
3. Get the AI behavior indemnity clause signed. The client signs off on the risk before you give the bot the keys to the castle.
4. Turn on write access last. Do not deploy a write-access agent without a signed waiver. Read-only agents can go first while the contract is in review.

For the operational guardrails that sit next to the contract, read 5 rules to stop an AI agent from damaging your live data. If you operate in the EU, check the EU AI Act Compliance Checklist as well.

The bottom line

Agents that act are useful, and they are here. But an agent with write access and no signed risk agreement is a ticking time bomb for the builder. Put the AI behavior indemnity clause in the contract first, then hand over the keys.

Build a calmer business,Yuri

Frequently asked questions

Who is liable if an AI agent deletes a client's database?

Liability is unclear when an AI agent with write access to a client's database gets tricked into deleting a table. Nobody can guarantee that an LLM will not hallucinate or fall for a prompt injection. For this reason, the builder and the client must agree in writing, before deployment, who carries that risk. An AI behavior indemnity clause records that agreement.

What is an AI behavior indemnity clause?

An AI behavior indemnity clause is a contract term in which the client accepts the risk that an AI agent acts wrongly, for example because the model hallucinates or an attacker uses prompt injection. Builders of autonomous agents should get the client to sign this clause before they give the agent write access to databases, CRMs or payments.

Can I guarantee that an AI agent will not hallucinate or be prompt injected?

No. A builder cannot guarantee that a large language model will never hallucinate or never be prompt injected. That risk stays with every AI agent, even a well-built one. The practical answer is to tell the client about the risk in writing and get a signed waiver before the agent gets write access to their systems.

Why are AI agents riskier than chatbots?

A chatbot only produces text, so a wrong answer stays on the screen. An AI agent takes action: it reads emails, updates CRMs and moves money. When an agent makes a mistake or gets tricked, the mistake changes real data. In November 2025, Anthropic also reported that attackers used agentic tools to run autonomous cyberattacks.