In most cases, your business is responsible. If an AI system you chose to use gives a customer wrong information, makes a bad decision or takes a wrong action, the customer, regulator or court will generally look to you, the business that deployed it, not to the software. You may be able to recover some losses from the AI vendor or developer under your contract, but AI liability toward the people you serve usually starts with you.
That's not a reason to avoid AI. It's a reason to deploy it the way you'd deploy a new employee: with clear limits, supervision, records and the right contracts. This post explains how responsibility typically splits between you, your vendors and your staff, then covers the practical controls that reduce your exposure.
Not legal advice: liability rules differ by country, by sector and by contract, and the law on AI is still developing. This is a general framework for thinking about the issue. Speak to a lawyer about your specific situation.
Why can't you blame the AI?
Because, legally and practically, AI isn't a person. It can't sign a contract, pay damages or be held accountable. When you put an AI system in front of customers or let it act in your systems, you're choosing to use a tool, and the consequences of that tool's output flow through your business.
An analogy: if a delivery van's brakes fail and it damages a shopfront, the shop owner claims against the delivery company. The delivery company may then claim against the van manufacturer or the garage that serviced it. The AI situation is similar. The person harmed deals with you; you may then look upstream.
Courts and tribunals in some places have already rejected the argument that a business isn't bound by what its chatbot said. Whatever the specifics, the direction of travel is clear: customers are entitled to rely on what your business tells them, through whatever channel.
Who carries what when AI goes wrong?
Responsibility usually spreads across several parties. A simplified picture:
| Party | What they typically answer for | How it's usually handled |
|---|---|---|
| Your business | Outcomes for customers, employees and regulators; choosing and supervising the AI | Your own controls, terms of service, insurance |
| AI model provider | The model service itself, within the limits of their terms | Their terms of service, which often limit their liability significantly |
| Developer or agency that built your system | Building what was agreed, to the agreed standard | Your development contract: warranties, acceptance testing, liability caps |
| Software platforms in the chain | Their own service availability and security | Their terms and SLAs |
| Your staff | Following procedures, reviewing what they're meant to review | Your policies, training and employment terms |
Two practical observations. First, standard terms from large AI providers usually cap their liability and disclaim responsibility for the accuracy of output, so don't assume the vendor will cover a mistake. Second, a custom build gives you more room to negotiate, so the development contract matters.
What kinds of AI mistakes create liability?
The type of mistake shapes the type of exposure:
- Wrong information to a customer. A chatbot quotes the wrong price, refund policy or delivery date. Typical exposure: honoring what was said, consumer complaints, reputational damage.
- Wrong action in your systems. An AI agent, meaning AI software that can take actions, issues a duplicate refund, sends an invoice to the wrong client or deletes records. Exposure: direct financial loss, data breach obligations, client disputes.
- Unfair or biased decisions. AI used in screening job applicants, approving credit or prioritizing claims disadvantages a group. Exposure: discrimination claims, regulatory scrutiny.
- Data leaks. Confidential or personal data sent to the wrong place or exposed through an AI tool. Exposure: privacy law obligations, contract breaches, notification duties.
- Professional errors. AI-drafted advice, reports or calculations used in professional work without proper review. Exposure: professional negligence, regulatory discipline.
- Infringing content. AI-generated material that copies someone else's work. Exposure: IP claims.
Our guide to AI hallucinations and business risk covers the most common source of the first and fifth categories.
How do you reduce your AI liability in practice?
You can't contract your way out of all responsibility to customers, but you can make mistakes rarer, smaller and easier to defend. The controls that matter most:
- Define what the AI is allowed to do. Write down its job, its limits and what it must hand to a person. An AI that can only draft refunds is a different risk from one that can issue them.
- Put humans at the right checkpoints. Approval steps for anything that moves money, changes important records, makes decisions about people or gives regulated advice. Our guide to human-in-the-loop AI shows where to place them.
- Build guardrails into the system. Spending limits, allowed actions, input and output checks, and refusal rules. See AI agent guardrails.
- Keep logs. Record what the AI saw, what it did, and who approved it. If a dispute arises, logs are your evidence. Our post on AI agent monitoring and observability explains what to capture.
- Ground answers in real policies and prices so customer-facing AI can't invent terms.
- Disclose AI use and offer a human route, which reduces the chance a customer feels misled.
- Test before launch and review regularly with real examples, and keep the test results.
What should your contracts say?
Contracts are where you allocate risk with the people upstream and downstream of you. Topics to raise with your lawyer:
With AI vendors and developers:
- Warranties about how the system will perform and what it was tested against.
- Responsibility for security and data protection, including breach notification.
- Data terms: training use, retention, location.
- Liability caps and exclusions, and whether they're proportionate to your exposure.
- IP indemnities for generated content or code, where offered.
- Support, fixes and response times when something goes wrong.
With your customers:
- Clear terms about what your AI service does and doesn't do.
- Disclosure that AI is used, where appropriate.
- Reasonable limitation of liability, bearing in mind that consumer protection law in many places restricts how far you can limit it.
If you're hiring someone to build AI for you, our guide to hiring an AI development agency covers contract and acceptance questions.
Does business insurance cover AI mistakes?
It might, partly, or not at all. Professional indemnity, cyber, general liability and errors-and-omissions policies were mostly written before AI agents were common, and wording varies. Some insurers are adding specific AI language, both covering and excluding.
Questions to ask your broker:
- Does my professional indemnity policy cover errors in work where AI was used?
- Does my cyber policy cover incidents caused by an AI tool or agent, such as a prompt-injection attack?
- Are there exclusions for automated decisions or AI-generated content?
- Do I need to tell the insurer I'm using AI in customer-facing work?
Get answers in writing, and revisit them when you expand what AI does in your business.
What should you do when an AI mistake happens?
Have a simple response plan before you need it:
- Stop the bleeding. Pause the workflow or remove the AI's permission for the affected action.
- Fix it for the customer. Correct the error, honor reasonable expectations, apologize plainly.
- Check obligations. Data breach notification, regulatory reporting or contract notices may have deadlines.
- Find the cause in the logs. Was it the model, bad data, a missing guardrail or a skipped review?
- Fix the system, add a test case for that failure, and document what changed.
Frequently Asked Questions
Can I put a disclaimer on my chatbot to avoid liability?
A disclaimer like "AI responses may be inaccurate" can help set expectations, but it generally won't protect you from responsibility for what your business tells customers, and consumer law in many places limits such disclaimers. Accuracy and human escalation matter more than the small print.
Is the AI vendor liable if their model gives a wrong answer?
Usually only to a limited extent, if at all, under standard terms, which tend to disclaim accuracy and cap liability. Custom development contracts can be negotiated differently. Read the terms and get legal advice on significant deployments.
Are employees responsible for mistakes made with AI tools?
Employees can be held to your policies, for example a rule that AI-drafted client work must be reviewed. But toward customers, the business usually remains responsible for its employees' work. Clear policies and training protect both.
Does the EU AI Act change who is liable?
The EU AI Act sets obligations for providers and deployers of certain AI systems, especially high-risk ones, and may affect you if you operate in or sell into the EU. Our overview of the EU AI Act for AI agents covers the basics; get specific legal advice for your case.
Should small businesses avoid AI because of liability?
No. The risks are manageable with sensible limits, human review, logs and contracts, and doing nothing has costs too. Start with lower-risk internal uses and expand customer-facing AI as your controls mature.
Want AI that's built to be accountable?
We design AI agents with clear permissions, approval steps and audit logs from day one, so you always know what happened and who signed off. Tell us what you want to automate, and see our approach to agentic AI development.