A business is legally responsible for what its customer-facing chatbot says and does, whether it built the bot or bought it from a vendor. The main exposures are liability for inaccurate or unauthorized answers, missing or late AI disclosure, privacy and data-retention failures, and the lack of records to prove how the agent actually behaved. The most reliable way to reduce all four is independent testing of the agent as a customer would meet it, with evidence you can keep.
This guide is for business leaders, compliance and regulatory teams, CX owners and technical teams who run AI agents in customer service, including US companies entering the European market.
TL;DR
- Courts and regulators treat a chatbot’s statements as the company’s own statements. Outsourcing the vendor does not outsource the liability.
- US exposure comes mainly from deceptive-practice and misrepresentation claims and a growing set of state disclosure laws. EU exposure adds the AI Act’s transparency duty, applicable from 2 August 2026.
- Privacy risk sits in what the underlying model providers retain and train on, and in what customers type into your bot.
- Vendor assurances and dashboards are not evidence. Independent, repeatable testing of real conversations is.
- Lexic audits conversational AI agents from any vendor across customer experience, security, technical reliability and regulatory compliance.
Table of Contents
- Why a chatbot’s words become your legal problem
- Disclosure: when customers must be told it is AI
- Privacy and data risk in chatbot conversations
- What regulators expect you to be able to show
- A practical control framework
- How independent audit fits
- Where Lexic fits
- Our perspective
- FAQ
- Sources
Why a chatbot’s words become your legal problem
The clearest early example is Moffatt v. Air Canada (2024), a decision of the British Columbia Civil Resolution Tribunal. A customer relied on the airline’s chatbot for information about bereavement fares. The information was wrong. The tribunal held the airline liable for negligent misrepresentation and rejected the argument that the chatbot was a separate entity responsible for its own statements. The amount was small. The principle was not: the bot is part of the website, and the company answers for it.
Official guidance points the same way. The UK government’s consumer-law guidance on AI agents states that organizations are responsible for what an AI agent does in the same way they are responsible for what an employee does, even when a third party supplies the agent. That is a UK source, but US and EU practitioners reason along the same lines.
In the United States, law firm analysis of the current landscape describes several routes by which a bot’s statements can bind or expose a company: misrepresentation and unfair or deceptive practices claims, and arguments of reliance, waiver, estoppel or contract, depending on the facts. A chatbot does not need formal authority to create an obligation. A customer only needs a reasonable basis to rely on what it said.
In Europe, practitioners flag unfair-competition and misleading-advertising claims when AI-generated customer communications are inaccurate.
The four ways a chatbot creates liability
| Failure | Example | Typical legal theory |
|---|---|---|
| Wrong answer | Misstated refund, price or eligibility rule | Misrepresentation, deceptive practice |
| Unauthorized commitment | Bot promises a discount, exception or outcome | Reliance, estoppel, contract |
| Missing disclosure | Customer believes they are talking to a person | State disclosure laws, deceptive practice, EU AI Act |
| Failed handoff | Customer in distress or dispute is never routed to a human | Consumer protection, sector-specific duties |
The last row is the one most teams underestimate. In Lexic’s audits, 85% of the agents assessed failed to escalate issues to a human representative appropriately. An agent that keeps a customer in a loop during a complaint, a cancellation dispute or a safety-relevant question turns a service problem into a documented compliance problem.
Disclosure: when customers must be told it is AI
Disclosure is where the legal rules are most concrete, and they differ by jurisdiction.
United States: a patchwork of state laws
There is no single federal chatbot-disclosure statute. Companies face state laws with different triggers:
- Utah requires a supplier using generative AI in a consumer transaction to disclose that it is AI and not human when the individual clearly and unambiguously asks. For regulated services, prominent disclosure is required for high-risk interactions: verbally at the start of a verbal interaction, and in writing before a written one.
- New Jersey and California have their own rules. According to a January 2026 analysis by DLA Piper, New Jersey requires clear and conspicuous disclosure at the start of a bot interaction involving sale or advertising of merchandise or real estate, while California’s BOT Act applies where a bot is used in a commercial transaction with intent to mislead about its artificial identity.
The same analysis notes that the federal standard is limited to the commercial context and applies where a bot’s presence is unexpected and material to a consumer’s choice. State consumer-protection laws against deceptive practices apply in the same way. In practice, the baseline expectation of clear disclosure remains whichever AI-specific state laws survive debates on federal preemption.
The practical consequence for a national US business: build to the strictest trigger you face, disclose at the start of the conversation, and verify that the disclosure actually appears in every channel and every flow, including the ones a vendor update changed last week.
European Union: Article 50 of the AI Act
For US companies entering Europe, this is the rule to understand first. Under Article 50(1) of the EU AI Act, applicable from 2 August 2026, providers of AI systems that interact directly with people must inform them that they are interacting with AI, unless this is obvious. The notice must be clear and distinguishable and given no later than the first interaction. The European Commission’s guidance says the duty applies when four criteria are met: the system is AI, it is designed for genuine two-way exchange, it communicates directly with people, and it interacts with natural persons.
Timing matters for related obligations. The Digital Omnibus on AI, effective from 27 July 2026, postponed other AI Act deadlines: Article 50(2) watermarking to 2 December 2026, stand-alone high-risk system rules to 2 December 2027, and rules for high-risk systems embedded in safety-regulated products to 2 August 2028. The chatbot disclosure duty under Article 50(1) was not among the postponed items listed there. Check the official EU pages for the current status before you finalize a compliance plan.
Privacy and data risk in chatbot conversations
Customers type sensitive information into chatbots: account numbers, health details, complaints, sometimes data about their children. Two layers of risk follow.
The model layer. An academic analysis of six US developers’ privacy practices found that chat data was used for training by default, that sensitive information could be included, that some developers appeared to retain chats indefinitely, and that four appeared to include children’s chats in training. If your agent is built on a third-party model, you need to know which of those terms reach your customers’ conversations.
The disclosure layer. The OECD describes a transparency gap: users are often unaware that conversational data feeds model training. It recommends just-in-time disclosures at the moment data is collected, not terms buried in a privacy policy. For a customer service bot, that means a notice in the conversation itself, before sensitive data is requested.
Technical teams can reduce exposure by minimizing what the agent collects and by masking personal data before it reaches a model. This guide on handling personal data in AI covers the engineering side. The legal side is proving the controls work in live conversations, which is a testing question, not a policy question.
A note on privilege
Teams often ask whether conversations with a chatbot are privileged. In Heppner, a federal court found that a user’s consumer Claude chats were not protected by attorney-client privilege: the AI was not counsel, the terms were not confidential, and the use was not directed by counsel. Work-product protection failed for a similar reason, though counsel-directed AI use might qualify. For compliance teams, the lesson is that employee and customer chats with a general-purpose bot should be treated as discoverable records. Questions about a specific matter belong with your legal counsel.
Governance of staff use matters here too. Training employees on what may and may not be entered into AI tools is a sensible baseline, and this overview of AI employee training shows how such programs are usually structured.
What regulators expect you to be able to show
The common thread across the UK guidance, EU transparency rules and US enforcement theories is that good intentions do not count. Organizations are expected to test and monitor AI agents and to apply regular human oversight so that outputs comply with consumer law.
That expectation translates into five questions a regulator, plaintiff or enterprise customer can ask:
- What did the agent actually say to customers on a given date?
- Did it disclose that it was AI, at the right moment, in every channel?
- Did it make commitments the business would not honor?
- Did it hand off to a human when it should have?
- Can you prove the answers to the first four, and show the records were not altered?
Vendor dashboards typically report the vendor’s own metrics on the vendor’s own system. They rarely answer questions two through five from the customer’s point of view.
A practical control framework
The following steps reduce legal exposure for any customer-facing agent, regardless of vendor.
1. Inventory every agent and channel
List each AI agent, the channels it runs on (web chat, voice, email, ticketing), the model behind it, and the vendor. Shadow deployments by individual teams are common and are still your liability.
2. Map the rules to each deployment
For each agent, note which disclosure, privacy and sector rules apply: state laws for US customers, Article 50 for EU users, and sector duties in financial services, health or insurance. A US company serving EU customers has to run both maps.
3. Define what the agent may not do
Write explicit boundaries: no pricing exceptions, no legal or medical advice, no promises of outcomes, mandatory escalation triggers such as complaints, vulnerability signals and requests for a human. Boundaries that are not written cannot be tested.
4. Test as a customer would
Test through the real interface with realistic and adversarial conversations: requests for refunds outside policy, questions about whether the user is talking to a person, attempts to extract data or override instructions, and emotionally charged complaints. Repeat after every model, prompt or vendor change.
5. Keep tamper-evident records
Store the transcripts and findings so they can be produced later. Evidence that can be verified as unaltered carries more weight in a dispute than a screenshot.
6. Put the requirements in the vendor contract
Specify disclosure behavior, escalation behavior, data retention and training restrictions, audit rights and incident notification. Contract terms are only enforceable if you can show a breach, which returns to step 4.
7. Monitor continuously
Agents drift. A prompt edit, a new model version or a changed knowledge base can break a compliant flow. Continuous monitoring of live conversations catches what a point-in-time test misses.
How independent audit fits
Lexic Compass is Lexic’s AI Agent Audit platform. It assesses conversational AI agents from any vendor by interacting with them as real users would, and it looks for blind spots across Customer Experience, Security, Technical Reliability and Regulatory Compliance, including issues that traditional monitoring tools may miss. Findings come as evidence organizations can use to govern agents in production and to hold vendors accountable. Lexic uses third-party cryptographic hashing to verify evidence integrity and support traceability, which speaks directly to the fifth question regulators ask.
This maps onto the framework above. Steps 4, 5 and 7 are the audit and evidence work: testing the agent from outside, preserving verifiable records, and watching it over time.
Where Lexic fits
Lexic offers engagement in the formats that match how far along a deployment is:
- Flash Preview and Audit Sprint for audit-based engagements on a specific agent.
- Enterprise Continuous Trust for ongoing oversight as deployments scale.
- Lexic Pulse for analysis of 100% of customer conversations across calls, chats, emails and tickets, with sentiment, churn and opportunity insights for CX, Product, QA, Operations and other teams.
Teams ready to see how an audit would apply to their agents can start with Lexic Compass or review the audit and continuous monitoring options.
Our perspective
Most chatbot legal risk is discovered the wrong way: through a customer complaint, a regulator’s question or a dispute. By then the only question that matters is what the agent actually said, and many organizations cannot answer it with evidence.
We built Lexic Compass on a simple premise: the party that sells an AI agent should not be the only party that vouches for it. Independent assessment, run through the real customer interface, produces findings a compliance team can act on and a vendor can be held to. Our audits found that 85% of the agents assessed failed to escalate issues to a human representative appropriately. That is a behavior that no policy document reveals and that ordinary monitoring tends to miss.
As AI deployments scale across channels and vendors, oversight has to scale with them. Audit-based assurance and continuous monitoring give business, compliance and technical teams one shared record of how their agents behave, and a basis for deciding what to fix first.
Sergio Llorens
FAQ
Is a company liable if its chatbot gives a customer wrong information?
Yes, it can be. In Moffatt v. Air Canada, a tribunal held the airline responsible for its chatbot’s misleading answer. US law firms also describe misrepresentation, deceptive-practice, reliance and contract theories. Using a third-party vendor does not move the responsibility away from the business that deploys the bot.
Do US companies have to tell customers they are talking to an AI?
It depends on the state and the context. Utah, New Jersey and California each have different triggers, and general deceptive-practice laws apply in addition. The safest approach for a national business is a clear notice at the start of every conversation, in every channel.
Does the EU AI Act apply to a US company’s chatbot?
Article 50(1) obliges providers of AI systems that interact directly with people to inform them they are dealing with AI, unless it is obvious, and it applies from 2 August 2026. Companies serving EU users should confirm their role and scope with EU counsel and check the Commission’s guidance.
Are chatbot conversations legally privileged?
Not by default. In Heppner, a federal court found that consumer chatbot conversations were not protected by attorney-client privilege. Treat chat logs as potentially discoverable, and take privilege questions on a specific matter to your lawyer.
What evidence helps most if a chatbot is challenged?
Dated transcripts, records of disclosure and escalation behavior, and results of independent testing, stored in a way that shows they were not altered. Lexic Compass produces this kind of evidence for agents from any vendor.
Sources
- Moffatt v. Air Canada
- Complying with consumer law when using AI agents (GOV.UK)
- Your Chatbot Does Not Need Authority to Create an Obligation (Jones Walker)
- Who’s liable for AI hallucinations in customer service?
- Utah Chapter 77 Generative Artificial Intelligence
- AI disclosure laws on commercial chatbot interactions (DLA Piper)
- Transparency obligations under Article 50 of the AI Act (European Commission)
- AI Act Round-Up: Recent Developments (McCann FitzGerald)
- User Privacy and Large Language Models (arXiv)
- Enhancing user control and privacy protections in the context of AI (OECD)
- Heppner Shows Attorney-Client Privilege’s Fragility in AI Era (Bloomberg Law)
- PII Handling in AI
- AI Employee Training