A customer-service chatbot answers a pricing question. A hiring platform ranks applicants. A marketing team generates an image and tagline. An employee pastes a customer record into an AI assistant. These may look like separate productivity choices, but each can connect data, people, decisions, contracts, intellectual property, professional responsibility, cybersecurity, and insurance. The useful first question is not “Do we use AI?” It is “What does the tool do, what goes into it, who may rely on the output, and what have we promised?”

AI risk begins with the business use—not the product label

An AI feature may be obvious, such as a standalone chatbot, or nearly invisible inside a customer relationship platform, office suite, security product, applicant system, design tool, accounting workflow, or software update. A useful inventory records the use case, business owner, vendor, model or version, access method, input data, output, audience, human-review role, known limitations, and change history.

The National Institute of Standards and Technology describes its AI Risk Management Framework as voluntary and organizes the work around GOVERN, MAP, MEASURE, and MANAGE.[1] Its Generative AI Profile adds practical questions about provenance, known issues, human oversight, intellectual property, sensitive information, model versions, access modes, testing, monitoring, and incident response.[2] These resources can help a business ask better questions. They are not universal legal mandates, certifications, or insurance endorsements.

For technology-dependent operations, the agency’s technology business insurance overview can help frame property, liability, professional-service, cyber, equipment, and interruption questions. The description supplied to an insurance professional should match the real operation. “Office services” may not communicate that a public chatbot provides instructions or that an automated tool ranks applicants.

Customer-facing chatbots can create more than a technical problem

A chatbot can provide false, outdated, incomplete, inconsistent, or overly confident information. Risk increases when a person may rely on the answer for price, availability, eligibility, suitability, deadlines, warranties, safety instructions, professional advice, or contract terms. The issue may involve what the business knew about the limitation, how the tool was presented, whether a person was available, and how quickly the answer was corrected.

Oregon Department of Justice guidance says existing consumer-protection, privacy, information-security, and equality laws can apply to AI. It discusses developers, sellers, deployers, and customer-facing chatbots and notes that known patterns of false or misleading information may raise disclosure questions.[3] That is agency guidance applied to facts—not a statement that every chatbot error violates the law.

Federal Trade Commission matters offer another caution. The FTC’s final order involving DoNotPay addressed unsupported claims that an automated service could perform like a human lawyer.[14] That order binds the named respondent, not every business. The broader lesson is narrower: test material claims, preserve support, avoid presenting experimental output as verified expertise, and give customers a clear path to a person.

An inaccurate answer can raise different policy questions depending on the allegation. A technology or errors-and-omissions review may focus on the service promised, professional or technology definitions, contractual liability, claims-made dates, prior knowledge, defense, and reporting. A general liability discussion addresses different interests and should not be assumed to replace cyber, media, professional, employment, crime, or management-liability analysis.

Automated decisions deserve a higher-consequence review

Decision tools may screen, rank, score, recommend, profile, route, price, approve, deny, or flag. The tool may support a person or make the final action automatically. That distinction matters, but the phrase “human in the loop” is not enough. Record what the reviewer sees, whether relevant information is available, whether the reviewer understands the output, whether the reviewer has time and authority to disagree, and how an affected person can ask for correction.

Federal employment-discrimination law applies when employers use AI and other employment technology. The Equal Employment Opportunity Commission explains that both intentional discrimination and unjustifiable disparate impact may be relevant.[6] U.S. Department of Justice guidance says employers can remain responsible when third-party hiring technology unlawfully screens out qualified people with disabilities and emphasizes accessibility and reasonable accommodation.[7]

California’s employment automated-decision regulations took effect October 1, 2025 and address definitions, records, and evidence connected with anti-bias testing.[8] Separate California privacy regulations became effective January 1, 2026; their automated-decision requirements for covered businesses making significant decisions have a January 1, 2027 compliance date, subject to scope and exceptions.[4][5] They describe meaningful human involvement as more than a rubber stamp.

Oregon’s privacy analysis differs. Oregon DOJ says the Oregon Consumer Privacy Act generally excludes employment and commercial contexts, while other employment-discrimination, consumer-protection, security, contract, and professional rules may still apply.[3][15] A business should not copy a California checklist into Oregon or assume one state’s exception answers another state’s question.

A professional liability insurance review and the organization’s employment-practices discussion should use the actual workflow, affected group, vendor contract, testing, records, notices, explanation process, and complaint route. Coverage depends on the issued form. Employment practices, professional liability, cyber, D&O, and contractual obligations can overlap without being interchangeable.

Generated content raises copyright, trademark, likeness, and permission questions

An AI-generated advertisement can contain more than words. It may combine an image, style, slogan, product description, logo suggestion, music, voice, face, code, or design. A content review should therefore ask separate questions: Who supplied the source material? What did a person create or change? What do the vendor terms say? Was the brand name cleared? Is a recognizable person’s voice or likeness authorized? What records show the final approval?

The U.S. Copyright Office says copyright protects human-authored expression in AI-assisted work. Prompts alone generally do not provide enough expressive control under currently available technology, while human selection, arrangement, or modification may qualify case by case.[9] A vendor’s statement that the customer “owns the output” does not itself establish federal copyright protection or third-party clearance.

Names and logos require a different analysis. The U.S. Patent and Trademark Office recommends searching federal and state trademark databases and the internet because common-law rights may not appear in the federal database.[10] California also has statutes concerning commercial use of a person’s name, voice, signature, photograph, or likeness. The facts, consent, exceptions, contracts, and current law require qualified legal review.

A product liability review may become relevant when AI-supported content, instructions, design, labeling, or recommendations connect to a product allegation. Media, advertising injury, professional liability, technology E&O, cyber, general liability, product liability, and contractual indemnity can treat the same event differently. Preserve source, permission, editing, clearance, consent, and approval records without assuming those records guarantee a defense.

Privacy and cybersecurity start with what enters the prompt

A business may focus on what the model produces and overlook what employees send to it. Prompts can contain customer records, employee information, health details, financial data, legal strategy, source code, pricing, trade secrets, incident reports, photographs, or credentials. The tool may retain, log, reuse, transmit, or expose information through a model host, application programming interface, connected plugin, or subprocessor.

The FTC has warned AI companies to honor privacy and confidentiality commitments and not hide material training uses or changes behind vague disclosures.[11] Oregon DOJ likewise says existing privacy and information-security laws can apply to AI and discusses notice and consent questions for covered uses.[3] California’s privacy framework adds its own covered-business, consumer-rights, vendor-contract, security, risk-assessment, and automated-decision questions.[4][5][16] Applicability depends on entity, data, purpose, thresholds, exemptions, and actual processing.

CISA’s voluntary AI data-security guidance recommends practices such as provenance, classification, least privilege, encryption, integrity controls, vendor diligence, multifactor authentication, logging, input validation, testing, and rollback planning.[12][13] These practices can improve resilience but do not create coverage, satisfy every law, or replace requirements in a cyber application or policy.

A cyber liability insurance review can examine privacy response, forensics, restoration, cyber extortion, system failure, vendor events, dependent interruption, social engineering, regulatory matters, retentions, sublimits, exclusions, and security conditions. Ask how the form defines covered data, system, privacy event, security failure, vendor, and interruption. Do not wait for a breach to learn whether costs require consent, an approved panel, or prompt notice.

Vendor contracts and insurance applications should tell the same story

AI vendors may limit warranties, cap liability, disclaim output accuracy, reserve training rights, use subprocessors, change models, or require the customer to indemnify the vendor for prompts and outputs. A contract can allocate risk between parties, but it does not erase consumer-protection, privacy, employment, professional, or third-party claims. Nor does an indemnity promise prove that an insurance policy will fund that promise.

Review training and retention settings; subprocessor disclosure; data location and access; security obligations; incident timing; audit and cooperation rights; deletion and return; service levels; model changes; testing; intellectual-property allocation; defense and indemnity; limitation of liability; insurance requirements; and exit assistance. Match those facts to the insurance submission. If the business promises a client that no data will be retained while the vendor contract allows retention for product improvement, the operational story is already inconsistent.

A business owners policy review can help organize core property and liability questions for an eligible business, but a BOP should not be assumed to address AI-related professional services, privacy, employment, media, crime, or vendor interruption. A coordinated program may involve several policies, endorsements, and contracts.

Insurance for AI use is a coordinated set of questions

There may not be one generic “AI insurance policy” that answers every use. The better approach is to identify the potential allegation and the policy that might be expected to address it, then read the form. A chatbot error, privacy event, discriminatory hiring allegation, copyright complaint, fraudulent payment instruction, regulatory inquiry, vendor outage, and director oversight claim do not present the same facts.

The agency’s directors and officers insurance resource can support governance and management-liability questions, including how leaders approved, monitored, documented, and responded to a material AI use. A commercial umbrella review may address higher-limit questions over qualifying underlying policies. An umbrella cannot be assumed to follow every cyber, E&O, employment, media, or contractual exposure.

Build a pre-rollout and renewal checklist

A documented AI process can be proportionate. An internal drafting assistant with no sensitive data and mandatory review may not need the same controls as a public chatbot or automated hiring tool. The point is to identify consequence, data, reliance, and change before deciding the review depth.

Prepare the incident path before an output causes a complaint

An AI incident may begin as a strange answer, customer complaint, rejected applicant, leaked prompt, suspicious payment request, takedown notice, vendor alert, inaccurate report, model outage, or regulator inquiry. The first operational step is not to debate which policy applies. Address immediate safety and containment, preserve the evidence, restrict or pause the affected workflow when appropriate, and notify the organization’s legal, privacy, security, management, vendor, and insurance contacts according to the facts.

Preserve relevant prompts and inputs, outputs, model and version information, configuration, logs, notices, source and permission records, tests, approvals, complaints, vendor terms, customer contracts, corrective actions, and communications. Avoid silently deleting or rewriting the record. Insurance forms may require prompt notice, consent before costs are incurred, cooperation, approved providers, or reporting of circumstances even before a formal lawsuit.

Review incident duties under the applicable law rather than relying on a universal deadline. Oregon and California have different breach frameworks, and other federal, state, contractual, sector-specific, or international requirements may also apply. Determine whether the event is a reportable breach, employment matter, consumer complaint, professional issue, intellectual-property claim, contract dispute, regulatory matter, crime, or several of these with qualified advisers.

Turn the AI inventory into an insurance-review agenda

Bring more than the product name. Useful materials may include the AI-use inventory, data-flow diagram, vendor terms, customer contracts, privacy notices, employee policy, decision and human-review procedure, test results, complaint records, content-approval files, incident plan, insurance applications, declarations, forms, endorsements, and current limits.

Describe the operation accurately. If a consultant uses AI to produce client recommendations, a retailer uses it to answer product questions, a manufacturer uses it to change instructions, or an employer uses it to screen applicants, those are different use cases. The insurance discussion may include cyber and privacy, technology E&O, professional liability, general liability, product liability, employment practices, media or advertising injury, crime, D&O, property, business income, and umbrella or excess. None should be assumed to respond without reviewing the issued wording and facts.

Frequently asked questions

Sources and further reading

This article uses official government, regulator, standards, and public-agency resources available on September 29, 2026. Laws, regulations, guidance, technology, and policy forms can change. Confirm current requirements and obtain advice for the organization, jurisdiction, profession, data, decision, and event involved.