Who is responsible when an AI answers your customer wrongly

The contracts that govern AI messaging tools are published, so here is what they allocate. Meta’s WhatsApp Business Solution Terms carry one sentence that decides most of this question: the business carries its AI vendor’s acts, not the vendor.

This is a reading of published terms and not legal advice. We have not had any of these documents reviewed by a lawyer, we say nothing about how they would apply to you, and you should take your own advice before relying on any of it.

The sentence

Hiring someone is expressly permitted: “You may retain a third party to manage your access to the WhatsApp Business Solution”. The terms call that party a Third Party Service Provider, require it to agree in writing to handle your data only “on your behalf, pursuant to your instructions and authorization”, and then place the risk:

You are solely and fully liable for all acts and omissions by your Third Party Service Providers, and any breach of these WhatsApp Business Solution Terms, the Business Terms, and other terms referenced or incorporated herein, or therein by your Third Party Service Provider, shall be deemed a breach by you, with you being liable with respect to such breach as if you yourself had breached these WhatsApp Business Solution Terms, the Business Terms, and other terms referenced or incorporated herein or therein.

Two mechanisms sit in one sentence: liability for the provider’s acts and omissions, and a deeming rule that turns the provider’s breach into yours. An indemnity follows immediately, running from you to Meta and covering “all liabilities, damages, losses, and expenses of any kind (including reasonable legal fees and costs) relating to, arising out of, or in any way in connection with any claim related to the acts or omissions of your Third Party Service Providers”.

An AI messaging tool on WhatsApp is normally in exactly that role, because the same terms give it no other one. Their AI Providers section bars AI companies from offering AI as the primary functionality, then adds that “you may retain an AI Provider as your Third Party Service Provider”. The arrangement the terms permit is the arrangement whose failures they assign to you.

What each document makes you answerable for

Document, and on what	The allocation, in its own words
Solution Terms, on your provider	“You are solely and fully liable for all acts and omissions by your Third Party Service Providers”
Business Terms, on your account	“Company is responsible for all activities occurring under its account.”
Business Terms, on your conduct	“Company is solely responsible for its conduct while using our Business Services”
Messaging Policy, on selling	“You are solely responsible for your transactions and providing any sales terms, privacy terms or other terms applicable to your interactions with users.”
Google, on profile content	“These policies apply to all of your Business Profile content.”

One detail in the Business Terms is easy to miss. Automation is not treated as a separate actor: each list of prohibitions opens “Company must not directly, indirectly, or through automated or other means”. A machine acting is the company acting.

Meta’s own side is stated in capitals. On control: “WE ARE NOT RESPONSIBLE FOR AND ARE NOT OBLIGATED TO CONTROL THE ACTIONS OR INFORMATION (INCLUDING CONTENT) OF OUR USERS OR OTHER THIRD PARTIES.” On money, an aggregate cap that “WILL NOT EXCEED THE GREATER OF ONE HUNDRED DOLLARS ($100) OR THE AMOUNT COMPANY HAS PAID US IN THE PAST TWELVE MONTHS TO USE OUR BUSINESS SERVICES”.

Automation is allowed, with an exit attached

The Business Messaging Policy permits a machine to answer inside the reply window, and attaches a condition:

You may use automation when responding during the 24-hour window, but must also have available prompt, clear, and direct escalation paths.

It then enumerates what counts, which makes the condition a specification rather than advice: “In-Chat Human Agent transfer”, a phone number, an email address, “Web support (on the business website)”, in-store visits, or a support form.

The same policy states what failure costs: “If content or actions violate our terms or policies, we may limit or remove your access to WhatsApp Business Services.” Access belongs to the account, so the penalty lands on the account holder whoever, or whatever, composed the message.

Google attributes the reply to the business

Google’s review documents reach the same place in different vocabulary. Its Maps User Generated Content Policy covers machine output by name: “These policies and standards apply to all Maps user generated content, regardless of whether the content is human-generated or AI-generated content”, and attaches the duty to whoever publishes, since “Any content you publish to Google platforms must follow these policies”.

Then attribution. Google screens replies before they appear, and its help page for managing reviews says what a published one looks like: “It will appear like your business replied, and your personal name won’t be shown.” Neither the author nor the software is visible to the reader. That screening is the same machinery that rejects a reply and exposes a reason, worth knowing before you automate an answer to an unhappy customer.

Where these documents say nothing

• Nothing apportions anything between you and your provider. The deeming clause makes the provider’s breach yours as against Meta. What you can recover from the provider is left to whatever you signed with it, and none of these documents mentions that contract.
• None of them addresses an answer that is merely wrong. The accuracy rules they do carry are about profile information, not replies, and the nearest conduct rule is about deception rather than error: “Do not confuse, deceive, defraud, mislead, spam, or surprise people with your communications.”
• Google names a vendor once, about menus. In the whole guidelines page the only third-party clause reads: “Third parties that manage Business Profiles on behalf of clients must notify and have the consent of the business owner to submit a menu URL for a business.” There is no general clause about software acting on a profile owner’s behalf.
• The Solution Terms carry no disclaimer or liability cap of their own. They incorporate the Business Terms of Service, where both live.

Our position, marked as ours

Everything above is quotation. This part is argument.

If the platforms attribute the reply to the account and make the account answerable for the provider it retained, the sharp product question is not how good the model is. It is whether anything reaches a customer without the owner seeing it. We read that as the case for drafting rather than autonomy: the software writes, a person releases, and accountability stays where the contracts already put it. That is an opinion, held by a company that would itself be the Third Party Service Provider in the clause above, so discount it and read the terms yourself.

Nor is it proven. No channel is connected to mesej, so nothing here has ever answered a real customer, and the gate we intend to put in front of publishing is a commitment rather than a measurement. Choosing between the tools that exist is a separate question from this one.

The checkable answer is short. Meta’s terms name the business twice over, for its provider’s acts and for its own account. Google’s policies attach to content published under the profile, cover machine-written content by name, and show the reply as the business speaking. What none of them settles is what you can recover from your vendor, which is the part you negotiate.