The Customer Privacy Promise When You Use AI

Why this matters

Most small shops have a privacy page that somebody copied from a template years ago, and somewhere in it is a sentence like "we never share your information with third parties." Then the office starts pasting job notes into a chat assistant, the phones go to an AI answering service that records every call, and that sentence quietly stops being true. Nobody lied on purpose. But a privacy statement is a factual claim to the public, and a claim that no longer matches what the shop does is the thing regulators and plaintiffs' lawyers look for first.

This card covers the promise itself: what the notice has to say once customer data touches an AI tool, and when the state privacy laws start to matter as the shop grows. What may and may not go into a tool in the first place is a separate question, owned by the shop's AI use policy. See related: The Shop AI Use Policy SOP.

The promise is a statement of fact, and the federal floor treats it that way

There is no comprehensive federal privacy law in the United States. What there is, at the federal level, is Section 5 of the FTC Act, which prohibits deceptive practices in commerce, and the FTC has applied it to privacy statements for decades: if your notice says you do a thing and you do not, that gap can be a deceptive practice regardless of whether any harm followed. The FTC has also said publicly that quietly changing data practices to feed AI tools, without telling people, is the kind of conduct it watches. State consumer-protection statutes (each state has its own unfair-and-deceptive-practices act) sit on top of that floor and often give customers a private right to sue, which the FTC Act does not.

So the practical rule is simple and runs in one direction: you are not required to say much, but everything you say has to be true. A short, accurate notice is safer than a long, aspirational one. The template line "we never share your data" is the most common false statement in a small shop's paperwork, because every shop shares data with its field software vendor, its payment processor, its email provider and its accountant. AI tools just add one more recipient, and usually the one the owner has thought about least.

What actually happens to text you put into an AI tool

Before you can describe your practice, you have to know it, and the answer lives in each tool's terms of service and data-processing terms, not in its marketing page. Four questions settle almost everything:

• Retention. How long does the vendor keep what you typed or the audio it recorded? Days, months, or until you delete it?
• Training. May the vendor use your inputs to improve its models? Consumer and free tiers commonly allow it by default; business tiers commonly do not, or let you switch it off. This is the single most consequential setting.
• Human review. Can the vendor's staff read conversations, for abuse monitoring or quality? Many say yes for a limited window.
• Subprocessors. Who else touches it? An AI phone service often runs on top of someone else's speech and language models, so your customer's voice may pass through two or three companies.

These terms change, sometimes several times a year, and the setting you checked in spring may have a new default by fall. Everything in this section is current as of 2026; re-read each vendor's data terms at renewal and whenever the vendor emails you about a policy update, because that email is usually the change.

The four things the notice has to cover

A notice that is honest about AI does not need to name products. It needs to tell a customer four things in plain words:

1. What kinds of information go into automated tools. Call audio, the text of messages, job notes, names and addresses. Say the categories.
2. What the tools are used for. Answering and routing calls, drafting messages that a person reviews, cleaning up invoice descriptions, scheduling. The purpose is what a customer actually cares about.
3. Whether the vendors may use the information for their own purposes, including training their models. If you have turned training off, say so. If you have not checked, you cannot say anything here yet, which is the point of the audit below.
4. How a customer can ask a question, ask for a human, or ask you to delete what you hold. A real address or phone line, not a form that goes nowhere.

What the notice should not do: promise a security outcome ("your data is completely safe"), promise something a vendor controls and you do not ("your calls are never stored"), or bury the AI use in paragraph eleven of a legal page nobody reads. If the phone line is answered by an automated agent, the notice is the backstop, not the disclosure. The disclosure belongs at the start of the call. See related: Telling Customers When AI Is Involved.

The notice, written plainly

This is the shape that works for a residential shop. It fits on one screen and every sentence is something the shop can verify about itself.

HOW WE USE YOUR INFORMATION

We use your name, address, phone, email and job
details to schedule, do and bill your work.

Some of our tools are automated. Calls to our main
line may be answered and recorded by an automated
assistant, and staff use writing tools to draft
messages and invoices. A person reviews anything
we send you.

The companies behind these tools process your
information for us. We have set them so they do
not use it to train their own systems.

We do not sell your information.

Questions, or want a person instead of the
assistant? Call the office or write to us. You can
ask what we hold about you and ask us to delete it.

Last reviewed: March 2026

Two lines in it are doing more work than they look. "We have set them so they do not use it to train" is only writable after someone has checked the training setting in every tool, and it has to be re-checked when vendors change defaults. "Last reviewed" is a commitment: a date more than a year old tells a careful reader the shop stopped looking.

Worked example: a fourteen-person plumbing shop

Say the shop runs eight trucks, has about 6,000 customer records accumulated over a decade, and completes roughly 1,800 jobs a year. Its privacy page, copied from a template in an earlier year, says "we do not share your personal information with any third party."

The audit. The office manager lists every place customer information goes and finds three AI uses that arrived over the past year without anyone deciding on them as a set:

What the audit actually found. The old notice was already false before AI arrived, since the field software, payment processor and email provider are all third parties. AI made it worse in one specific place: two office staff on free personal accounts, where the shop has no contract at all and the vendor may train on the input. One of the job notes they pasted said a customer was on home oxygen and the side gate code. That is the finding that matters, and it is a use-policy problem first (those accounts should not exist) and a notice problem second.

The order of fixes. First, move the two staff onto the business tier with training off, so the practice becomes describable. Second, switch the phone service's improvement setting off. Third, rewrite the notice in the shape above. Doing it in the reverse order produces a notice that is true on paper and false the day it is published, because the free accounts are still in use.

What would change the answer. If the shop had found it could not turn training off on the phone service, the honest notice says so ("the company that runs our answering service may use recordings to improve its service") and the owner decides whether that vendor stays. Writing the comfortable version and hoping nobody checks is the one option that is not available.

When the state privacy laws start to apply

As of 2026 roughly twenty states have comprehensive consumer privacy laws, and the list grows most years. Nearly all of them gate on size, and a residential shop is usually below the gate, but the gate is not revenue alone, so check rather than assume.

• California (the CCPA as amended by the CPRA, Cal. Civ. Code 1798.100 and following) applies to a business that meets any one of three tests: annual gross revenue above a threshold the statute sets and adjusts for inflation; buying, selling or sharing the personal information of 100,000 or more California consumers or households a year; or deriving 50 percent or more of revenue from selling or sharing personal information. "Sharing" in California includes passing data for cross-context behavioral advertising, which is why a website ad pixel can matter even for a shop that never sells a list.
• Virginia, Colorado and most of the states that followed use a count of consumers whose data you control or process, commonly 100,000 a year, or a lower count (commonly 25,000) combined with revenue or benefit from selling data.
• Texas has no numeric threshold but exempts businesses that qualify as small under the SBA's size standards, with one exception that survives the exemption: a small business still may not sell sensitive personal data without the person's consent.

On the worked example's numbers, 6,000 customer records sit far below any 100,000 count, and the shop does not sell data. It is very likely outside every comprehensive state law today. What changes that: a growing website audience tracked by advertising tools, an acquisition that merges two customer lists, or expanding into a state with a lower count. Re-check when any of those happens, and re-check the list of states yearly, since the thresholds are set state by state and some have been amended downward.

Two obligations apply regardless of size. Every state has a data-breach notification law, so an AI vendor breach that exposes your customers' information can trigger notice duties even for a five-person shop, on deadlines that vary by state. And the deception rule above applies to every business that publishes a notice at all.

Where to stop and call a lawyer

This is not legal advice, and the fork is clean enough to name. Handle the audit and the plain notice yourself. Stop and talk to a privacy attorney before acting when any of these is true: you think you may be near a state threshold; a vendor's terms say it trains on your data and you want to keep using it; you hold health, financial or children's information in any volume; or you learn a vendor has had a breach involving your customers. Walk in with the audit table above, each vendor's current data terms printed with their date, your current notice, and your customer and website-visitor counts by state. That turns a general consultation into a short, specific one.

References

• FTC Act Section 5 (15 U.S.C. 45), deceptive practices, and FTC business guidance on privacy statements and AI
• California Consumer Privacy Act as amended by the California Privacy Rights Act (Cal. Civ. Code 1798.100 and following); state privacy laws as enacted in Virginia, Colorado and Texas
• See related: The Shop AI Use Policy SOP
• See related: What Not to Paste Into an AI Tool
• See related: Telling Customers When AI Is Involved