Using AI with customer data: a privacy checklist

AI tools are most useful when they can see your real data: the lead’s history, the deal notes, the last three texts. That’s also what makes them a privacy question. Customer names, phone numbers, call recordings and invoice details are exactly the information your customers trusted you to look after.

This checklist is for owners and sales managers at small businesses who want to use AI with customer data without creating a problem they discover later. It covers the questions to ask vendors, how to set permissions, what to stop collecting and what to tell customers. This is general information, not legal advice. Privacy laws differ by state and country, so check the rules that apply to your business.

Why this matters now

US regulators have been clear that AI doesn’t come with special rules that let you skip the old ones. In a January 2024 post, staff in the FTC’s Office of Technology wrote that “there is no AI exemption from the laws on the books,” and that AI companies that fail to keep their privacy commitments to users and customers may be liable under laws the FTC enforces (FTC, 2024). The same post notes the FTC has previously required companies that unlawfully obtained consumer data to delete products built with it, including models and algorithms.

A month later, FTC staff warned that it “may be unfair or deceptive for a company to adopt more permissive data practices,” such as sharing data with third parties or using it for AI training, and to tell users about the change only through a hidden, retroactive update to its terms or privacy policy (FTC, February 2024).

Those posts are aimed mainly at AI companies, but they matter to you twice: once when you choose a vendor, and again because the same principle applies to your own privacy promises to customers.

Questions to ask AI vendors

Get answers in writing, ideally in the contract or data processing terms rather than a sales email.

Question	What a good answer looks like
Do you use our data, or our customers’ data, to train or improve models?	A clear yes or no, with any opt-out spelled out
Which outside AI model providers process our data?	A named list of subprocessors, with notice before changes
How long are prompts, outputs, recordings and transcripts kept?	Specific retention periods, and whether you can shorten them
Can we delete data, and what happens to it when we cancel?	A deletion process with a timeframe
Does the AI see everything, or only what each user is allowed to see?	Access follows the user’s role and permissions
Can the AI change records or contact customers by itself?	Actions require a person to confirm them
Is there a log of what the AI and users did?	Audit logs you can review
How are credentials and integration keys protected?	Encrypted at rest; keys hashed or otherwise protected
Will you tell us before changing these terms?	Advance notice, not a silent update

The FTC’s general guide for businesses, Protecting Personal Information, adds a practical point: put your security expectations in writing in contracts with service providers, and verify they’re being met rather than taking their word for it.

NIST’s voluntary AI Risk Management Framework (AI RMF 1.0, January 2023) makes the same point in its own language: organizations should have policies for AI risks associated with third-party entities (GOVERN 6.1). For a small team, that just means knowing who touches your data and on what terms.

Permissions: least privilege for AI

The FTC guide recommends the “principle of least privilege”: give each employee access only to what their job requires. Apply the same rule to AI.

• The AI should inherit the user’s access. If a rep can only see their own leads, an AI assistant they use shouldn’t answer questions about everyone else’s.
• Separate reading from acting. Letting AI summarize a deal is low risk. Letting it send messages, change records or create automations needs a confirmation step. We cover this in why your AI assistant should ask before it acts.
• Limit who can configure AI. Prompts, voice agent scripts and automations should be editable by owners or admins, not everyone.
• Offboard fully. When someone leaves, remove their access to the CRM and any AI tools connected to it on the same day.

Our guide to CRM roles and permissions walks through how to design roles for a small team.

Minimization: collect and keep less

The FTC guide’s five principles start with “take stock” (know what personal information you have) and “scale down” (keep only what you need for your business). Every field you don’t store is one less thing an AI tool, or anyone else, can expose.

• Audit your custom fields. Do you need date of birth, or just whether someone is a homeowner? Do you need a full ID number, ever?
• Keep payment card details out of the CRM. The FTC guide specifically warns against keeping card numbers and expiration dates without an essential business need.
• Don’t paste sensitive data into general chatbots. A staff rule like “no customer phone numbers, addresses or financial details in personal AI accounts” prevents the most common leak.
• Set retention periods. Decide how long you keep call recordings, transcripts and closed-lost leads, write it down and follow it. The FTC guide recommends a written records retention policy.
• Dispose properly. Deleting old exports and CSV files from laptops and shared drives counts.

NIST’s 2024 Generative AI Profile lists data privacy among the risks specific to or made worse by generative AI, including leakage and unauthorized use or disclosure of personal information. Minimization is the cheapest control you have against it.

Disclosure to customers

Customers should not be surprised by how you use their information.

• Update your privacy notice. Say that you use service providers, including AI services, to manage customer relationships, and describe what for. Then make sure your practice matches the words.
• Disclose AI on the phone. If an AI voice agent answers your calls, say so in the greeting. Our buyer’s guide to AI voice agents covers the FCC rules for AI voices on outbound calls.
• Disclose recording. According to the Reporters Committee for Freedom of the Press, about 11 states primarily require every party’s consent to record a conversation. Our post on AI call summaries summarizes the one-party vs all-party rules.
• Respect channel consent. Consent to email isn’t consent to text or to AI calls. Track it per channel.
• Don’t change the deal quietly. If you start using customer data in a new way, tell people first. That’s the lesson of the FTC’s February 2024 warning, and it applies to small businesses as much as AI companies.

A one-page checklist

Use this as a quarterly review:

1. We have written answers from every AI vendor on training, retention, subprocessors and deletion.
2. AI tools see only what each user’s role allows.
3. AI can’t send messages, change records or contact customers without a person confirming.
4. We review audit logs for unusual activity.
5. We’ve removed fields and files we don’t need, and we follow a written retention policy.
6. Staff know what not to paste into personal AI tools.
7. Our privacy notice describes our use of AI services accurately.
8. AI-answered and recorded calls are disclosed to callers.
9. Consent is tracked per channel, with evidence.
10. We have a plan for what to do if customer data is exposed.

The last item is the FTC guide’s fifth principle, “plan ahead”: create a plan to respond to security incidents before you need one.

Where Exa fits

Exa’s AI assistant respects each user’s role and permissions, and it creates tasks, notes, leads, messages, appointments and automations only after the user confirms an action card. Roles run from owner and admin to manager, rep and viewer, and reps see only records they own or that are unassigned. Audit logs record creates, updates and deletes with old and new values, the user, IP address and user agent. Integration credentials, calendar tokens and webhook secrets are encrypted at rest, API keys are stored hashed, and consent is stored per channel with its source and evidence. You can read more on the security page and the AI page. Ask us, and every other vendor you use, the questions in the table above.

Final thoughts

Using AI with customer data is less about new rules than about applying old ones with discipline: know what you have, keep less of it, limit who and what can see it, get vendor promises in writing and don’t surprise your customers. Do those five things and AI becomes a tool your customers never have a reason to worry about.