On this pageJump to a section
Connecting your CRM to an AI tool takes about ninety seconds. Deciding whether you should takes longer, and that is the part most people skip.
The setup screen says "Allow access." It looks like every other permissions prompt. But when the thing you are allowing is your customer list, your inbox or your intake forms, you are not changing a preference. You are making a decision about other people's information. This is educational, not legal advice. If you handle health or financial data, take it to counsel before you connect anything.
Why is connecting a tool a data-sharing decision?
Because the connection moves other people's information.
A modern AI tool does not only answer questions. It can be given access to your files, mail and customer records so it can work on them. The industry has moved quickly toward standard ways of doing that (I covered the basics in Claude Integrations): when Anthropic donated the Model Context Protocol to the Agentic AI Foundation in December 2025, it called MCP the open standard that links AI applications to outside systems and cited more than 10,000 active public MCP servers.1 More connection makes tools more useful. It also means more doors, and each door needs someone who decided it should exist.
What the framework asks
NIST's AI Risk Management Framework is voluntary, and it is organized around four functions: Govern, Map, Measure and Manage.2 You do not need the whole document. Map is the useful one here: before you can manage a risk, you have to know where your data goes.
What should you ask before you connect?
Three questions. Write the answers down.
- What goes inExact fields, records and folders
- Where it goesRetention, access and training use
- What it can doRead only first, then widen one step at a time
- Decision recordTool, data, terms, owner and review date
1. What goes in?
List the exact data the tool will receive. "My CRM" is not an answer. Which fields? Which records? Does it include notes? Attachments? Past conversations? For an inbox, does it read everything, or only a labeled folder?
2. Where does it go?
Find out what the vendor does with the data: how long they keep it, who can access it, whether it is used to train models, and whether you can switch that off. The FTC has warned AI companies that promises about customer data are binding, wherever they are made, including terms of service and marketing materials, and that using data for new purposes without clear notice and consent is risky.4 Read the actual terms you are accepting, not just the summary.
3. What can the tool do?
Read-only is very different from read and write. A tool that can only summarize your inbox is a smaller risk than one that can send mail, delete files or change records. Start with the narrowest permissions that do the job, and widen them one at a time.
What customer data should stay out?
Sort it by sensitivity, not by convenience.
| Usually reasonable to share with an approved tool | Keep out of general-purpose tools |
|---|---|
| Names and business emails for routine follow-up | Passwords, API keys and login links |
| Order history and public product questions | Payment card data and bank details |
| Draft marketing copy built from public information | Government identifiers |
| Anonymized or redacted examples | Health, medical or intake information |
| Notes you would be comfortable showing the customer | Anything covered by a confidentiality agreement |
Redact first when you can
Often the tool does not need the identity. Replace names, account numbers and contact details with placeholders before you paste a message or a record. You get the same help without handing over the person.
Use a sandbox account
Test a new connection with dummy data or a throwaway account before you point it at the real thing. The first connection is when you learn what the tool actually reads, and you want that lesson to cost nothing.
What about health information?
Treat it as its own category, with its own rules.
Health information carries extra legal obligations. If you are in healthcare, telehealth or wellness, the question is not "is this tool good?" but "am I allowed to send this data to this vendor, under this agreement, for this purpose?" That depends on your role under HIPAA and state law and on what the vendor has signed. I cannot answer that for you, and neither can a blog post. Ask counsel. The same caution applies to tracking pixels.
The FTC rule many small businesses miss
Not every health-related business is covered by HIPAA. The FTC's Health Breach Notification Rule applies to vendors of personal health records and related entities that are not covered by HIPAA. In April 2024 the FTC finalized changes clarifying that a breach of security includes an unauthorized acquisition of identifiable health information, whether it comes from a data security breach or an unauthorized disclosure, and that related entities offering products or services through a vendor's online services, including mobile apps, can be covered.3 The practical meaning is that sharing health data in a way customers did not authorize can create obligations, even if no hacker is involved.
How do you document the choice?
One page per connection. Keep it where you can find it, next to your AI policy.
- Name the tool and the date you connected it, and who approved it.
- List the data it can access, field by field or folder by folder.
- Record the permissions granted: read, write, send, delete.
- Note the vendor terms: retention, training use and whether you opted out where possible. Save a copy of the terms or a link with the date.
- State the purpose, in a sentence. If you cannot, you may not need the connection.
- Set a review date, and review sooner if the vendor changes its terms.
Why the record matters
When a customer asks, or a regulator does, "what happened to my data?", a dated record is the difference between a calm answer and a scramble. It also forces you to notice when a connection has grown past what you meant to allow.
What do you tell customers?
Whatever is true, in your privacy policy and in plain language.
The FTC's warning is specific: surreptitiously changing a privacy policy or terms to use data in new ways is risky, and so is burying disclosures in fine print.4 The same logic applies to any business. If you tell customers you do not share their data with third parties, do not then connect their records to a tool that does. If your practices change, update your notice before you change the practice.
The decision is yours
I work with AI agents every day, including in the system I build for clients, so this is not an argument against using them. It is an argument for deciding on purpose. Every connection is a decision about someone else's information, and the person it belongs to is not in the room.
Convenience is a bad reason to hand over something that is not yours.
Keep building.Ekai
Questions people ask
Is it safe to connect my inbox or CRM to an AI tool?
It depends on what the tool can read, where the data goes and what you have promised your customers. Treat the connection as a data-sharing decision: define the data, the tool's permissions and the terms you have accepted before you switch it on.
Can I put customer health information into an AI tool?
Be very cautious. Health information carries extra legal obligations, and whether a particular tool may receive it depends on your role under HIPAA or other rules and on the vendor's agreement. Keep it out of general-purpose tools and ask counsel before sharing it.
What is the FTC Health Breach Notification Rule?
It requires vendors of personal health records and related entities that are not covered by HIPAA to notify consumers and the FTC after a breach. The FTC's April 2024 final changes clarified that a breach includes unauthorized acquisition of identifiable health information from an unauthorized disclosure, not only from a data security breach, and extended coverage to related entities that offer products or services through a vendor's online services, including mobile apps.
Do I need to tell customers I use AI?
If your privacy policy or promises say something about how data is used or shared, they must stay accurate. The FTC has warned that quietly changing terms or using data for new purposes without clear notice is risky. Update your notices when your practices change.
Sources
- 1
- 2
- 3
- 4
