The Three-Bin Data Sort: an AI data privacy checklist for service businesses

This is the one-page sort from the video, plus the questions to put to any AI vendor before your customer records go anywhere. You can fill it in at a desk in an afternoon. You do not need a server.

The three bins

1. Bin one: low risk. Marketing copy, your public price ranges, blog drafts, job descriptions. Any tool on your approved list is fine.
2. Bin two: customer and job data. Names, addresses, phone numbers, repair orders, estimates, job notes. This goes to cloud AI under business terms, or to a hybrid setup, with limited access.
3. Bin three: money and identity. Payment card details, bank records, anything that would trigger a breach notice. It stays off AI tools unless there is a clear need, and then it runs only on systems you control.

And the rule over all three: running everything on your own machine does not decide who sees what. Every agent still needs narrow access and a log.

Why sort the data first

The usual trigger for this conversation is an uncomfortable moment. In 2023 Samsung engineers uploaded internal source code to ChatGPT, and the company then restricted staff use of generative AI, worried that data sent to outside platforms may sit on servers where it is hard to retrieve or delete. In a service business the small version looks like an estimator pasting a customer's insurance details into a free tool to get a letter written.

The FTC's guide to protecting personal information, as summarized by the law firm Manatt, starts with two habits: take stock of the personal information you hold, and scale down to what the business needs. The three bins are those two habits applied to AI. Once each kind of data has a bin, choosing tools becomes a short conversation instead of a debate.

How to fill in the sort

Open a sheet with these columns and list every kind of record your office touches in a week:

Data	Where it lives today	Bin	Allowed AI route	Who may use it
Website and ad copy	Website, ad accounts	1	Any approved tool	Anyone
Customer name, address, phone	Field service software	2	Cloud AI under business terms, or local	Office staff, the intake agent
Job notes and estimates	Field service software	2	Cloud AI under business terms, or local	Office, the follow-up agent
Card and bank details	Payment processor, accounting	3	None by default	Owner, bookkeeper
Employee records	Payroll	3	None by default	Owner

The rows are an example layout. Your own list will be longer.

Three steps:

1. List the records, not the software. "Customer phone numbers," not "our CRM."
2. Give each one a bin. When in doubt, pick the higher bin.
3. Write the allowed route and the people. That column is your AI policy. Print it and put it by the office printer.

What each bin allows

Bin one needs a list of approved tools and nothing more. Pick two or three, tell the team, and stop people from signing up for a new one every week.

Bin two needs business terms you have read. Providers' API and business terms differ from their free consumer apps. OpenAI's API documentation, for example, says data sent to the API is not used to train its models unless you opt in, and abuse monitoring logs are kept for up to 30 days, with zero data retention available only to approved customers. Other providers have their own terms. Read the ones for the tool you actually use, and keep a copy with the date.

Bin two can also run locally. Ollama, a common way to run models on your own hardware, states that it does not see your prompts or data when you run it locally. That is the hybrid answer most service businesses land on: routine, every-night work on hardware you control, harder work in the cloud. Our private AI page describes how Cortex sets that up, and the Fifty Example Test shows how to check which jobs a local model can handle.

Bin three stays out by default. If a real job needs it, such as reconciling payments, keep that job on systems you control, give the agent read-only access to exactly the records it needs, and put a person on every action.

Hosting doesn't decide who sees what

This is the part people skip after buying hardware. A local box with a shared password and no log is not private. Whatever bin and whatever host, every agent gets:

• Narrow access. It reads only the systems its job needs.
• Approval on sensitive actions. Sending, discounting and anything that spends money stays with a person. Cortex's defaults are on our trust and security page.
• A log. What it read, what it drafted, and who approved it.
• An off switch. Pause the agent or pull a permission without calling anyone.

NIST's AI Risk Management Framework is voluntary and organizes this work into Govern, Map, Measure and Manage. A small company does not need the whole document, but the four words make a good quarterly check: who owns it, what data it touches, whether it behaved, and what you changed.

The vendor questions

Send these in writing to any AI vendor or implementer before bin two data goes to them, and keep the answers:

1. Where does my data go, and which models see it?
2. Is it used to train anything? Put the answer in writing.
3. How long is it kept, and who on your side can see it?
4. Can each agent be limited to the records its job needs?
5. Is every action logged, and can I see the log?
6. Can I run part of this locally later without a rebuild?
7. Can you show me a real deployment, and what stayed on site?

Cortex answers those for its own work: Detail Driven, its named deployment, runs dedicated Linux hardware on site with hybrid local and cloud models, and Tekmetric and CCC ONE stay the systems of record, as described on the private AI page.

What to watch out for

• Sorting vendors before data. You end up debating brands instead of deciding what may leave the building.
• Free consumer apps for bin two. Their terms are not your business terms.
• One shared login. It makes the log useless.
• A policy nobody sees. If the sheet is not where staff work, it does not exist.

How to tell it worked

Every record type your office touches has a bin, an allowed route and named people. Every AI tool in use is on the approved list. You have the vendor answers in writing, and you can pull up a log of what each agent did yesterday. If you want help sorting your own systems, that is part of the free AI Operations Audit.