How to add AI to your business without leaking data

Where the data actually goes

When someone uses an AI tool, the text, photo or document they give it is usually sent to the company that runs the AI model. That company processes it on its servers and sends an answer back. What happens to the data after that depends on the provider, the product and the settings: some keep it for a period, some may use it to improve their models, and business and developer plans often have stricter terms than free consumer apps.

So "is AI safe for our data?" really means three questions. What are we sending? Who receives it? What are they allowed to do with it?

The most common leak is a person, not a hacker

The usual way business data ends up somewhere it shouldn't is an employee pasting a customer list, a contract or a spreadsheet into a free chatbot on their personal account, because it was quicker. No one did anything malicious, and the business has no record of it.

The fix is partly a policy and partly giving people a better tool. If your team has an approved AI tool that does the job well, they stop reaching for random ones.

Seven practical rules

1. Decide what never leaves. Make a short list: for example, customer payment details, staff records, passwords, and anything covered by a contract or a law that applies to your business. That list doesn't go to an outside AI provider, full stop.
2. Send the least you can. If the AI only needs a question and three numbers, send that, not the whole database. Remove names and contact details when they don't matter to the answer.
3. Use business or developer access, not free consumer apps. Paid business and developer plans from the major providers generally offer clearer terms on retention and training. Read the provider's data terms and choose the settings that fit your needs.
4. Keep sensitive work on the device when you can. Some AI now runs directly on an iPhone or Mac. For jobs it can handle, like sorting photos or removing a background, the data never leaves the device at all.
5. Ask before sending. If customers use an AI feature in your app, show them which provider will see their data and let them agree before anything is sent. It's respectful, and it builds trust.
6. Keep keys and access locked down. The keys that connect your software to an AI provider belong in secure storage, like the phone's keychain or encrypted secrets, never in the code or a shared document.
7. Log what happens, without logging the private parts. Record which feature was used, when, and what it cost, so you can see usage and spot problems. Don't record the customer's actual data in those logs.

Don't let the AI touch the numbers

Data leaks aren't the only risk. AI models are good with language and bad at arithmetic, and they can sound sure while being wrong. If your team asks "how many hours did we bill last week", the answer should not be a guess.

The safer design is to let the AI work out what is being asked, then have ordinary, tested code look up and calculate the answer from your records. The AI only puts the result into words. That also means the AI sees less of your data: it gets the question and the final figure, not the whole spreadsheet. I built a manager assistant for a workforce platform this way, and it's in daily use.

Keep a person in the loop where it matters

For anything that goes to a customer or affects money, have AI prepare and a person approve. AI drafts a reply; someone on your team reads and sends it. AI reads a receipt; someone checks the record before it counts. This catches mistakes and keeps responsibility where it belongs.

Sometimes the most private option is no AI at all. Tasks with clear, fixed rules are better done by ordinary code: it's exact, it's cheaper to run and nothing leaves your systems. I'll tell you when that's the better choice.

A safe first project

If you want to try AI without much risk, start with something internal and low-stakes:

• An assistant that answers staff questions from your own policies and price list, where nothing private is involved.
• Draft replies to routine customer emails, with a person reviewing every one.
• Reading receipts or invoices into records that someone checks before they're saved.

Each one is useful from day one, uses data you're comfortable sending, and shows your team how AI behaves before you trust it with more.

Questions to ask before you build

• Exactly which data will be sent, and to which provider?
• What does that provider's plan allow them to do with it?
• Could any part of this run on the device instead?
• Who checks the AI's output before it reaches a customer or a payment?
• What will it cost each month, and how will we see that cost?
• Can we switch providers later without rebuilding everything?

I've used AI in production code since December 2022. In my own apps, users agree to each AI provider separately before any photo is sent, provider keys live in the iPhone's secure keychain, and costs are tracked per feature. See AI integration for business for what I build, or the AI tools and automation case study for examples.

What it costs

AI projects start at $5,000 and get a fixed quote after a free call. For larger ideas, a 60-minute strategy session ($250) maps out what data is involved, where it goes and what it will cost to run, before anything is built. AI features also have a running cost from the provider; I estimate it up front and build in tracking so you can see it.