Handling client data safely: a builder's checklist for Israel

Handling client data safely comes down to five questions answered before anything is connected: what is collected, where it goes, who can see it, how long it is kept, and what happens if it leaks. In Israel, personal data is covered by the Privacy Protection Law and its security regulations, and the Privacy Protection Authority publishes guidance builders should read. This checklist is editorial, not legal advice; it exists so that a first automation doesn't quietly become a privacy problem for a small business that trusted you.

Why does this matter for a small automation?

Because the automation touches exactly the data that is regulated: names, phone numbers, health hints in a clinic's messages, payment details in an invoice flow. A lead-reply bot processes personal data on every run. The business owner is responsible for it and usually doesn't know what the automation does with it. You do, so the choices are yours to make well.

Question 1: what is collected?

Map the fields the automation touches. Then remove every one the task doesn't need. A reminder needs a first name, a phone number and an appointment time; it does not need the reason for the visit. Minimisation is the cheapest control there is.

Question 2: where does it go?

List every system the data passes through: the form vendor, the workflow tool, the AI model provider, the messaging provider, any sheet. For each, note the company and where it stores data. Sending text to an AI model is sending data to a processor; check the provider's data-use terms and prefer settings that exclude your inputs from training. Prefer providers with clear data-processing terms, and tell the client in plain words which vendors are involved.

Question 3: who can see it?

No shared logins. Each person with access has their own account and only the access they need. Your own access should be removable in one step when the project ends; put that in the hand-over. Logs are data too: a workflow that logs full messages has copied the data somewhere else.

Question 4: how long is it kept?

Set retention deliberately: run history in the workflow tool, copies in sheets, transcripts in logs. Delete what the task no longer needs. "Forever" is a decision, and usually the wrong one.

Question 5: what if it leaks?

Agree in advance who tells whom, and how fast, if credentials leak or a message goes to the wrong person. The regulations set obligations for the business; your job is to make sure the owner can act within them, which means they need to know within hours, not weeks.

Israel-specific basics

Treat this table as a map of what to read, not as the rules themselves; they change and a lawyer should confirm anything that matters.

The AI-specific risks

The OWASP list for LLM applications names the ones that hit automations: prompt injection through customer messages, leaking data from one conversation into another, and over-broad tool permissions. Practical defences: never let customer text change the automation's instructions, keep each run's context separate, and give the model only the actions the task needs, with a person approving anything irreversible.

Best fit and not a good fit

Best fit: every builder before their first project with customer data, and especially those working with clinics, finance or HR. Not a good fit: as a substitute for legal advice on a sensitive database; use it to ask the right questions, then ask a lawyer.

What to do this week

Run your current project through the five questions and write the answers on one page. Give it to the client. If any answer is "I don't know", that's the task before the next feature.