AI in public services: where it saves a wait and where it creates risks

You open a public authority's website, and a small window pops up in the corner: "Hello, how can I help?" You type a question, and within a second a clear answer appears in straightforward Hebrew, with a link to the right form. No queue, no hold music, no "all our representatives are busy." It sounds like the future. Then you ask something a little more complicated, about eligibility or a case that does not quite fit the form, and the answer is still clear, still confident. It is just wrong.

Authorities in Israel are already experimenting with automated responses, sorting inquiries and translation using artificial intelligence. The public usually encounters it as a chat window on a website. The question is not whether the technology is good or bad. It is exactly where to draw the line: between saving you an hour and potentially costing you a right.

Where the machine excels

Let us start with the good news, because it is real. There are tasks a bot simply handles better than a telephone queue. Repetitive forms. Booking appointments. Explaining an existing process. Above all, making information accessible when it has already been published but is written in hard‑to‑read legal Hebrew.

In all these cases, the machine is not inventing policy. It is translating it and shortening the route to it. There is another benefit that is easy to miss: every "where is the link to the form?" inquiry the bot answers frees a human employee to handle the unusual case, the one that really requires judgment. That is the use that justifies the system.

Dangerous confidence

Now for the troubling part. Language models write confidently. Even when they confuse a legal provision or mix up two similar procedures, the answer sounds authoritative and polite. Unlike an official who hesitates when unsure, the bot never hesitates.

What happens when a resident trusts the answer? They submit the wrong document, miss a deadline or abandon an application they were entitled to make. The most painful part is that sometimes they discover the mistake only after the appeal window has closed.

The red line: decisions about people

There are decisions for which the machine's output cannot be the final word. Eligibility for a benefit. Rejecting an application. A risk score. Closing a case. Each requires the name of a person responsible for the decision and a clear way to appeal it.

Here is a simple rule every authority should adopt and every resident should demand: if a human cannot correct the decision, the machine cannot be allowed to make it alone. A system that sorts inquiries by urgency can help. A system that rejects an application without a person seeing it is something else entirely.

What happens to the text you typed

Almost nobody asks one question while typing into a bot: where does this go? An inquiry to an authority often includes an identity number, income details or information about health or family circumstances. Sending that content to an external model without a clear agreement on its retention and use is not "innovation." It is handing over information.

A resident is right to ask where the text is stored, for how long and who can see it. An authority that cannot answer simply probably did not ask itself that question often enough before making the window live.

Three things to ask of every bot

What should you do next time that little window appears? Use it, certainly. Just keep your eyes open. Here are three things to look for and request if they are missing.

First, a route to a person: a button, a telephone number or at least a way to leave an inquiry that will reach an employee. Second, a reference number, so there is a record of what you were told and when. Third, a reference to the official source. Not just the bot's summary, but a link to the provision, procedure or form itself, so you can check for yourself.

If the bot cannot transfer you to a person, the website's job is unfinished. Convenience without an emergency exit is a trap. Built properly, this technology really can shorten queues and make rights more accessible. The question is whether those building it remember that on the other side of the window sits a person, with a real problem and a right to an accurate answer.