A chatbot is the right answer to one narrow problem: candidates asking things your careers page should already have told them. It is the wrong answer to almost everything else, and the boundary is where the risk sits.
What it should answer
- The process — what happens next, how many interviews, who they will meet.
- Timings — when they will hear, how long it usually takes.
- Practical things — location, format, remote policy, what to bring, adjustments.
- The salary range, if you publish one.
- Application status, where you can look it up reliably.
All of those have one thing in common: the answer is the same for everyone and already exists in writing. That is the test.
What it must never answer
- Anything about how they are doing. How they compare, whether they are in the running.
- Anything about their chances. Even hedged, it will be quoted back at you.
- Anything with legal weight — visa eligibility, contract terms, notice periods, salary promises.
- Anything it does not know. A bot that guesses at process details is worse than no bot, because the candidate acts on the answer.
Asking questions up front without secretly screening
Asking a few questions before an application is fine. Using the answers to reject people automatically, without that being visible as a decision, is not — and in several places it is specifically regulated.
Three rules keep it on the right side:
- Only hard, objective, legally relevant facts. Right to work, a required licence, willingness to work the advertised pattern. Not experience, not skills, not "fit".
- A person can see the answers, and they are stored with the application rather than used and thrown away.
- If an answer ends the application, that is an automated decision. Say so, record it, and make sure someone can review it on request. Calling it "just a form" does not change what it is.
The pattern to avoid is a friendly chat window doing screening the candidate does not know is happening. The friendliness is what makes it a problem — it invites people to say things without showing them the consequence.
Getting to a human
Every bot conversation needs a visible route to a person, available from the first message rather than after three failed attempts. Put a time on it: say when someone will reply, and meet that. A bot that traps someone in a loop creates more work than it saves, and produces the complaints that get the whole project cancelled.
Telling people, and not excluding them
Say that they are talking to software. Some places require it; everywhere else, candidates work it out within two exchanges and think less of you for the pretence.
Access matters more here than most teams expect. Chat windows work badly with some assistive technology, badly under time pressure, and badly for people who are not comfortable writing in the language. Always offer another way to get the same outcome, and make it just as easy to find.
The scope test
If the honest answer would be different for two candidates, a bot should not be answering it. Everything else — process, timing, practical details — it handles better than you do, at two in the morning.