Unstructured interviews mostly measure how much someone reminds you of yourself. They feel informative and predict performance poorly, which is why confident hires so often disappoint.
What actually predicts
- A small paid trial on real work. The single best signal available.
- Work samples with a conversation about the choices behind them.
- The same structured questions for every candidate, scored as you go.
- Reference calls where you ask what the person needs to do well, not whether they were nice.
The paid trial
Pay for a small, real, self-contained piece of work — a few hours to a day, at their normal rate. You learn how they ask questions, how they handle ambiguity, whether they hit a deadline, and what their work looks like when it is not a portfolio piece. Never ask for unpaid spec work; the good candidates decline and you learn nothing.
Ask past-behavior questions rather than hypotheticals. "Tell me about a deadline you missed and what happened" beats "How do you handle pressure," because the first has to be remembered and the second can be performed.
Try it
Design your paid trial: the task, the time box, the rate, and the three things you will judge. Write the judging criteria before anyone starts, so you are scoring rather than reacting.