AI Assessments
An AI assessment is an adaptive technical interview conducted by a model. The candidate answers questions one at a time; each answer changes what comes next, so a strong candidate is pushed harder and a struggling one is not marched through fifteen questions they cannot answer.
AI Assessments require a Professional plan.
How it differs from a written assessment
| Assessment | AI assessment | |
|---|---|---|
| Questions | Fixed set, same for everyone | Chosen per candidate as they answer |
| Difficulty | Whatever you wrote | Adapts up and down during the session |
| Length | Every question, always | Up to a maximum, ending when there is enough signal |
| Output | Scores per question | A report with a level, skill breakdown and narrative |
Use it for breadth — establishing roughly where someone sits across a technology before you spend an engineer's hour on them. Use a written assessment when you need every candidate answering identically comparable questions, and an interview when you need to see someone think.
How a session runs
The candidate states the level they consider themselves, which sets the starting difficulty. From there:
- A question is selected for the skill area that is furthest behind its target coverage, at the current difficulty.
- The candidate answers.
- The answer is evaluated as correct, partial, incorrect or skipped, and the difficulty moves.
- Repeat until the question limit or the time limit.
The self-declared level is a starting point, not a claim being tested. Starting a senior candidate at the bottom wastes the first third of the session, and starting a junior at the top produces a demoralising sequence of blanks.
What it costs
AI assessments consume AI tokens, which are metered. Your plan includes an allowance; usage beyond it bills as overage at the end of the cycle if you have enabled it, and is refused if you have not. Both the questions and the final report consume tokens.
Billing → AI usage shows consumption. See Plans and billing for how overage works, and note that AI tokens are one of only two meters that bill past the allowance.
What the candidate is told, and what they are not
During the session, candidates see whether each answer was correct and an explanation. That is adaptive feedback on an answer, and it is what makes the session feel like a conversation rather than an exam.
They are never shown the outcome. No score, no assessed level, no skill breakdown, no strengths or weaknesses, and no recommendation — not at the end, and not through any endpoint. The report goes to you.
That boundary is enforced in code by an automated contract test, not by convention, because it is the line that keeps the hiring decision with the employer rather than with a model. See Reading the report.
What to read next
- Templates — technology, focus areas and session length
- The question pool — where questions come from and how to keep them good
- Reading the report — the output, and how far to trust it