Assessments
Tests that measure judgement with AI, not memory without it.
Candidates are told AI is allowed, the test gives them an assistant, and what they do with it becomes evidence. The reviewer sees the answer and the transcript side by side.
Question types
Six ways to see how someone really works.
| Type | What the candidate does | What it measures | Scoring |
|---|---|---|---|
| Knowledge check | Multiple choice, one correct answer | Baseline domain knowledge | Automatic |
| Review the AI | Reads a scenario and an AI answer seeded with one or two deliberate errors; marks it right or wrong, explains the error, gives the correct treatment | Spotting confident but wrong AI output | Automatic for the verdict, rubric for the explanation |
| Prompt and verify | Uses the built-in AI to solve a task, then submits the final answer and what they checked or changed | Prompting quality, verification habits | Rubric |
| Scenario judgement | A short written answer to a realistic, company-specific situation with no single right answer | Practical experience, priorities | Rubric |
| Practical task | Uploads a file or screenshot from a real tool | Hands-on skill | Rubric, a person only |
| Live follow-up | In the interview: variations on the candidate's own answers | Whether the understanding is theirs | Interviewer scorecard |
Try one
Can you catch the AI out?
This is a review-the-AI question as a candidate sees it, shortened. The AI answer is fluent and sure of itself. Mark it right or wrong.
In the real test the clock is visible, questions come one at a time, answers save as you type, and a dropped connection resumes where you were.
Scenario
A company in the UAE invoices a client in India for consulting delivered entirely online. The client has no presence in the UAE.
"Charge UAE VAT at the standard rate of 5%, because the supplier is established in the UAE and VAT follows the supplier."
Is the AI right or wrong?
The error: the place of supply for a service exported to a customer outside the UAE is outside the UAE, so the supply is zero-rated, not standard-rated. The correct treatment: zero-rate it and keep evidence that the customer is outside the UAE.
In a real test the candidate also writes the error and the correct treatment in their own words. The verdict is marked automatically; the explanation is scored against a rubric by a person, with an AI suggestion beside it.
Every prompt, paste, edit and submission is kept with its time, and shown to the reviewer next to the final answer.
The AI sandbox
The same assistant for every candidate, and every word kept.
- A chat panel inside the test. The candidate sees clearly that every prompt and reply is recorded and will be reviewed.
- The reviewer sees the full transcript next to the final answer, with a timeline of when the candidate asked, pasted, edited and submitted.
- The sandbox model and its settings are fixed per test version, so every candidate gets the same assistant.
- Question authors can mark a question as no AI; the panel is switched off for it.
The AI Collaboration Profile
Four sub-scores, shown as evidence, never a verdict.
For each candidate, Kasauti shows prompting, verification, error detection and final correctness. Each one links to the answers and transcript lines behind it. There is no single pass or fail number, because judgement does not reduce to one.
Four sub-scores, each linked to the answers and transcript lines behind it. Never one pass or fail number.
Scoring
Rubrics first, AI suggestions second, a person last.
- Every rubric question has criteria and level descriptions written by the author before the test goes live.
- AI suggests a score per criterion with a short reason that quotes the candidate's answer. A reviewer accepts, changes or rejects it, and the final score records who set it.
- Calibration mode: two reviewers score a sample independently, and the platform shows where they disagree before scores are released.
- The AI scorer is evaluated on a labelled set before any model or prompt change goes live, and differences in accuracy between groups are reported.
Fairness and integrity
Integrity without surveillance.
- Time limits per question, question pools, shuffled order, one attempt per candidate, expiring invite links.
- Adjustments such as extra time or a different format on request, handled without asking for a diagnosis, and the reason is never recorded.
- No webcam proctoring, no face matching, no "AI-written text" detection. Integrity is checked in the live follow-up instead.
- Company-specific scenarios and seeded errors are checked by a subject expert before a test is published; the studio records who approved each question and when.
- Test versions are frozen once a candidate starts, so edits never change a test someone is already taking.
The Assessment Studio
Write it yourself, import it, or let AI draft it for an expert to approve.
Build
A question bank with pools, time limits, points and author notes. A live preview shows exactly what a candidate will see; answer keys and rubrics never leave the server.
Import
Bring questions from a CSV template or a plain-text layout. Every row becomes a draft to review; anything missing is flagged, never guessed.
Draft with AI
Describe the role and the AI drafts questions, seeded errors and rubrics, labelled as AI and cited to your brief. Nothing is published until a named expert completes the approval checklist.
Next step
See Kasauti on one of your own roles.
Bring a role you are hiring for. We walk through the hire end to end: the test, the evidence and the decision.