# How to check an AI-generated Swiss job shortlist

Treat each suggested job as a lead. Check the original source, record what supports each constraint and keep missing information unknown. This worksheet helps applicants, advisers and developers review a shortlist without mistaking fluent recommendations for verified vacancies.

By Panta. Published 2026-11-06.
Canonical: https://withpanta.com/insights/audit-ai-job-shortlist/

## Start with the question the shortlist must answer

“Find me jobs in Switzerland” leaves several decisions to the assistant. Before checking results, write down the occupation you want, where you can work, your actual languages, the workload you need and any other essential constraints. Keep preferences separate from requirements. You do not need to put a CV, identity document or private employment history into a public worksheet.

For a short review, freeze the first five results **before** opening them. Keep their original order, the exact request, the tool or product, visible model setting if supplied, and the time. Save that run information in a separate note alongside the sheet; its rows describe individual results. If fewer than five results were returned, delete the unused numbered rows and record the actual count. Do not quietly replace a broken link with a better result and report five successes. A five-result task review is useful for fixing a workflow; it cannot establish an assistant's general accuracy or a market-wide ranking.

[Download the blank review worksheet (CSV)](/research/ai-job-shortlist-review.csv). The file contains a header and five numbered rows, with no real job or personal data. Protocol version: **panta-shortlist-review-v1**. It works in a spreadsheet or as a record structure for an agent. It does not send applications or change a Panta workspace.

## Check one claim at a time

| Check | Evidence to retain | What to leave unanswered |
| --- | --- | --- |
| Original destination | Suggested URL, final employer or authorised ATS URL, and check time | A mirror or HTTP 200 response alone does not confirm a readable employer advert. |
| Actual role | Employer, title and role-specific responsibilities on that page | A matching keyword in an unrelated advert does not establish occupational fit. |
| Place of work | Stated workplace and, separately, any explicit countries eligible for remote work | Swiss headquarters or a remote label does not establish permission to work from Switzerland or abroad. |
| Working languages | Brief relevant wording, classified as required, preferred, alternative, conflicting or unstated | English-written text does not establish English-only work. |
| Workload and schedule | Advertised percentage/range and any stated days or shifts | 80–100% does not guarantee an agreed 80% schedule or four-day week. |
| Contract and experience | Explicit term, training pathway, experience requirement and accepted alternatives | Full-time does not mean permanent; “junior” does not mean no requirements. |
| Availability and application | Explicit closure notice, relevant application destination and what you could read | An Apply button alone does not prove the employer still accepts applications. |

If a page is blocked, unavailable, redirects without readable details or requires access you do not have, record that state. Do not bypass access controls or replace missing evidence with an assistant's guess. A later check can add a new observation without rewriting what was available earlier.

## Record support, contradiction and uncertainty separately

The worksheet has one row per suggested result. `source_access` uses **readable**, **unavailable**, **blocked**, **redirect_only** or **not_checked**. `role_match` uses **supported**, **contradicted** or **unknown** against your declared occupation. A single global “verified job” label would hide which questions remain open.

The language, location, remote eligibility, workload, schedule, contract and experience columns hold concise evidence notes. Use **unknown** when the source does not settle a question. Quote only the short relevant wording you need; do not republish a whole advert. Do not store recruiter email addresses or applicant details in a shared sheet.

For `availability_evidence`, distinguish **explicit_closed**, **no_explicit_closure** and **unknown**. “No explicit closure” records what you saw, not a guarantee that the vacancy is open. Put the relevant source destination in `application_url`; following it does not establish that an application was submitted or accepted.

Use `decision` for **follow_up**, **exclude** or **unresolved**, with a reason. This is your review decision, not an automated hiring or eligibility judgment. A non-essential preference can remain unknown while a role is worth asking about; a clearly contradicted essential constraint can justify excluding that result.

## Two fictional examples

**Example A:** an assistant recommends an English-only engineering role at 80%. The fictional employer advert is written in English, explicitly requires German and offers 80–100%, without specifying days. Record the language contradiction, the range as stated and the schedule as unknown. Do not call it an English-only match or an agreed four-day job.

**Example B:** an assistant calls a fictional role fully remote from Switzerland. Its link leads to a readable advert for a role in another country with a general homeoffice benefit, but no Swiss remote-eligibility statement. Preserve the role's stated location, keep Swiss eligibility unknown and ask for clarification if that is essential. The absence of an explicit prohibition is insufficient positive evidence.

These examples illustrate the protocol; they are not observed vacancies or benchmark results. Panta's separate [language audit](/insights/english-tagged-jobs-switzerland/) and [remote-label investigation](/insights/remote-hybrid-jobs-switzerland/) explain why those distinctions matter in dated observed records.

## Summarise the task without inventing a success rate

Report how many results you attempted to review, how many original sources were readable, which essential constraints were supported or contradicted, and what stayed unresolved. Keep blocked results in the denominator as unknowns. Separate task completion, readable sources, personal fit and application outcomes.

If you repeat the task, retain each run's request, product settings, date and access conditions. Do not fill yesterday's missing observations with today's answers or infer which sources an assistant searched unless the product exposed them. The [research method](/insights/methodology/) describes Panta's wider evidence and review limits. The [English-speaking jobs guide](/guides/english-speaking-jobs-switzerland/) offers broader search channels once you know what you need to verify.