What is an ATS?

The applications screen runs on a single route: the filter strip and the list on the left, the detail of the selected candidate in the middle, their document on the right. Picking a candidate does not change the route, the selection is written to a parameter in the address bar. In practice that means the list does not reload from scratch when you set up a filter and click a candidate, the same candidate is open when you refresh the page, and you can send that address to a colleague.

Internship applications use the same shell but sit in a separate tab with three extra filters: university, insurance and internship type. That tab does not come ready on every setup, unlike job applications it has to be switched on.

Application status carries nine keys in code and shows five stages in the panel. Two of the nine keys belong to the same stage, which is why saying "there are five statuses" is wrong and "nine keys, five stages" is right. Two more statuses belong to life after the hire, and those never enter the funnel.

One warning, because this is where demos go sideways most often: the same stage is not worded identically on every surface. The stage badged Interviewed in the list reads Interview on the detail status button and in the audit log. When you walk someone through a screen, use the label that screen carries.

Interview exists as an application status on every plan. Scheduling an interview with Google Meet is a separate capability and is open on the Professional and Enterprise plans. Putting the two in one sentence creates an expectation the product does not meet.

Rolling an application back from Accepted or Hired does not work directly while that application has open document or information requests. The system counts how many are open and warns you. Forcing past it requires a written justification, and the justification lands in the audit log as its own field.

If two people update the same application at the same time, the second one does not silently overwrite the first. The update has to match the last change stamp on the record; if it does not, the request comes back and the user is asked to refresh the page. That is exactly what gets lost in a shared spreadsheet.

Sending a status email to the candidate is a separate decision, and it is not available at every stage. On the candidate side a template exists only for Accepted, Rejected and Reviewing. The email is skipped on Interview and Hired transitions.

If no evidence for a rubric item is found in the CV, that item is not counted as half a point. Its weight is taken out of the denominator and the item is flagged for human review. The rule fits in one sentence: the absence of evidence is a gap in the measurement, not in the candidate.

The reverse is closed off. The denominator of the score is a constant of the position; a criterion the candidate was not evaluated on keeps its weight and earns no points from it. The old behaviour dropped a criterion the model failed on out of the denominator and pushed the score up: on a four criterion rubric, a candidate with coverage of 1, 1, 0, 0 who should have scored 50 was read as 67 once the weak criterion left the denominator.

The required criterion check is never asked of the model. Dates and technologies are pulled out of the CV, overlapping ranges are merged, and the year arithmetic runs in code. The gate returns three values: passed, closed, unknown. "Unknown" is never treated as "closed", because the cost of a false elimination was accepted as higher than the cost of keeping an unnecessary candidate on the list.

That choice came out of measurement. Across four policies, the binary judgement asked of the model was right in 7 of 8 cases and produced 1 false elimination, while the deterministic gate produced 8 correct calls and 0 false eliminations.

Edge cases were measured separately: 12 of 12 difficult cases correct, zero false eliminations. The same measurement also caught a policy mistake. Counting part time experience at half weight was removed, because it turned reduced hours into a discrimination vector.

A person can override the gate with a written reason. Who overrode it, when and on what grounds is stored, and it is not deleted when the score is recomputed: refreshing a score does not erase a human decision.

AI never changes the status of an application on its own. It produces suggestions, shows the decision band and lists the evidence items. The operation that changes a status always runs with a user identity and is written to the audit log. Even a bulk status change asks for confirmation.

The system does not learn from past hiring outcomes either. Who was invited or hired never enters scoring. The reason is not a preference but a known failure: the hiring tool Amazon shut down in 2018 reproduced its own past decisions precisely because it learned from the outcome label.

Every score record carries its own governance block: rubric and scoring versions, model name, temperature, the three valued call of the gate, and the supporting ranges per criterion. A re-analysis refreshes the old score without deleting its previous state. The history row is append only, and the reason for a change is derived from a hash comparison rather than from what the user says it was.

The most useful part of a definition page is often the list of what is missing. GoTeam has no video interviews, no skills tests, no candidate sourcing and no social media screening. There is no LinkedIn, Indeed, HR information system or Slack integration either. Posting views are not counted, the data model carries no such counter; the top measured stage of the funnel is the application that arrives.

Candidate uploads accept PDF only, up to 5 MB. DOCX is valid only in the Google Drive import on the Enterprise plan, never on the candidate side. There is no automatic deletion when a retention period ends either: you do the deleting, the record drops into the bin and can be restored from there.