Chosen uses AI in several places, and it doesn't behave the same way in all of them. Some of it acts on its own; some of it waits for you. This page is the straight version of what each piece does, where it can be wrong, and how to keep it honest. The short rule: AI reads and suggests on its own — anything that changes your data or sends a message waits for a person.
Resume parsing — extracts, sometimes wrong
When a resume comes in, an AI model reads the document and pulls out structured fields: name, contact details, work history, skills, education. It runs automatically, in the background, the moment a resume is uploaded.
It is occasionally wrong. Unusual job titles, dense two-column layouts, and resumes that are really portfolios trip it up. That's expected — it's a model reading a document, not a guarantee. Every parsed field is editable, so the fix is always to correct the field on the candidate. Resume parsing covers the failure modes.
Match rating — derived, and it recomputes
The match rating is a 0–100% score for how well a candidate fits a job. It's not a number an AI picks by feel. Chosen derives concrete requirements from the job description, scores each one against the resume with a cited line of evidence, and combines them.
Because it's derived from the job description, it recomputes whenever that description changes. Edit the job and every candidate's score for it is rebuilt. The score is a sorting aid, not a verdict — a low score often means the job post didn't capture what you actually want. To keep it honest, write a specific job description; a vague one produces a vague score.
Deep search — reads resumes, has a recall cap
AI deep search plans your query, pulls a candidate set, and has a model read and score each resume. It's the search for queries that need judgment rather than a field match.
The honest limit is recall. Deep search scores at most 150 candidates per run. If your pool has more plausible matches than that, the strongest 150 are scored and the rest are never seen by the scorer. On a specific query this rarely matters; on a broad one it quietly leaves good people out. The fix is a narrower query — and a narrower query is also a better one.
The scheduling AI — drafts, then waits
The AI scheduler coordinates interviews: it drafts the emails, reads candidate replies, and proposes times that fit your calendar and your availability windows.
What it does not do is finalize an interview behind your back. It drafts and proposes; you confirm. It only ever offers times inside the windows you set, so its judgment is bounded by your input — if it's offering bad slots, the calendar isn't connected or the windows are wrong. The email assistant follows the same pattern: it prepares, you send.
HQ — gated behind your approval
HQ, the in-app assistant, is the clearest example of the split. Ask it a question and it answers immediately — reads change nothing, so they're not gated. Ask it to change something and it doesn't act. It writes a plan: a rationale and the exact steps. Nothing happens until you approve it.
That's the safeguard. HQ can propose any change, but a person approves every change before it runs. Read the plan before approving — skimming it defeats the point, and an approved plan runs as written with no one-click undo.
Where the AI acts, and where it asks
| AI feature | Acts on its own | Asks first |
|---|---|---|
| Resume parsing | Extracts fields when a resume arrives | — (you correct mistakes) |
| Match rating | Computes and recomputes scores | — (a score is read-only) |
| Deep search | Runs and scores when you start a search | — |
| Scheduling AI | Drafts emails, proposes times | You confirm interviews and sends |
| HQ assistant | Answers questions | Every change waits for your approval |
The pattern holds across the product. AI that only reads or suggests — parsing, rating, searching, drafting — runs on its own, because the worst case is a suggestion you ignore. AI that would change your data or contact someone — a status move, an interview booking, a sent email — waits for a person. You get an assistant that's genuinely useful without one that acts behind your back.
Keeping it honest
Three habits make every AI feature in Chosen more reliable:
- Feed it specifics. A detailed job description produces sharp match claims; a clear search query produces relevant results. Vague input is the most common cause of a disappointing AI result.
- Check the evidence. Match ratings cite the resume line behind each requirement. HQ plans list every step. The receipts are there — read them.
- Correct, don't work around. A wrong parsed field is editable. Fix it on the candidate so everything downstream — search, match rating — works from the corrected data.