Help Center · AI & Trust
Auditable AI and the Decision Audit Export
What 'auditable' means in practice: AI decision logs, an exportable decision history, and what PlacementFlow does not measure.
3 min readOn this page
We say the AI is auditable — deliberately not "unbiased." No vendor can honestly promise unbiased screening; what we can promise is the evidence to check for yourself: every AI call logged and an exportable history of AI-assisted decisions.
Auditable, Not "Unbiased"
"Unbiased" is an outcome claim nobody can prove. "Auditable" is a process claim we can back up:
- Every AI call (screenings, interview analyses, reply classifications, drafts) is written to an audit log with a timestamp, a summary of the prompt and response, the model used, and any score it produced
- For screening and interview-analysis decisions, whether a person then acted on the result is recorded alongside it
- You can export the whole history, on demand, in a format you can hand to counsel, a client, or a regulator
- Candidates can download their own AI decision report from the candidate portal
If a bias question ever comes up, you're not asking us to vouch for the AI — you're pulling the records.
The Decision Audit Export
- Go to Settings → Compliance & AI Governance
- Find the AI Decision Audit Export card
- Pick a date range, click Preview Count to see how many decisions it covers
- Click Download CSV
Each row is one AI decision:
| Column | What it contains |
|---|---|
| Decision ID and type | Screening or interview analysis |
| Date | When the decision was made |
| Candidate | Initials only — the export is privacy-safe by design (no emails, phones, or profile links) |
| AI score and outcome | What the AI concluded |
| Must-know flags | Any must-know checks that fired — flags for recruiter review, never automatic rejections |
| Human decision recorded / date | Screening: whether and when a recruiter reviewed the response. Interview analysis: whether a recruiter logged an outcome; the date is the interview's last update, so treat it as approximate |
| Role | The role the decision related to |
What the Export Is Not
The export is a decision log, not a bias analysis. PlacementFlow does not collect demographic or other protected-characteristic data about candidates, so it does not compute selection rates across groups and does not apply the four-fifths (80%) rule for you. If your agency runs that kind of analysis, you would combine this log with demographic data you hold lawfully yourself. The "Selection-rate (four-fifths rule) monitoring" row on Settings → Compliance & AI Governance says "Not provided" for this reason.
One bias guardrail does run underneath:
- Client-feedback calibration is bias-checked — if a feedback pattern looks like a proxy for a protected characteristic, the AI refuses to apply it and flags it to you instead (see Client Feedback and AI Calibration)
Note: candidate matching itself is keyword- and tag-based with recruiter judgment — PlacementFlow deliberately does not rank candidates by machine-learned similarity to past hires, precisely because that pattern is a well-documented proxy-discrimination risk.
Why This Framing Matters
Vendors that promise "unbiased AI" are making a claim they can't defend — and one that regulators and courts increasingly test. Our position: the AI does the operational work, a human approves high-stakes actions, and everything is logged so you can verify rather than trust. When a client's procurement team asks "how do you monitor your AI for bias?", the honest answer is the decision log plus what we deliberately do not collect, not an adjective.
Related articles
Related articles